AMD (Advanced Micro Devices) Stock

Okay. Great. Good morning, everybody. Welcome to the final day of the Goldman Sachs Communacopia and Technology Conference. My name is Jim Schneider. I'm the semiconductor analyst here at Goldman Sachs. It's my pleasure to welcome AMD to the stage today. With us from the company, we have SVP and General Manager of the Compute Enterprise AI, Dan McNamara; and Corporate Vice President of Financial Strategy and Investor Relations, Matt Ramsay. Welcome, guys. Thanks for being here.

Thank you.

Thank you.

I think the topic almost every session at this conference is AI. So you're a key enabler of that trend with your infrastructure products. Maybe before we kind of get into those products, how has the AI adoption kind of progressed inside AMD over the past several years from a corporate perspective? What areas have seen the biggest productivity gains? And what lessons from AMD's own AI journey are applicable to enterprise customers today?

I can...

Do you want to start?

I can jump with that. Yes. So look, it's a great question. And I think that when I think about our journey, it's very similar to a number of enterprises. But our team started out -- I think it's a multilayer approach to the infrastructure, right? We started out, first and foremost, with the data layer and optimized that. And a lot of -- and we talk to a lot of enterprise customers, and this is often overlooked is how do you structure your data such that you can actually employ agents effectively. And we actually open source our solution. It's called Optima. So it's -- we started there, and then we've been on this journey about 4 agents driving what I would call automation for efficiency, and that's gone very, very well. And now where I would say is we're really in the domain-specific type applications, right?

So if you think about it for us, domain-specific is EDA. So we're seeing a tremendous amount of upside across coding, debug and those 2 key areas along with Kernel development and just software development in general, very, very strong returns there. And then, of course, across all of the businesses, we're seeing very, very strong automation and efficiencies across each of the businesses. So I would say that -- and it's interesting because we were in New York City last week, and I was with our CIO, and we had a roundtable with a number of top enterprise customers in New York City. And he started out and just walked them through the journey. And it was a very good conversation about where each one of them are on this journey. So I would say that we have -- we're advanced in this area. I would say that we took it on very, very aggressively. And we're also looking at how do you balance sort of token costs with the value, and we're really doing some advanced things across that, too. So overall, very, very strong adoption. You got to look at us as both a provider and a major adopter of AI.

Yes, Jim. The only thing I would add there is you guys saw us work with a big framework that we put together with Anthropic about, obviously, them buying up to 2 gigawatts worth of MI450. There's also a lot of work of not just the Open AI tools, but the Anthropic cloud tools being adopted across our engineering organizations and unlocking a much faster flywheel of software development, debug, time to production of chip programs, optimizing where our software people are spending their time. And we have a huge software organization and trying to figure out what they need to be working on, where can they use tools to accelerate that flywheel versus doing anything manual. I mean my boss, Jean, our CFO, has benchmarked us versus a whole bunch of leading semis and tech companies. And I think we're on the bleeding edge of AI adoption internally. And it's come with an increased token cost, but it's come with an even a much, much greater productivity gain across the organization, and it's allowed us -- I think you'll see it allow us to bring hardware and software products to market much more quickly as we go forward.

Yes. Actually, just one last point. I want to just emphasize that, right? So you've got domain-specific and then you've got sort of what I would call general IT automation and one is for efficiency. But when you can drive a faster time to market, that's where the real rubber hits the road. And that's what we're after. As we go to external enterprises, our goal is to get them time to value very quickly, right, with ROCm and with some of our solutions. So it's -- again, we always say we eat our own dog food, right? Everything we build is deployed in our data centers first. And it's going very well in terms of driving our time to market with our engineering teams.

Yes. And then with respect to your customers from their perspective, how do you think this plays out in terms of model evolution over 3 to 5 years in terms of the landscape, I mean, do you think frontier models still going to kind of be leading the charge here? Do we see small language models kind of like do a lot more kind of task-specific things? Or do you think open weight, open source models are going to have a larger role to play?

Do you want to start or I can.

Yes. I think, Jim, the answer is yes. It's not a very helpful answer, but it actually is the answer. I mean our goal is to make sure that our combination of CPU and GPU road maps are very differentiated in terms of driving tokens per dollar outcomes regardless of whether it's open weight models, frontier models for our largest customers. I think those -- I mean, obviously, the industry is evolving quickly around what are the right use cases, how should we say this? How to apply the right tokens to the right problem relative to the cost of the token versus the return of the token. And that's a very large continuum.

I think our goal is to make sure that on the GPU side, our hardware and software are driving the right efficiencies regardless of whether it's open weight models or closed-loop models or frontier models and that Dan's business is the right CPU to run agents to drive all of those models regardless of where they come from. I mean that's kind of our goal. I don't know, Dan, if you want.

Yes. I would just say, look, we rolled this out -- I showed this at our Advancing AI Day, right? And Matt's right, it's all of the above, right? Clearly, frontier will continue to be the cutting edge, but open weights are very, very valuable. And then you've got sort of what I would call SLMs for some of this domain-specific stuff, right? And what we showed was intelligent routing, right? So if you think about it, you've got frontier. Every enterprise is going to have some distributed model around frontier, probably GPU as a service, right, in the cloud, most likely an on-prem server that can service and run open weight models.

And you have a policy-based router, right, depending on the task. And you're looking at performance latency, you're looking at, obviously, security. That's one of the key areas where -- what I hear mostly is cost and security from the enterprise, right? And what you can do is you route this and you can manage your costs, you can manage -- if it's a policy-based router, if it's highly secure, it stays on-prem. So we see a lot of enterprises trying to build this out. And it's very interesting because it will -- enterprise are infinitely hybrid, and we believe that will continue.

Great. Now I want to dive straight into your business. First, the AI business and also the server CPU business as well. Your AI data center business has grown very rapidly over the last few years. If you think about the biggest strides you made in product development across silicon, software, customers, ecosystem, where do you think you can make the biggest stride going forward? And kind of like what are your key focus areas from here?

Yes. That's a great question because first and foremost, I'm I always say this because it's very, very important. Our vision for many years now has been you build the right compute engine for the right workload. And that's across CPUs, that's within -- not only across the product lines, but within the product lines, right? So -- and we'll talk about server at some point, but we optimize for workloads. But most importantly is we feel like we're in very, very good shape across the different product lines, right, from server to GPU to networking, right?

And now ROCm is coming online. So I think the biggest part for us is we have now shifted from this sort of individual product lines to a full system provider. So providing the full rack, all of it interworking. And we're also driving a different road map cycle, right? It used to be -- 3, 5 years ago, it was like you're optimizing for your product. Now it's a combined data center road map steering group, right? So whatever I'm doing, our -- you have to make trade-offs across all of the products. So I think that's the biggest change. And what you'll see is like getting rack scale solutions at scale is the biggest thing we're focused on right now.

I think from my perspective, just listening to -- I mean Dan spoke about it just now, but listening to Lisa and others speak about we don't necessarily have to force ourselves to be -- if you step back and think about the top -- I mean, there's a long tail of customers that we're going to continue to support. But if you think about the large top 15 or 20 or consumers of high-performance computing cycles, in the world, we don't need to necessarily be their CPU partner or GPU partner, FPGA partner or semi-custom partner. We can walk into a room strategically and say, how can we, at scale, be your high-performance computing partner.

And that might look differently at different customers, but it's a very powerful thing to be able to say, hey, we just want to be your high-performance computing partner and let's think strategically about what you want to do over the next number of generations and put solutions together that can support that across endpoint, across inference at the edge, across server, on-prem and in the cloud, AI deployments in massive data center scale or in PCs or there's a huge continuum of how can we be your high-performance computing partner and being able to bring those pieces of IP to the market at significant scale is one of the things that I think is unique about what we bring is it's not a push approach. It's how can we be your partner and let's decide how we're going to work together to bring significant amounts of high-performance computing to market. I think that's the biggest change that's happened. And now that AMD has this full breadth of portfolio and the scale that we have, that's a conversation that I think is valuable.

Great. Now the company has outlined some pretty healthy revenue growth targets, 60% CAGR over the next several years in data center revenue, 80% CAGR in AI data center revenue. Talk about 2 elements of that. One is how diverse does this get between the hyperscalers, CSPs, enterprise, AI labs over time, even sovereigns. And then -- so how diverse does it get? And then secondly, what should we be thinking about in terms of markers for more of the short term going into 2027?

I'll start. Yes. Maybe I'll start and Dan can add a bunch of detail on his business in server. Yes, Jim, we have outlined -- we started at the Analyst Day back in November, and it's amazing how long ago that seems given how fast this industry is moving now. But we talked about more than 60% growth of the data center franchise, more than 80% of growth of the AI business. And at that time, we thought we were well above where the market was and talking about a $60 billion server TAM, and we've now more than tripled that so we're -- at that point in time, I talked about the company growing at more than 35% annually. Lisa and the team have updated the TAM for AMD to be more than -- around $2 trillion by 2030, and that's a 40% growth rate of the TAM, and we expect to grow faster than that as a company.

And we talked about getting to more than $20 in earnings over the sort of strategic time frame. And I think we've updated that to be significantly more than $20. So we're excited about the growth, the leverage in the model. And we've given a few data points on 2027, much more than doubling the data center business and those are things that we feel really good about. And now it's just putting our heads down and making sure that we scale the AI business in terms of building racks and Dan's business is in a very, very different place than it was 12 or 24 months ago in terms of growth. So we feel it's a very, very exciting time at the company, but at the same time, we're heads down and trying to execute. So I don't know, Dan, if you want to expand on that.

Yes. I would just say, look, the way I look at our AI business is very similar to the way I looked at the server business 5 years ago, right, is very deliberate approach, you get in. And if you look at what we did in server, it was strong in cloud and national labs and then we evolved into the enterprise, right? And now we're seeing very, very strong growth in the enterprise. I think you'll see the same thing happen. Like Matt said, we're very focused on delivering to our top customers right now with Helios. But the spread will happen, just like I just talked about, the enterprises are really thinking through what -- how their -- what their overall infrastructure needs to look like.

It will include cloud. But if you think about AI, it's the exact opposite of what happened in general purpose compute. General purpose compute started on-prem and went to the cloud. It's the exact opposite. And we are seeing many of the mainstream enterprises look at building sub-rack scale, whether it's PCI card type deployments or 8-way server UBB-based deployments to do exactly what I just talked about in terms of what is the right balance and what's the distributed architecture that you need for the long term. So I think what you'll see is the shift happen over time. But right now, like Matt said, we're pretty -- we're concentrated from an AI standpoint. However, with server, we really are -- Lisa and Jean talked about the results we're seeing across the enterprise as well as cloud. And it's growing quite dramatically right now in terms of share gains across all of the mainstream enterprise and the channel. We've invested very, very heavily over the last few years to go drive the channel and the enterprise, and it's really starting to pay off. So I just -- like there's no sort of fixed ratio, but it's more of a -- I see the same evolution happening across the AI business.

Fantastic. I want to dive into server CPUs next, your home turf, so to speak. So for investors less familiar with the technical details of Agentic AI, maybe help us understand why Agentic workloads actually drive higher attach rates for CPUs. And sort of as you do that, maybe talk about the changes in sort of system architectures that occur as the customers move from sort of simple inference to sort of more autonomous multistep AI workloads.

Yes. Look, this is a hot topic. And I think I would start with saying that this is more of a distributed systems architecture problem as opposed to a simple linear problem, right? If you think about the world of ChatGPT from November '22 to probably into last year, very linear, right? It was a SaaS-based data center. You have your servers for web serving, you've got your application servers, you've got your database storage, you've got caching. And then you've got sort of this GPU server, right, which everyone understands the GPU server, right? You know the ratios. Everyone can calculate that very easily. And that was very linear, prompt response, right? That's what it was built for.

Well, with Agentic, as you all know, it's an entirely continuous flow. It's a completely different compute paradigm. It's 24/7 churning whether within the sandbox spawning numbers of different agents. So if you just think of the picture I tried to just draw for you, if you think of your traditional servers here and your big GPU servers here, you kind of open it up and you pull in a whole new class of compute, which is for agentic, control plane, API calls, database queries, tool execution. And that is pure CPU-based. So that clearly will do RL with the GPU servers. So GPU servers grow also. But if you think about those general purpose servers, those get uplifted too, because more and more calls to those. So you're seeing an uplift in a whole new class plus the traditional general purpose.

And we're just seeing that dramatically grow, right? And at our FAD in November, I said that, look, this there's multiple areas of growth for the CPU. We called it, but we called it too low, right? So we've upped it now. And I think the growth we're seeing across both the enterprise and the cloud is very, very exciting. And then lastly, what I would say is with Venice, we are hitting on the 3 main focus areas for CPU, right? You've got your GPU server that everyone knows and loves in terms of started out 1 to 4, a CPU to GPU. Then you've got this Agentic sandbox CPU, where with Venice with our high-core-count 256 core device, that is -- if you think about Agentic, it is really threads per watt with the right level of per core performance.

If you think about the head node, it's really about IPC and high frequency, driving and keeping the GPUs busy. And then the general purpose servers, we've been very, very strong there for many years, and we're going to continue. So when you think about it, we feel like with -- not only with Turin today leadership, Venice as we launched it already and as it comes online here through the back half of this year, we are extremely well positioned to capture this growth. But I'd say one last thing. If you're trying to find a number to plug into a model, it's very, very hard because there are so many things. If you just think of a gigawatt of power, right, and then you factor in your PUE and you come up with your IT power, it's all about the addition of the CPUs.

Again, the host node, you know, we all know that's an easy calculation. but it all depends on what you're trying to run. It's really workload dependent. And that's why it's so hard to plug a number in. But that's why we tried to capture sort of, hey, this is the growth we see. And when we show it for Agentic, it is also pulling in the uplift in those general purpose servers that I talked about. So I don't know if I confused you more or not, but just trying to give you the picture of what we're seeing.

Yes. Dan, maybe I'll just add one thing. I mean we did take a $60 billion TAM out to 2030 and upped that now to $120 billion and then $220 billion. And the company -- I know what Lisa is expecting of you Dan is for your business to be over 50% of that TAM as we grow. And it will be -- we can do a relatively small number of chiplets and put them together in configurations that can be a significant number of SKUs and a full coverage of the platform. So I mean you guys can do the math on more than 50% of the $220 billion. I mean it is -- I've been following and now part of AMD's server business for a very, very long time and to talk about building a $100 billion server business is pretty exciting. No pressure, Dan.

But the -- that's what we see coming is a significant growth of Agentic sandbox CPUs for which we have very large core count multi-threaded parts of strong growth of head node CPUs where we have really high-frequency focused high bandwidth, high single thread performance parts and then the broad range of the server market. One of the things that stuck out to me seeing the results of Dan's businesses in the second quarter, I mean, it seems like forever ago, we talked about the second quarter, but even the enterprise part of the server business grew more than 70%. That -- the industry has not seen those type of growth rates in enterprise server basically ever. So we're very excited about all parts of the server business and the breadth of SKUs and the breadth of platforms as we roll out Venice and then move into the Florence generation is something that we're really excited about.

Yes. Now server CPU market, as I said, talking about it's also becoming increasingly competitive even as it's growing. So what advantages do you think the x86 ecosystem continues to provide for the enterprise specifically? And how do you think about the durability of x86 in the hyperscaler environment, especially for some of these internal workloads where customers are developing their own silicon?

Yes. This is a common question, right? So I mean, first and foremost, we always talk about this, right? This is not an instruction set architecture problem or concern, right? There's no fundamental differences in the ISA between x86 and ARM. It's really about delivering to different optimization points, right? It's perf per watt per dollar ultimately. And we know that if we continue to drive along the 3 swim lanes that we just talked about and optimize for that performance per watt, we're in a very, very good position. And if you think about from an ecosystem standpoint, if you go back to that picture, I tried to draw with my hands, all those general purpose servers that I talked about, x86 based today. Lots of software built for x86.

So the ecosystem is built around x86. So all that growth comes on x86. Now if you look at sort of the hyperscalers, each one of them are doing some form of their own. And what we see is if we continue to drive just what I talked about, which is the highest throughput and core density per watt and then we hit these other points, we feel extremely good about the design-in that we have right now across all of the major cloud vendors in the world across -- from an Agentic standpoint at 256 core, from a high-frequency standpoint at 96 core and then just across other SKUs for high-performance computing.

And even though -- I'll just give you a good example. Like recently, Amazon came out with RDS, which is their database service, which is a first-party property, right, that we would classify. It's on Turin. And the reason why is performance. So we just know that, yes, would they -- there is a focus for them to try and get their first-party properties on their homegrown, but it doesn't fit for everything. And again, it's -- even when you go high density, it's that perf per core sweet spot and that optimization point on the VF curve that we very -- we pay close attention to. So we really feel like where we are today with coming out with Venice, well, turn today with Venice coming out as we speak and ramping. And then I just -- I was looking at -- we had a review earlier this week on even Zen 8 in terms of what our engineering teams are targeting. So I feel very, very good about where we are in terms of delivering the optimization points. That's the key, right? It's really optimizing for the different workload and the deployment model.

I think, Dan, I agree. I mean, from my perspective, it's not watching the teams internally, the investor focus tends to be much more around instruction set. And it is important for the enterprise pieces of the server market, whether that's on-prem deployment or in cloud deployment. But the economics of rolling out the server market to so unprecedented scale that we talked about with the TAM, it's about building the best server parts, period. Never mind the instruction sets. And I think that's what we -- from a scale and supply chain point of view, from a -- I think you said optimization points and the number of SKUs that we can roll out, the number of platforms that we can roll out, the significant amount of optimization you can do for different places in the road map, I feel really good about where we are.

And you can it's not just what we think about the market. We can see the demand pull from customers for different optimization points. And so when we think about, okay, this is where the demand pull is and these are conversations that are multi-generation in nature. I think we feel really good about where the server business is.

I would just -- final point on that is for Venice, and I'm pretty sure Lisa talked about this at our last earnings, but with each generation we built builds on the next, right? And you get more and more of the ecosystem coming along with you as you go. And we've been very, very focused on that. But with Venice, it's the broadest true launch that we've had in terms of OEMs, ODMs, cloud vendors, the ISVs on day 0 support. We're just very -- the demand is very, very strong due to the 3 swim lanes that I talked about. I think our customers in the ecosystem are seeing that one SKU doesn't solve every problem, right? And that's kind of what we're seeing from a merchant arm standpoint. It's really just sort of singular SKUs or 1 or 2 SKUs. So we're pretty excited because it really is in a -- we're in a very good spot from a market opportunity standpoint and our product portfolio leadership across really, I would argue, 3 generations straight.

Excellent. One of the things that's striking me over the past couple of years, we've kind of changed the parlance of how we talk about this market. We're not talking about server counts or counting accelerators. We're talking about counting gigawatts of capacity. And every single presentation at this conference has done that. Maybe as you think about these multi-gigawatt AI deployments, how should investors be thinking about CPU content per gigawatt?

Do you want to?

Maybe I'll start. We I think the focus that we have at AMD broadly in our data center business is to make sure that we provide very compelling tokens per dollar and TCO for our GPU business, and we're right in the throes of ramping and launching Helios and MI455. And you'll see us be a very large partner to some of the leading model companies in the world to run their inference workloads. Separately, Jim, regardless of whether the inference runs on our GPUs or NVIDIA GPUs or TPUs or whatever XPU. I mean, Dan can expand on this, but I think what we're focused on in the server business is to make sure that AMD's Venice portfolio and going forward are the -- there's a differentiated and right place for the industry to run agent code.

And so we haven't been super specific about what that ratio is in terms of gigawatts of deployment because it does look different depending on what customer it is. But we want to grow a very large AI business. And I think Dan's business is positioned to be a significant majority of the industry running agents to power Agentic AI. And so we haven't been super specific on the gigawatt comments in terms of CPU.

Yes. I would just -- really simple. It depends. Because the challenge is -- so take a gigawatt, you do -- again, do your PUE, you've got this IT and then you've got to break it down where, okay, I've got clusters of GPUs over here training. I've got clusters here doing inference, then I've got to support it with a general purpose and then the agents, and it just really depends on what you're trying to accomplish with that gigawatt. And it's very hard to just say, oh, here's a fixed ratio. I would say that it's growing, right? Like if you think about it today, we're saying 1:1 sort of ratio, and it's going to continue to grow. But it's just very hard to pinpoint, plug this into a model and you'll get what you're looking for. It's very highly dependent on what the end customer is trying to accomplish.

I spend a lot of time plugging numbers into the models. Okay. We're almost out of time, but let me leave you with the last question for you. If we are -- we covered a lot of ground. If you think about your position in AI, compute, data infrastructure -- data center infrastructure, et cetera, if we're up here on stage again in 5 years and we look back, what do you think the 1 thing or 2 things that investors remain most surprised about in terms of the performance of the company?

Dan, you want to start off?

Look, I think maybe I'll start with maybe what people may be missing about us, right? And it's what I talked about earlier is we have fully transitioned from a very, very good silicon provider across multiple products to -- and we are transitioning now to full rack scale. And our software has come even over the last 6 months, the gains we've seen, we're becoming more of a software company and a systems company today than we were even 6 months ago. So I would just say that, that will be, I think, if you look forward 12 to 24 months, I think it will become very clear how we have made that transition quickly, and we've driven a software stack that is truly focused on time to value for our customers. I think that's where I would -- I'd leave it in terms of what you'll see over the next few years.

I mean from my perspective, we're -- the goal is to -- I mean, Dan started the conversation this way, Jim, where we want to provide the industry that consumes high-performance computing with the right type of computing for the right type of workload. And I think that, that will serve us well across the breadth of our markets. And we're right now at one of the more exciting times that the industry has seen and more exciting times for the company. We've talked about much more than doubling our data center business next year and driving gross margin dollars very significantly faster than expenses, right?

So we're at that inflection point. And I think it's important for the investor community to understand that Lisa and the whole team -- it's funny, we were doing a meeting in a room at the conference just an hour before we came on stage here, and Dan was on an execution meeting with Lisa and the team, right? So it's like the team is focused on making sure that we have a cadence of execution at the company. And despite all the excitement there, the focus remains on making sure that we execute. And if we do that, then I think investors will be really pleased with where things end up.

But we -- it's not about driving -- for us, it's about driving outcomes for customers and then that will translate into outcomes for the investment community, not the other way around. So we're just going to put our heads down and execute because it's a super exciting time. But as we wrap up here, the little blinking light is on. But thank you all for spending time with us, and thank you, Jim and the team at Goldman for hosting us. We really appreciate it.

Dan, Matt, thanks for being here.

Appreciate it. Thank you.

Thank you.

Welcome to day 1 of Citi Global TMT Conference. My name is Atif Malik. I cover U.S. semiconductors and semiconductor equipment stocks. It's my pleasure to welcome Jean Hu, EVP, CFO and Treasurer; Matt Ramsay, CVP, Financial Strategy, Friendly Neighborhood IR from AMD. I'll kick it off with my questions first. If you have a question, save it towards the end, we'll have the mic come to you and you can ask your question.

Welcome, guys.

Jean, if you can take us through the lay of the land. It's been an exciting year. What has changed from January till now, you guys had your big AI day as well. You raised your TAM forecast. So just kind of walk us through how the demand picture has evolved from January onwards.

Yes. First, thank you for having us, and it's great to be here. Thank you, everyone, for joining us. Yes, it has been a really exciting year. They have so many different changes in our industry, and it's really about AI. And we do believe this is the most consequential technology transformation. When you think about the pace, the scale, and the rise of the AI, it's unprecedented, right? When you look at this year, not only the model capabilities continue to advancing very quickly. We also see inferencing outpace training from AI compute demand perspective.

And agentic AI, that's another thing, which really have another step change in demand, not only just for GPUs, but for CPUs -- so during our Advanced AI Day, we did update our market opportunities from in the past, a small number to now $2 trillion in 2030. We do see the demand for our product, not only just GPUs, but CPUs and AI PC, and in the future, physical AI for our adaptive compute across all different areas.

And the most exciting thing is we just reported the Q2 results, and we talked about our data center business are expected to double next year. Our GPU business continue to ramp; MI450 we're going to launch this quarter at the very beginning of the ramp, and then we are going to see a very significant ramp in Q4 and into next year. And on the CPU side, it was supply-constrained, but we have been increasing supply.

So we do expect the second half, the CPU business is going to expand more than 80% year-over-year and into next year, continue to expand more than 70%, which should continue to be supply-constrained. So it's a very exciting time for AMD. When you think about it, the company has been very aggressively investing in high-performance compute for more than a decade.

As you know, we have been building methodically the compute platform from CPU, GPU to adaptive compute across all different end markets. We think the most different thing, this AI super-investment cycle is at the very beginning. And over time, we're going to continue to see strong demand for AMD's product. And the portfolio we have built will benefit from this AI super-investment cycle.

Great. Matt, as you talked about inference being a big driver of demand, and we've been hearing about this disaggregated compute kind of shift in the market, the Hot Chips conference. And you guys struck a partnership with Cerebras. More recently, you've acquired Taalas. Can you kind of talk us through what your strategy is on the inference side and how you're thinking about these different pieces of the different type of inference units.

Sure. And thank you all for -- and Atif, thank you and the folks at Citi for hosting us and for everyone to come and see us. And then maybe this ties a little bit into some of the stuff that Jean talked about in your prior question about what's changed in the market in the last 12 months. It seems like a lot. We were waiting for inference to become the majority driver of AI computing, and I think that's happened. And at the same time that, that's happened, we've seen this radical change from what I kind of call chatbot inference to agentic inference.

And that's really created big opportunities for both inferencing silicon and for the CPUs to run the agents. We've talked about in a number of forums, obviously, the breadth of the inference market will be -- the majority of it driven by GPU-led computing, and our MI455 product is going to be ramping now with Helios, and we have a road map to continue to innovate there. There's obviously different ASICs and XPUs in the market that are going to do some of the inferencing work.

And there's this new sort of market for disaggregated inference and ultra-fast response time tokens. It's a relatively small piece of the market today, but I think depending on economics to grow into a larger piece over time. So we've taken a multi sort of pronged approach to the problem. One is a partnership that we have announced with Cerebras where our Helios systems will be in their cloud and compute alongside their Wafer-Scale Engine racks to expand the utility, not just of ultra-fast inference, but take it to a broader range of more general-purpose inference in their cloud for their customers.

And then longer term, we do have -- we haven't given a ton of details yet, but we do have some of our own internal silicon ambitions for ultra-low latency inference that would fit into our architecture via chiplets. And the Taalas team brings a lot of really good talent. We know the folks well. Some of them were former AMD, ATI folks back in the day and are going to integrate really quickly into the team and add a lot of technology and horsepower behind the internal silicon work we're doing in that area.

Awesome. Let's talk about Helios. I believe you guys already had very high expectations on Helios exiting last year. But just in terms of your shipments and where you stand on Helios, how should we think about the ramp of Helios in Q4 and then into Q1 next year? How is that ramp going? I do hear from clients around questions around execution. If you can just help us understand.

Yes. I would say the Helios ramp, MI450 ramp is going very well. Q3 will be at the very beginning of our production shipment. We expect revenue in Q3, but Q4, we'll see a very significant step-up and then another step-up in Q1 2027, and the ramp through 2027. it is a scale level, rack-scale level, right? It's very different, very complex. So we have been very methodical and deliberate how we design the ramp process, start to ramp in the end of Q3 and the step-up in Q1 -- Q4, and Q1 into 2027.

We are working with all our partners, ODM partners, and our overall supply chain to ensure we have all the components. So, it's not about the GPU, CPU, or HBM memory. There are also a lot of other small components that we need to ensure we have. And with the ODM partners, it has been a long really, really durable work with them to make sure the manufacturing process can execute. That's why it's a methodical process. It's our team have been working with the ODM partners.

We also have been working with all the customers. We're going to have a production shipment to ensure not only all the component works, all the mechanical, all the software stack, everything works. So that has been ongoing. It's a weekly execution process. We actually feel pretty good about the ramp, both on the supply side. The demand and volume for 2027 has certainly go above our original initial expectation. We need to continue to expand the supply. We feel good about the financial plan. We have talked about it, but it's absolutely the case. We can use more supplies to make sure we meet the customers' demand.

And just on the customer profile, you guys had Meta and Microsoft as 2 customers last year. You're growing that demand further this year, and then you added Anthropic as a customer. Can you just talk about that customer funnel, particularly touch on the neo-cloud opportunity, if that's an emerging area for you? And also touch on the dollar-per-gigawatt across that range of customers?

Yes. So we are really pleased with our strategic long-term partnership with Anthropic -- we announced -- when you have Anthropic, basically, we have 3 major anchor customers, Meta, OpenAI, and Anthropic. And all 3 of them, the deployments are going to be multiple gigawatt scale deployment and multi-generational engagement with all of them.

When you think about our customer pipeline, not only all 3 of them give us the forecast more than what we expected in the beginning of the strategic partnership. Of course, we need to make sure we have supplies to support them. But more importantly, MI350, we have seen a tremendous demand from all the other third-party customers, the model builders, the new AI companies, even the enterprise customers.

So we do think the neocloud continue to be the area we're going to work with to make sure we meet all the other end-customers' need. The pipeline is quite significant. For us, we really need to prioritize. We want to make sure we support the 3 major anchor customers for large-scale deployment. But at the same time, you should expect us to work with neocloud to make sure we support all the other customers in the market.

Atif, one thing I would add to what Jean mentioned is it's not -- when we think about these programs, we don't just think about it as what's going to happen with Helios and MI455 over the next 12, 18 months. It's the engineering level and technical engagement that's influencing what the MI500 program will look like, the MI600 program will look like, and what the design of the future racks will look like.

The amount of technical engagement across these the 3 leading model companies that we've announced as customers so far, that's -- it's really heartening to see that influence of the road map over a multigenerational period. And I think that gives us confidence in not just where our road map is going, but what that engagement level is going to look like over the -- through the end of the decade, right, over multiple generations.

And we saw the same thing happen. It's a very different time in a different market, but the same thing happened on the CPU side 6 or 7 years ago, right, where people were announcing partnerships with AMD on the wrong generation, but it was really influenced by what they saw and their level of influence in the road map of multiple generations going forward from there. And I think that's kind of where we are now with the AI business.

Jean, the MI Series is below corporate average on gross margin side. And how should we think about the ramp and the move to MI355 to MI500 in the future, and the impact to your gross margin expectations next year?

Yes. Thank you for the question. I think data center AI, MI450 and the future generation, it's one of the most significant growth drivers in our data center business. What it is going to help us is to drive very significant incremental revenue and gross profit growth in next year and beyond. And even though the gross margin percentage right now is still a bit below corporate average, but the way to think about it is we actually have a broad portfolio at the company level.

So, we always talk about the gross margin is being driven by the mix of our different product portfolio. And one of the things is, going into Q4 and 2027, even though we are ramping MI450 significantly, but we do have some tailwinds on the gross margin side to help us. I think the first one is the server CPU business. When you think about the server CPU business right now, within our data center, it's still a larger portion of our business, and it's actually going to continue to grow.

We expect second half to be more than 80% year-over-year increase and next year, more than 70%. So from that perspective, that business is gross margin accretive to corporate average. So, we do see that is going to help to offset some of the dilution from Data Center AI business. Secondly, our Embedded business, after 3 years of inventory digestion, we have seen a significant double-digit year-over-year increase, not only in Q2, we also guided in Q3 and going forward.

So the business recovery is very broad-based in Embedded business. And we are also winning a lot of design wins with our embedded x86 business in data center, in networking, and all those business are margin accretive to us. Third thing is probably small, but our Gaming business, which tends to be the lower gross margin, which is at a later stage of product cycle and the memory cost is pretty high right now, which also impact the demand side.

So the mix from a gross margin perspective, actually, we do have all those tailwinds. In general, the way to think about it is when we ramp MI450 in Q4 and 2027, the gross margin will be slightly lower than what we guided Q3. Q3, we actually guided our gross margin at 56%. You have seen we are expanding gross margin since last year when we actually are ramping MI350.

So going forward, gross margin in 2027 quarter-over-quarter will be different. But the most important thing is the way to think about our business model is we are driving very significant Data Center revenue expansion and the gross margin dollars expansion, which our investment and OpEx increase is slower than revenue and gross profit increase, which is going to drive very significant operating leverage and earnings-per-share expansion.

Very clear. Let's talk about the server CPUs. When I visited you guys in, I believe it was January, something for a bus store, some bumps going on in terms of the CPU demand this year, and you guys were talking about not finding enough CPUs internally to do agentic AI yourself. So kind of walk us through -- and now the TAM is like $220 billion by 2030. And can you just talk about your aspirations around market share in the server CPU market? What have you seen so far x86 versus ARM? And just your aspirations on market share?

Yes. I'll start, and Matt can add. It's actually astonishing when you think about the server CPU market expansion. I think when we had our last Financial Analyst Day in November 2025, we talked about server CPU market TAM to be at $60 billion in 2030. At that time, agentic AI was at the very early beginning and really agentic AI adoption, the diffusion into the business enterprise has been -- the curve has been tremendous.

So we see the agentic AI adoption very significantly starting January and just continue to be like almost like a vertical in enterprise market, which that's when the demand for CPU continue to increase because as Matt mentioned earlier, when you think about agentic AI, it's about workflow execution. In enterprise that really requires retrieving data, execute orchestration, all those are being done on the CPUs. So not only you needed to go back to your foundational CPUs to run all the tasks, you actually have increased the layer agentic AI sandbox, we think it's a new segment, which is to executing all the agents' tasks to make sure they coordinate with every other compute.

That market has been just continued to increase. We can see the demand continue to go up. That's why we're we have just updated our TAM opportunities to more than $220 billion from $25 billion in 2025. That is how significant the expansion has been. We see more than 50% CAGR in the next several years, largely driven by agentic AI. They are different segments. The way to think about it is you do have the foundational CPU segment, which handles all your enterprise applications, your SAP, your database, that continue to grow, but the growth is probably not as big as the other segment.

Then there's the high node part coordinate with GPUs, that continue to grow. But the largest is agentic AI sandbox, which is very small today, but it's going to be more than 50% of that $220 billion market. And we do think it's not only about ASP increase, the unit will increase very significantly, too. So the market opportunity is tremendous, and we are very well positioned as a company from an investment perspective.

Yes. Just to add a couple of points, Jean. We feel like the road map is in about the best place that it's ever been in the server business. I think Atif, you asked the question around instruction set of x86 versus ARM. Just to be clear, I mean, there are areas in -- particularly in the enterprise server applications where x86 legacy is very important. There are broad applicability, not really of instruction set, but of experience in security features, reliability, serviceability, like really enterprise and cloud-grade features that we have in the sixth or seventh generation of the road map that are quite important.

But we also feel like our own differentiation and our competitive lead in the server business will expand with Venice pretty significantly. I mean, it's a very, very compelling program. And the amount of demand and visibility that we have in the server business is very significant from what it's been historically. And I think we get asked about competition a lot. And I think our approach to this is not an x86 thing or an ARM thing. It's a build the best server parts period. I mean that's the priority of the business.

And I think if we do that, we can give differentiation in the agentic world of threads per rack or threads per megawatt. We can give over 5 gigahertz products that go into head nodes. We can have broad applicability across the enterprise stack, whether that's on-prem or in the cloud. And one of the fascinating -- you mentioned some of the growth rates, but one of the fascinating stats to me, having been in the server industry since 2000 is, I mean, our enterprise server business grew more than 70% in the second quarter.

I mean, it used to be heroic if you had double-digit growth in enterprise server, and now we're talking about 70% growth. So it's a really broad-based portfolio. And I think going from the Venice generation into Florence and Ravenna and beyond that, I think that we do intend to -- the much larger TAM that Jean described, we do intend to grow to 50% of that dollar TAM inclusive of all instruction sets of competition. And when you do the math and you talk about building a $100 billion server business, and that's what we're intending to do.

Great. Jean, let's talk about supply. This topic never dies. You guys have talked about $70 billion in Data Center sales next year, maybe low $40 billion GPU and the remaining CPU. You saw purchase commitments, $29 billion, $30 billion. Just -- and then we can all track and hear about TSMC's allocation, and you guys are seeing the biggest jump next year. But just help us understand; what are the limitations around supply and your ability to upside next year?

I think overall, supply is very tight. There are multiple areas, not only wafers, advanced process node, HBM tight, packaging, substrate, and some of the components are very tight. We do have an excellent supply chain team. Operationally, we have been working with the whole supply chain to ensure we can support our top-line revenue growth. On the data center AI side, MI450, we have been preparing for the ramp for a long time.

So the way to think about it is we have been working with the whole ecosystem to ensure not only we have wafers, HBM, and also advanced packaging capacity, as well as all the different components to support the Helios rack-level solutions. That has been ongoing and now the demand continue to go up. So we absolutely needed to get more supplies. On the server CPU side, as Matt, and I talked about it's actually the demand start to accelerate this year. So we didn't need to catch up.

During the process of the first half, we have continued to increase the supply from wafers perspective, from advanced packaging capacity perspective, and we continue to invest aggressively. We talked about our CapEx increase. The primary increase for CapEx is to build the capacity to support server CPU ramp, like especially lines. As Matt talked about, we do see very significant ramp next year, and we are actually -- capacity is not enough.

We need to build ourselves to buy the equipment, do the consignment to support the ramp. So right now, we actually feel really good about all the supply chain work we have done to ensure we can support the data center revenue to double next year and also to ensure the embedded get the supplies to still to be very significant growth in 2027. PC and gaming, we prioritize, but we absolutely want to make sure we continue to gain share in the PC market, too. So we feel pretty good. But I would say we can have more supplies, demand continue to increase.

Let me pause there and see if there are any questions in the audience. If you have a question, please raise your hand. All right. Let's move on.

The CPU side, can you talk about maybe price? I think like Intel has been taking more price because they had older products. And just wanted to get your thoughts on price and how you guys see price evolving over the next couple of years?

Yes. I think first case, when we talk about our significant CPU growth, it increased both because of unit increase and the ASP increase. Unit increase probably is more than ASP increase. When you think about the ASP increase for us, the core count have been going up for each generation. So in general, that will drive the ASP increase. And your question into the future, when you really think about it is the way AMD, we think about how we work with our customers is we absolutely need to make sure we get the gross margin to continue to invest in the future.

But in general, if there's a component cost increase, we do need to make sure we share that cost increase with the customers. So for instance, if there's a wafer price increase, we absolutely want to make sure we share that with the customers. And if wafer price increase, you should expect us to increase PO price too. But in general, that is how we're operating. We really want to make sure we provide customers the best TCO and have long-term strategic relationship with them. And our gross margin really is important to us, but we are not going to increase the price just to expand the gross margin.

As scale-up domain gets bigger, at what point does copper stop being good enough? And when does that -- when does optics have to move inside the package?

Yes. Thank you for the question. We have not disclosed a ton about the road map in this area, but we have disclosed some. And at our Advancing AI conference a couple of months ago, we did talk about in the MI500 series program, which would be sort of second half of 2027 introduction and be kind of the primary product for the company in 2028, that we will have scale-up domains that are larger than we're offering now. So we haven't set a number, but greater than 72 and that we would have both copper-based and near-packaged optics-based options for scale-up connectivity.

We're still going to run our Infinity Fabric traffic, which is a coherency protocol that we use to communicate between CPUs and GPUs. So that technology was donated into the industry consortium, the UAL. And so we're still going to be doing traffic tunneled over both Ethernet and other means on copper and over optics. I don't think you'll see it be a point-in-time thing where everybody decides as an industry, okay, no more copper, now we're going to go optics.

It's not going to be a student body left type of decision. You're going to see the technologies run in parallel, and different versions of optics will be introduced with different risk tolerances over time. So I guess that's a long way of saying we're going to start on that path in the products in 2027, and we'll give you guys more updates as we go forward. But you should think about copper and optics living side by side for a number of generations. It's not going to be a binary shift. We have partners...

And especially what we're seeing in the private markets.

On the M&A environment, right, we have been doing bolt-on acquisitions, which really focus on the software capabilities we want to add. Of course, we acquired ZT System a while ago, which was to prepare the system level expertise to get the talent capabilities there. I think you should expect us to continue to focus on those areas to make sure we not only increase the capabilities of just building the system solutions, but also software side, the stack, the model to make sure we continue to invest for the AI.

Is there any scenario under which you could -- would consider using Intel as a foundry partner?

Thank you for the question. So I think as most of you know, right, we have a long strategic partnership with TSMC. We have been working with TSMC for a very long term, not only just from there as a supplier, but on the R&D side, if you look at a lot of our technology, we actually co-developed it with TSMC.

And for us, the most important thing when you look at our scale and the volume, the most important thing is we want to make sure the quality of the product, the advanced process technology, the 3D packaging, everything we are doing, it can be supported by our suppliers. We definitely -- we have a fab in Arizona. We're diversifying from geolocation perspective with them. But you should expect us to continue to view TSMC as our primary supplier on the wafer side.

I think we should -- you should view us as -- anything that's advanced packaging or anything that's advanced wafers, we're going to evaluate and due diligence on technology from every vendor in the market. But as you said that for the foreseeable future, I mean, the partnership we have with TSMC is going to be a significant one and will continue to be for a long time.

Think about your competitive positioning versus NVIDIA and proprietary ASICs. Where do you feel AMD has the clearest performance cost or availability advantage at this time?

I think what you're seeing right now in market is -- and Lisa and Vamsi and others at the company have talked about this very publicly is we feel like we have for -- in today's generation of product, tokens-per-dollar advantages for large-scale inference that we're going to be bringing to market and deploying in large volume with OpenAI, with Meta, and with Anthropic. And over time, and through multiple generations, we're expanding our training capability with the MI450 Series, we'll expand it further with generations beyond that.

But as you think about things in market today and what we're going to be ramping significantly over the next number of quarters, we feel like there's a -- our customers are great, great partners with us, but they do expect us to generate economic -- differentiated economic returns for them in terms of tokens per dollar, and that's what the market is demanding of us, and that's what we think the product delivers.

But when we talk about the entire breadth of the AI market, whether that be some level of customization, whether it be merchant computing, whether it be the server CPUs that run the agents, we intend to participate in all of that. And it might not look the same at every single customer. But I think we can bring at significant scale technologies across DPU, CPU, custom scale-up optics, and networking and system-level design across the board. So -- but where we differentiate the most today is in large-scale inference.

Great. We're almost out of time. Jean and Matt, thank you for coming to the Citi conference.

Thank you.

Thank you, guys.

Great. Good morning, everybody. I'm John Vinh with KeyBanc Capital Markets. I cover semis here. We're pleased to have AMD with us this morning and pleased to have Matt Ramsay, Corporate Vice President of Financial Strategy and Investor Relations. Welcome, Matt.

Thank you, John, and thank you for all your colleagues at KeyBanc hosting us. And I think we got -- I live in Atlanta, so we got a little bit of warm weather here too, but humidity is, I think, a factor of 6 below where I'm used to. So this is great. So thank you guys for having us.

Great. Maybe where we could start off our conversation, Matt, is server CPU sounds like it's on fire for you guys. I think you talked about 80% growth in the second half, 70% revenue growth next year. And you talked about having secured enough capacity to support that growth and potentially even upside to that number.

Maybe you can talk through what's been the primary constraint that you've had to work on to secure that sort of capacity. And then for the upside to the 70% number, I've got to imagine you've got a lot more demand than that. What needs to happen in order for you to be able to raise that number going forward?

Thank you for the question, John. And it is a remarkable time in the server CPU market. I know there was a couple of years where the server market maybe grew a little bit less than it had historically as CapEx quickly shifted towards AI systems. And with the -- we've always at AMD had the belief that server CPUs were going to -- and CPUs in general were of paramount importance across our business and have been investing in multiple generations of CPU architecture over a very long period of time.

We're just about to launch our 2-nanometer Venice CPUs. We're sampling them to everyone today. They're going to ship in our Helios AI racks and they're going to ship the family of Venice CPUs are going to ship broadly across all of our server markets. What's happened in the last 9 months is absolutely phenomenal in the server market. And I think maybe I can describe a little bit about that big picture, and then we can get to some of your supply chain-oriented questions.

What we've seen over the last 6 or 9 months is, I think many of this audience and many in the industry have been waiting to see when the dominance of AI spend around AI training of large models was eventually going to shift towards being much more heavily on inference. And I think we, at our Advancing AI event a few weeks ago, put out some models that we've done internally where the shift is happening right now towards inference being the majority of the AI computing spend.

What's happened on top of that shift, and that shift is happening right now is -- and this is maybe my terminology versus the company's terminology. But at the same time, the shift of spend is going from training to inference. Chatbot inference is becoming agentic inference at the same time, right? And that's a phenomenal thing for our business because we supply what we believe are very differentiated products for inference on the GPU side, given our memory footprint and bandwidth and also the best CPUs in the industry. So what you need when you're running agentic inference, you'll hand off the task to the big XPU or GPU cluster to actually run the intelligence of the inference.

And then in a very automated way, the agents will take the result of the prior inference, figure out what to do next and what to ask the inference and the AI model to do next, figure out where do I get the data to support the next step in the inference. Some of it comes from the cloud, some of it comes from enterprise systems, some of it comes from the web, some of it comes from wherever, reorganize the data and then hand it back off to the next step of the inference.

And you do that a whole slew of times in a very automatic way and you end up with a big agentic automated inference flow. And agents are nothing but simulated automated workers. And the computing that these workers and agents do is very diverse. I just described some of the pulling data from here there and everywhere to hand off to the next inference task. Sometimes it's running code that was just generated by the prior inference task. And that requires really high thread count, high-performance CPUs that can do many, many different tasks. And we're going to push at AMD to make sure that our Helios AI racks are doing as much of this inference computing as we can. But there's a market out there where inference is run on many different accelerators and all of those need agentic racks of CPUs.

And that's what we've seen in select markets. So we reported that we grew our server business more than 50% in the first quarter of this year. And many of you guys might remember when growing 15% to 20% in the server business was a phenomenal result. So we grew more than 50% in Q1. We grew more than 70% in Q2.

And interestingly, both our cloud and our enterprise business both grew more than 70% in the second quarter. To John's point, we've talked about growing greater than 80% in Q3 and Q4 in the back half of the year. And then a really early view of 2027 is on top of that much larger base, at least 70% growth next year. To the constraints that you were asking about, John, I think there's a few, right? We need -- we obviously need to wait for support from our great partner in TSMC, and Lisa and our supply chain team have been working with the folks at TSMC directly, and they've been absolutely phenomenal partners of giving us additional supply.

And we have to actually -- obviously, we need to earn it by delivering the products. But then there's advanced packaging. Venice is the first server product in the market to use advanced packaging, and we've invested Lisa was in Taiwan 1.5 months ago or so, and we announced a $10 billion ecosystem investment in the Taiwanese ecosystem. Much of that is oriented on back-end capacity. So I think we feel really good about where we are there. And then the industry obviously needs to have memory to support these servers as we sell them. So we're working really closely with all of our OEM and ODM partners and our hyperscale partners to make sure that they have matched set DRAM to support the server volume.

So I mean it's a pretty remarkable stat, right? The early view of 2027 server revenue for our company is roughly 20% larger than the whole server market was in 2025. So it's a pretty phenomenal thing that's happening, and we're going to continue to innovate. We talked about not just Venice, but the whole Florence lineup of CPUs that comes in 2028 in our event a couple of weeks ago. So we're going to continue to push across really high-performance single thread type SKUs to the highest core count and thread count CPUs for agentic racks and sort of everything in between.

But anyway, John, that was a long-winded version, but there's a lot going on in server. So I just wanted to give a little bit of the lay of the land.

Great. Just a quick follow-up there, Matt, is do you think there's opportunities for you to secure additional capacity through the rest of the year to maybe grow at a faster rate than that?

Well, we're certainly going to try. As you guys know, there's a lot of things in the supply chain that are tight. We happen to be very well positioned where depending on what quarter it is. We're the third or fourth largest customer at TSMC and been a very loyal partner with them for a long time. So we're -- one of the things that's been interesting is when Lisa speaks about this, the industry is quite good. If you give accurate forecasts and sufficient lead time, the industry is quite good at getting you supply. It's when you come and ask for stuff way underneath lead time where things get a little more complicated.

So we're doing our absolute best to continue to upside on supply in for the second half of the year. But I think we feel as we look into 2027 and into '28 that the industry, including ourselves and our partners have had much more time to adjust supply higher to support the growth. The nearer term has been where we've been having to work hard to get additional supply.

Great. Maybe switching to Helios. You talked about Helios going into mass production with maybe shipments starting in September. I'm just curious, can you talk about just the feedback you've gotten from your customers so far on Helios? What's kind of surprised them the most about it?

I guess I would start, John, you're absolutely right. We're going to start to ship MI450 and then CPUs and some of our Pensando networking products into our ODM partners that are building Helios starting in the month of September, and then we're going to have a fairly large ramp of that revenue in the fourth quarter, another fairly large jump of revenue in Q1.

And the business is going to be pretty phenomenal from a growth perspective. We talked about the server business growing more than 70%, but the data center business, inclusive of AI growing much more than 100% next year. So that will leave quite a lot of -- I don't know what much, much more than 100% is, but the math has to do that to get there for 2027.

The feedback from customers has been phenomenal. We've been working with the customers hand-in-hand on the designs and the spec of Helios for a very long time. And when you get to a product that is as complicated as this one going and launching full racks with our partners, the goal is no surprises.

And so when you get folks that are running full model code on sampled systems and getting ready to scale those systems, and the feedback from them is, wow, this thing really works the way that you told us it was going to work. That's the feedback that you want. Every day is a little bit of a different challenge as we're trying to get a product as complicated as this off the ground into significant scale I mean we're going to go from a standing start to billions of dollars of revenue in the first quarter that it's shipping for the full quarter, right? So it's quite a ramp.

And when you have that, there's something every day that happens, but the feedback from the customer base is phenomenal. We're very, very proud to have 3 of the premier model companies in OpenAI, Meta and Anthropic as our sort of workhorse customers for this generation of product, each of which want to go to gigawatt scale with our first rack scale solution. And I think that's a testament to the performance of the rack, the performance of AMD as a partner and also the performance of the software stack that's allowed them to get to that point.

Great. Maybe just to follow up on that, right? You talked about your 3 strategic partners, OpenAI, Anthropic, Meta. It seems like you've got multi-gigawatt commitments from all of them. It seems like the expectation there is they're each going to roughly deploy about 1 gigawatt next year. How do we think about that ramp next year? Is kind of a gigawatt per strategic partner, the right way to think about it? And then what about the rest of your nonstrategic -- you've got your core customers such as Oracle and Microsoft. It seems like they could also maybe account for another gigawatt of capacity.

Yes, John, I think it's -- we want to take this ramp in -- it's going to be a very fast ramp when you look at the revenue dollars, but we also want to take it in sort of a methodical approach because it is a complicated system. And the focus is on getting stable systems into market that our customers can run production code on as quickly as possible.

And so if you think about MI455, the MI450 Helios series being the primary driver of our [ AI ] business, that really starts in next month in September and probably runs through the first quarter or so of 2028, so a 6- or 7-quarter period of time, right? And so we've gotten commitments for 6 gigawatt arrangements, but 1 gigawatt commitments from both OpenAI and Meta. We've gotten a gigawatt commitment and a 2-gigawatt ambition from Anthropic over that generation.

And as you mentioned, at OCI, there are other OCI customers that are going to be running on Helios at Microsoft. There's Azure customers and Microsoft's own internal AI workloads that are going to be running on it. There's a number of Neoclouds that are there as well. So I think Lisa made a comment a couple of weeks ago that we would love to be -- there's certainly demand there to do what you described. It's now it's a matter of us executing and making sure we have land and power and shell and capital commitments for all the folks to actually deploy this stuff. We'd love to be able to do the first full gigawatt with Anthropic in 2027, whether we get all the way there or not, there's some variables there, but we're on a really exciting growth trajectory to much more than grow our AI business by -- it's going to double and a good bit more than that next year. Whether it's exactly the number of gigawatts per customer, I think that's a little too precise for today, but we're there's -- we have a significant ambition to ramp supply. The customer demand is quite strong across those 3 and others, and we'll see where we get.

That's great. There is, as you know, quite a bit of excitement from the investor community for you guys in terms of the Helios opportunity going forward. But there's a little bit of angst around this will be the first time you're going to go to rack scale. One of your peers, as you recall, when they first went to rack scale, there were quite a few growing pains. Maybe talk about what are you ensuring -- what are you doing to ensure that there's a kind of a smooth ramp here into the back half of the year?

I think the first order answer to that, John, is a lot of work. The second answer to that is, we've been taking a lot of feedback from the customer base over the last couple of years as we've been designing and getting ready to ramp Helios and setting up the ODM partners and the supply chain. Third, we did a significant acquisition of ZT Systems to bring in system-level talent into the company. And those folks have proven invaluable to hardening the design and derisking the different design points. There's a very large Gantt chart to ramp a complicated products such as this, as you might imagine, and every day is a little bit of a different battle. But we've tried to be methodical about it. We've tried to design the system such that we've taken risk out of the design. We've been -- we're going to be fairly focused on the initial ODM partners to ramp. We're not going to ramp everyone to massive scale all the same time.

We're going to have a couple of focused partners to start and then spread it out into the ODM ecosystem much wider as we go forward. And once we've gotten sufficient scale to sort of copy exact the success we have with the first couple of partners into a broader system. So I think there's a number of things where -- as I said, the first focus is to make sure not that we're just shipping racks, but we're shipping racks that are running code, production code for customers as quickly as possible.

And so that -- I guess the way that I would describe it right now is there's no smoking guns. We've gotten over that. We've shipped sampled racks, people are running code. They're very, very happy with it. We've proven that we can build racks. Now the question is the vast amount of blocking and tackling that we need to do to build the racks at the scale that we're talking about. That's the next step and where the team is focused.

Great. Any questions? Okay. At your Advancing AI event, I thought one of the most interesting announcements you made or comments that you guys have made is that on ROCm.ai that -- this is probably the biggest leap that you've made from a software perspective. Can you just unpack that a little bit and just talk about what you guys are most excited about there? It seems like you feel pretty confident you've been able to kind of close the gap with CUDA there?

No, I think it's a great observation. The team there with Bose and his leadership and Nushen the software team on ROCm have done, the progress that they made in the last 18 months has been phenomenal. And it's been accelerated significantly in the last 6 to 9 months of using AI tools in software development.

One of the things that we announced in our multi-gigawatt partnership with Anthropic that a lot of people focus on maybe the hardware pieces, but some of the software bits are just as important. I think we're not only using Claude across AMD's engineering teams, broadly, but we're working with Anthropic to make sure that any other customers that use Claude for their AI model work can automate and land on top of ROCm, on top of our Instinct platforms, their code that's automated by Claude.

And I think that's an important step. The Anthropic people were kind enough to tell the story on stage with Lisa so we can repeat it. But one of the things that they did going back in the really early part of this year is they actually rented a cluster of MI355s and got their premier inference model up and running on 355 and tuned in a weekend.

And I think that gives you some -- we can give you all the kind of stats about ROCm closing the gap with CUDA that you want. But the fact that a premier model company out of the gate without AMD's help or even our knowledge at that time can get up and run in a weekend. And that was a pretty phenomenal result and shows you where the software is now. Software is a battle every day, right? ROCm.ai is going to launch with Helios. There'll be another ROCm version that launches with the 500 series next year, and we're always in a refresh battle there.

But I think we -- for the largest customers that are spending the majority of the CapEx and are the most sophisticated in terms of their model work, we feel like we've taken the friction out of the system for ROCm to be a great place for them to do their work and as they ramp Helios and that gap has, I guess, narrowed to a point where it's not really a conversation with the top customers now. It's -- they know what they want to do in their application at their level, and they know that they need to run it through ROCm to give it to our hardware, and that's what we're optimizing for. And you can tell with the first-generation product that they're using from AMD on Instinct and our first rack scale product running software on top of ROCm that a company like Anthropic has ambitions to do up to 2 gigawatts with us in the first generation is a pretty good testament to where the software stack is.

Great. I thought one of the most interesting things you also talked about is kind of updated the server CPU TAM to $220 billion by 2030. I really like how you kind of broke out kind of the key kind of workloads within that stack with Agentic representing roughly about 2/3 of that TAM by 2030, which is pretty interesting. And I think you reiterated expectations of getting to 50% market share.

When I think about kind of the 2 camps of competition for you, it's your other x86 peer and then you got ARM if you look at the ARM results, it does look like they are gaining share, albeit off of a smaller base. Can you just talk about those 2 camps of competition and how you think about AMD kind of faring against those 2 camps?

Sure. I think the first thing I would say, John, like just philosophically in AMD, the discussion about what we want to do in the server market does not start with x86 versus ARM. It starts with go build the best server parts. And that -- I think that's the most important piece of this conversation is and do it with huge platform support and because of the chiplet architecture that we bring with a relatively small number of actually taped out chiplets, we can make a large, large number of optimization points and SKUs across the server business relatively easily.

So what we're seeing, you mentioned the way that we broke out the TAM we're seeing the, say, medium core count, really high frequency up to sort of 5 gigahertz with high bandwidth, emerge as a market for AI head nodes, head nodes for GPUs or XPUs, and that market be very distinct in its characteristics relative to the CPU only racks that run agents that we talked about earlier in the conversation where that demand is predominantly for our largest core count, largest thread count products that Venice goes up to 256 core and 512 threads. And it's how many agents can you run in a megawatt or how many agents can you run in a rack footprint?

And then we have the enterprise market and the cloud market in between traditional server workloads, like these agents will make a lot of calls to CRM systems or ERP systems or databases or whatnot. And those will run either in the cloud or on-prem, depending on -- that's really not a workload decision. That's a deployment decision. And the optimization points for those 3 buckets are very different, and we're seeing demand pull from customers for different SKUs to support those different buckets, right? If you think about one of the advantages that we've had as we've gained x86 share, I can remember when AMD's share was 0.4 and now it's in the high 40s of the x86 market.

One of the advantages that we've had is because we've had so much expertise on this chiplet technology. We've been able to push the SKUs and the core counts very rapidly versus our competitor. And I think we can continue to do that.

If we look at that versus some of the ARM competition, again, build the best CPU regardless of instruction set. And some of the x86 security and reliability, availability, serviceability features that we've hardened in our EPYC road map from servicing all the enterprises and all the hyperscalers over the last 5 or 6 generations. Those -- without being put through those paces, I think that's going to be difficult to replicate. And if you think about the importance of having security around running agent codes, I mean, an agent is what an autonomous worker with access to your enterprise data.

So the differentiation in RAS and security features and the ability to run all of the X86 enterprise workloads in addition to run the agent code, we feel strongly that -- not only are we going to be participating in a TAM that's much, much larger as you described, but the ambition to get to over 50% revenue share of that much larger TAM is certainly still there.

And the indications that we're getting from customers as we partner with them on what the road map looks like for the -- not just Venice, but the Florence generation, the Ravenna generation or the engagements there are really, really deep.

Last question for me is there's a little bit of debate about which architecture ARM or x86 is better optimized for Agentic workloads. I think you guys said that Venice was CPU that was built for Agentic. Obviously, you probably feel that your CPU is superior. What's the key metric that we should be paying attention to that you think suggests that maybe your server CPU is better optimized than ARM for Agentic workloads?

I think the first -- I'll make 2 points to start, John. One is agentic is not a workload. It's a very diverse set of workloads. There's not one design point or one SKU point or one optimization point that is going to be the right CPU for AI. I talked a little bit before about the big divergence that we're seeing in characteristics of head node CPUs and sort of traditional workload CPUs and what's going to happen in agents.

One of the metrics that keeps coming back to us on the agentic Rack piece is agents per megawatt or threads per megawatt. So where we're seeing the demand pull on the agentic side is for our highest SKUs and our highest core count SKUs. And so I don't know that it's an instruction set conversation. It's a capabilities conversation. We've been -- we -- as we ramp Venice, which is the first server part in the industry to use really advanced packaging and have really high core counts, that's where we're getting the demand pull in the agentic side.

So I think the first point is probably the most important, which is agentic AI from a CPU perspective, is not a monolithic workload. It's a very diverse set of workloads for which there are very many optimization points. And the way that we bring together our road map with the configurability that we have through chiplets to address all of those things with a large number of SKUs at scale, I think really does differentiate the road map.

Right. I think we're out of time. Thank you, Matt.

Yes. Thank you, John.

Thanks, everyone.

Greetings, and welcome to the AMD Second Quarter 2026 Conference Call. [Operator Instructions] And please note that this conference is being recorded. I will now turn the conference over to Matt Ramsay, VP, Financial Strategy and IR. Thank you, Matt. You may begin.

Thank you, and welcome to AMD's Second Quarter '26 Financial Results Conference Call. By now, you should have had the opportunity to review a copy of our earnings press release and the accompanying slides. If you have not had the chance to review these materials, they can be found on the Investor Relations page of amd.com. Today, we will refer primarily to non-GAAP financial measures during the call. The full non-GAAP reconciliations are available in today's press release and slides posted on our website.

As a reminder, our second quarter 2025 results included approximately $800 million of inventory and related charges associated with U.S. export control restrictions on MI308 shipments to China. Unless otherwise noted, comments making year-over-year comparisons exclude the impact of those charges to provide a more comparable and meaningful view of our underlying business performance today conference call are Dr. Lisa Su, our Chair and CEO and Jean Hu, our Executive Vice President, CFO and Treasurer. This is a live call and will be replayed via webcast on our website.

Before we begin, I would like to note that AMD will participate in the following events for the financial community. Key Bank's Technology Leadership Forum on Tuesday, August 11. Citi's 2026 Global TMT Conference on Tuesday, September 8, and the Goldman Sachs Communacopia and Technology Conference on Friday, September 11. Today's discussions contain forward-looking statements based on the current beliefs, assumptions, expectations including forward-looking statements regarding financial projections, business and industry trends that speak only as of today and as such, involve risks and uncertainties that could cause actual results to differ materially from our current expectations.

Please refer to the cautionary statement in our press release for more information on these factors that could cause actual results to differ materially.

With that, I will hand the call over to Lisa.

Thank you, Matt, and good afternoon to all those listening today. We delivered another outstanding quarter with record revenue and profitability as adoption of our leadership products continue to expand. Revenue increased 50% year-over-year to $11.5 billion, driven by significantly higher sales of EPYC, Instinct, Ryzen and embedded processors. Data center revenue more than doubled year-over-year and now represents 58% of total revenue, up from 42% a year ago, reflecting the rapidly expanding scale of our server and data center AI businesses.

Our record results mark another clear step up in AMD's financial performance and demonstrates the strength of our product portfolio and execution. We are still in the early stages of a multiyear AI adoption cycle as deployments grow across a broad set of markets and workloads. Driving demand for more compute and creating a clear path to significant revenue growth and earnings power in the years ahead.

Turning to our segments. Data center revenue grew 107% year-over-year to a record $6.7 billion, driven by strong demand for EPYC processors and Instinct accelerators. In server, we delivered our fifth consecutive quarter of record server CPU revenue with cloud and enterprise sales each growing more than 70% year-over-year, exceeding the outlook we provided last quarter. We gained x86 server revenue share year-over-year as customers expanded deployments of both fifth-gen EPYC turn and fourth gen EPYC general lease.

In cloud, hyperscalers continued expanding EPYC across their internal infrastructure and public cloud offerings, including AWS, Microsoft, full Oracle and others. Fifth-gen EPYC picture now powers nearly 1/3 of the more than 1,600 EPYC public cloud instance types available globally as providers broaden their offerings with new database storage and AI workloads. That expanding footprint is translating into growing adoption of EPYC in the cloud with health care, financial services, media and technology companies, adding tens of millions of instances last quarter.

In enterprise, we delivered record sales in our fourth consecutive quarter of record sell-through as on-prem adoption accelerated driven by the leadership performance and TCO advantages of our EPYC portfolio. Growth was broad-based as we won large deployments with leading financial services, manufacturing, telecom, retail, and technology companies.

More than 230 fifth-gen EPYC platforms are now in market from HPE, Dell, Lenovo, Super Micro and others, our broadest enterprise portfolio to date. Looking ahead, agentic AI is creating a new growth vector for server CPUs, spanning high-frequency AI host nodes, high-density agentic servers and general-purpose cloud and enterprise workloads.

Our sixth gen EPYC Ventas family is purpose-built for this expanding range of workloads and delivers 1 of the largest generational performance gains in EPYC history. Built on our all-new N6 core and 2-nanometer technology, Metis extends EPYC leadership in performance and efficiency, delivering more than twice performance per watt of leading x86 CPUs. And up to 3.3x the performance per watt of leading ARM-based CPUs.

The Venice family includes more than 30 processors that combine leadership per core and per-socket performance with a broad range of memory configurations, giving customers greater flexibility to optimize performance, efficiency, and TCO across the most widely used cloud, enterprise, and HPC workloads. Venice is in production now with every major OEM on track to launch platforms and the leading cloud providers planning deployments beginning later this year.

Customer demand for Venice is stronger than for any prior EPYC generation, and we expect to continue growing market share across cloud and enterprise in the coming quarters. Turning to our data center AI business. Revenue more than doubled over year driven by strong demand for Instinct accelerators. MI355x adoption continued to broaden as leading AI companies scale deployments across a growing range of inferencing and training workloads and cloud providers expanded MI350-series availability.

At our advancing AI event, we launched Helios, our rock scale AI platform combining EPYC Venice CPUs, MI450-Series GPUs, Pensando networking and ROCm Software. Across a broad range of inferencing workloads. Helios delivers up to 15% more throughput at the same rock power and up to 30% more tokens per dollar than the competition. Customer pull for Helios is very strong and tracking ahead of our initial forecast.

In addition to our multi-generation gigawatt scale deployments with OpenAI and Meta, we announced a new strategic partnership with Anthropic. Anthropic will deploy up to 2 gigawatts of MI450 series GPUs in Helios with deployment of the first gigawatt beginning in the first half of 2027. The partnership includes a multiyear joint engineering collaboration using Claude to optimize workloads for instinct GPUs and accelerate ROCm software development.

We also expanded our long-standing partnership with Microsoft. Microsoft will deploy Helios at scale on Azure for Frontier model inferencing across Microsoft, its AI customers and ad AI services. Together, these commitments broaden the group of leading AI companies and cloud providers building their next-generation infrastructure on AMD. Helios is now in production with initial shipments on track to begin later this quarter and ramp through the fourth quarter and into 2027 to meet very strong customer demand.

Looking beyond Helios, we plan to launch a new rack scale AI platform every year with each generation delivering significant performance, efficiency, and TCO gains. In 2027, our next-generation platform combines MI500 series GPUs, Verano CPUs and Pensando networking with expanded scale-up domains in both copper and optical-based interconnects. Customer engagement on MI500 is very strong with multiple customers working closely with us as they plan their next-generation AI infrastructure.

We expect MI500 to deliver the largest generational leap in instinct history is on track to increase inferencing performance more than 2,000 times in just 4 years. Turning to our AI software stack. ROCm has reached an important inflection point with the performance, capabilities and developer experience customers need to deploy AI in production at scale. The breadth of the ecosystem also continues to expand.

More than 3 million models now run out of the box on AMD, the leading open models launched with day 0 support for Instinct and open source contributions to ROCm have increased more than tenfold over the past year. We introduced ROCm, our new AI-assisted development platform for AMD GPUs last month. ROCm.ai-led developers use today's leading coating agents, including Claude, Codex, and Cursor to create, port and optimize code for instinct, making it significantly faster and easier to bring new models and workloads to AMD.

ROCm. AI delivers more than twice the trading performance and won 3x the inferencing performance of ROCm 7 across a broad range of models. We are also working closely with the leading AI labs, including OpenAI, Anthropic, Meta and others to co-optimize ROCm for their models with the improvements benefiting the entire AMD ecosystem. Taking a step back, the overall data center market opportunity is expanding far more rapidly than we projected just 6 months ago.

As AI moves into production across a broader range of applications and workloads, Demand for both accelerators and CPUs is growing well above our prior expectations. We now expect the data center AI accelerator market to grow more than 45% annually to approximately $1.4 trillion by 2030 and we expect the server CPU market to grow more than 50% annually to approximately $220 billion by 2030.

For AMD, this larger opportunity, combined with the strength of our portfolio and growing customer visibility is creating a steeper growth trajectory for our data center business. In data center AI, the growing number and scale of Helios and MI450 Series instinct deployments, position the business for significant growth in the second half of the year with growth accelerating in 2027.

In server CPUs with very strong customer demand and improved supply, we now expect server revenue to grow more than 8% year-over-year in the second half of 2026 and more than 70% for the full year 2027, off a much higher. Taken together, we now expect data center segment revenue to more than double year-over-year in 2027. Turning to client and gaming. Segment revenue grew 6% year-over-year to $3.8 billion.

In client, revenue increased 23% year-over-year to $3.1 billion, driven by record mobile processor revenue and continued share gains. Commercial adoption continued to expand in the quarter, with Ryzen Pro sales growing more than 50% year-over-year as we close new wins with large health care, technology, automotive, and financial services companies.

To build on this momentum, Dell, HP, Lenovo, Asus and others launched a broad portfolio of new commercial PCs powered by our latest generation Ryzen AI Pro or 100 series processors. Demand was also strong for our Ryzen AI halo developer systems, which went on sale in the quarter. In July, we introduced our next-generation Ryzen AI Halo platform powered by our new Gorgon Halo processor, featuring an industry-leading 192 gigabytes of unified memory and can run models with up to 300 billion putters.

And to make it even easier for developers to build and test large AI models locally, we are partnering with Hugging face to include 1 year of Hugging face Pro with every Ryzen AI halo system beginning later this year. Looking to the second half of the year, we are planning for a softer PC market as higher memory and component costs weigh on demand. Against this backdrop, we expect our client business to perform better than the market, driven by the strength of our Ryzen portfolio and growing commercial adoption.

In gaming, revenue declined 31% year-over-year to $779 million, primarily due to lower semi-custom sales at this stage of the console cycle. Gaming graphics revenue also declined year-over-year as higher industry-wide component costs contributed to higher graphics card prices and weighed on overall demand. Turning to our embedded segment. Revenue increased 19% year-over-year to $977 million, our strongest growth in more than 3 years.

Demand was broad-based with strength across networking, aerospace and defense, test measurement and emulation and communication customers. Our embedded X86 business grew significantly in the quarter as hyperscalers and networking customers increasingly adopted our CPUs to power critical networking and control plane functions in the data center. We also continued to expand our portfolio, introducing Ryzen AI embedded X100 processors for demanding real-time edge AI workloads and the Crea AI robotics platform for physical AI.

Looking more broadly, the strategy we have been executing over the last few years is now delivering strong results. Embedded x86 is becoming a significant growth driver for the segment. Our overall embedded portfolio is outgrowing the market and gaining share, and our embedded semi-custom engagements are expanding. Design win momentum also remains very strong. We are tracking towards another record year with more than $18 billion of new design wins, led by major wins with networking, data center, communications test, and aerospace, and defense customers.

In summary, we delivered record revenue and profitability in the second quarter, reflecting our strong execution and the growing adoption of our leadership products. We entered the second half with strong momentum across our businesses. With MI455x now in production, initial Helio shipments set to begin this quarter, Ryzen Pro CPUs driving continued commercial share gains and our embedded segment returning to strong year-over-year growth.

More than a decade of focused investment has given us the strongest and best product portfolio in the industry, deep strategic relationships with the companies driving the future of computing and a proven ability to deliver multigeneration road maps and ramp complex products at scale. At the same time, AI is driving demand for dramatically more compute across all of our markets.

We now see the overall market for high performance in AI computing growing approximately 40% annually over the next several years, approaching $2 trillion by 2030, and we expect to grow well above the market. As a result, we are tracking materially ahead of the long-term financial model we shared at our Financial Analyst Day last November. We now expect revenue to grow substantially above our prior target of greater than 35%, and we expect to significantly exceed our $20 annual EPS target within our strategic time frame.

We are still in the early innings of a multiyear AI adoption cycle and the opportunity ahead is enormous. We are exceptionally well positioned to capitalize on this opportunity and deliver significant growth in the coming years. Now I will turn the call over to Jean to provide additional color on our second quarter results. Jean?

Thank you, Lisa, and good afternoon, everyone. I'll start with a review of our second quarter financial results and then provide our current outlook for the third quarter of fiscal 2026, we had an outstanding second quarter, marking our sixth consecutive quarter of greater than 30% year-over-year revenue growth. Revenue increased 50% year-over-year and 13% sequentially to a record $11.5 billion, driven by continued momentum across our businesses. .

Importantly, data center accounted for approximately 58% of total revenue, underscoring the continued shift in our business mix and its increasing contribution to AMD's growth. On comparable basis, diluted earnings per share increased approximately 82% year-over-year, significantly outpacing our revenue growth and demonstrating our earnings power as we scale our business.

Gross margin for the quarter expanded to 56%, up over 200 basis points year-over-year and 80 basis points sequentially reflecting a favorable product mix and a growing contribution from our data center business. Operating expenses were $3.4 billion, an increase of 40% year-over-year as we continue to invest in R&D to expand our AI silicon systems and software capabilities to support our long-term growth opportunities. Operating income was $3.1 billion, representing a 27% operating margin.

Now turning to our reportable segment, starting with the data center segment. Revenue was a record $6.7 billion, more than doubling year-over-year and up 16% sequentially. Growth was driven by greater than 70% year-over-year increase in EPYC sales with record enterprise sell-through and robust demand across our cloud customer base. Both unit shipments and ASP increased significantly year-over-year reflecting the continued mix shift powered to our latest SM5 generation processors.

Instinct sales more than doubled year-over-year with the continued ramp of our MI350 Series product -- as adoption broadened across the largest AI labs, cloud providers, leading AI start-ups national labs and the sovereign AI deployment. Data center segment operating income was $2.1 billion or 31% of revenue. Client Gaming segment revenue was $3.8 billion, up 6% year-over-year and 7% sequentially.

The client business revenue was $3.1 billion, up 23% year-over-year and 6% sequentially, led by record mobile revenue. The gaming business revenue was $779 million, down 31% year-over-year, primarily due to lower semi customer revenue. Sequentially, gaming revenue increased 8%, driven by higher semi customer sales, partially offset by lower Radian shipments. Client gaming segment operating income was $582 million or 15% of revenue compared to $767 million or 21% a year ago.

Reflecting continued investment in go-to-market activities and expanding product road map. In base segment revenue was $977 million, up 19% year-over-year and 12% sequentially as demand continued to improve across end markets and the new design wins with [indiscernible]. Embedded segment, operating income was $386 million or 40% of revenue, compared to $275 million or 33% a year ago, driven by higher revenue and favorable product mix.

Turning to the balance sheet and the cash flow. During the quarter, we generated $2.4 billion in cash from continuing operations and $1.6 billion in free cash flow. Inventory increased sequentially to approximately $8.5 billion to support strong data center demand. At the end of the quarter, cash, cash equivalents and short-term investments were $13.1 billion. Now turning to our third quarter 2026 outlook. We expect revenue to be approximately $13 billion plus or minus $300 million.

At the middle point of our guidance, revenue is expected to be up 41% year-over-year, driven by very strong double-digit growth in our data center segment. Strong double-digit growth in our embedded segment and the decline in client gaming segment with growth in our client business more than offset by significant double-digit decline in gaming. Sequentially expect revenue to be up approximately 13%, driven by strong double-digit growth in both our data center and embedded segments and a modest decline in our client gaming segment with a slight growth in clients, offset by strong double-digit decline in gaming.

In addition, we expect third quarter non-GAAP gross margin to be approximately 56%. Non-GAAP operating expenses to be approximately $3.65 billion. non-GAAP other income and expense to be again approximately $55 million. Non-GAAP effective tax rate to be 13% and the diluted share count is expected to be approximately 1.66 billion shares. In closing, we delivered another outstanding quarter of revenue growth and the significant earnings expansion, reflecting the strength of our execution and the strong momentum across the business. As data center growth continue to accelerate, we entered the second half of the year, very well positioned to deliver continued profitable growth.

With that, I'll turn it back to Matt for the Q&A session.

Thank you very much, Jean. John, we'd like to go ahead and start the Q&A session now. Please do poll the audience for questions. We ask that each caller ask 1 question with 1 brief follow-up. Thank you.

Thank you, Matt. [Operator Instructions] And the first question comes from the line of Tom O'Malley with Barclays.

Lisa, you laid out the Analyst Day $1.4 trillion on the Accelerator side, you laid out the $220 billion on the CPU side. But the share dynamics within those are very different. So I think the genesis of the question here is when do you see that crossover point happening in the data center GPU business and the CPU business. Is that something that comes this calendar year? And when you look at the relative contributions between those 2 buckets, into calendar year '27 from a dollars perspective, where are you seeing the most strength?

Yes. Tom, thanks for the question. So look, we are certainly seeing a very, very strong compute market demand over both the data center accelerator as well as the server CPU. I think the server CPU is a little bit newer, and so we've been giving updates on that. As we go forward, certainly, looking at our Q3 guide our Q4 guide, we see very strong growth across both server and data center AI. So we're seeing server grow over 80% in the second half of the year year-over-year.

We're seeing very strong data center AI growth and that reflects into our 2027. The way to think about it is both businesses are going to grow a lot. The data center AI TAM is certainly larger. And as a result, as we ramp the large strategic customers that we have on Helios in 2027, you would expect substantial growth in that business in 2027. But both businesses are very significant drivers of our 2027 growth and beyond.

And then maybe under that framework with both growing very strongly into next year. Maybe this 1 is for you, Jean. You obviously have some headwinds in the -- on the GPU side as you're ramping Helios -- but also on the server processor side, you would imagine that you're getting above corporate gross margins there. Maybe as you look at the trade-offs to gross margins into next year, is 1 kind of offsetting the other? Or could you maybe think that directionally, just given some of the tailwinds you're seeing with pricing directionally with the mix shift to more CPUs, could you see a tailwind more so than a headwind into calendar year '27?

Yes, Tom, that's a great question. Thank you for that. I think as we talked before, our gross margin is primarily driven by the mix -- business mix. We're actually very pleased with our gross margin progress in the first half of 2026 and our guidance for Q3 2026. I think the way to think about the 2027 is there are a few puts and takes -- the first thing is, right, as Lisa just mentioned, the Altera business is expanding very significantly with the increased TAM and the pace of our business continue to grow, which from a gross margin perspective is accretive to our overall gross margin. .

On the other side, you are right, the pace of data center AI ramp is also very important because from a revenue opportunity perspective, we see large incremental revenue opportunity from data center AI once it ramps not only added tremendous revenue for the company, but also gross profit. The gross margin is slightly below copper average, I think the mix will determine how we go through 2027. But overall, we actually feel good about how we can really navigate through the transition of the business and continue to improve our gross margin to manage the balance of the gross margin plus our embedded business is recovering significantly. We actually think the Embedded business will give us additional tailwind on the gross margin side in 2027. So I feel really good about that.

And the next question comes from the line of Timothy Arcuri with UBS.

Lisa, I wonder if you could give us any color just what's embedded in the guidance in terms of data center. So -- the question really is like of the growth between Q2 and Q3, which will grow more on a dollar basis CPU or GPU? And I ask because if I take your up 80% half-on-half number, it sort of implies that the server CPU might not grow that much in Q4. So I'm not sure if I'm doing that wrong. Can you just give us some color there?

I'm just looking at that, Tim. Look, I think we believe that the server CPU is going to grow substantially in Q3 and Q4 -- the way I would describe it in terms of dynamics is we have been -- we have very strong demand across all of our customers in hyperscale and enterprise. And we have been adding additional supply as we've increased supply throughout the year. So we would -- we're guiding the Q3 segment to grow sequentially double digits the both server and the data center AI will grow nicely within that.

And then we should expect more growth for each of those businesses in Q4 with more supply coming on for server as well as for data center AI, the Helios ramp is just starting at the end of Q3, and it will be much more substantial in Q4, so that should sort of give you the picture of -- we think Q3 is certainly a strong quarter as we look at the strong -- very strong double-digit growth going into in data center segment. And in Q4, it will be higher than that.

Okay. And then just as a follow-up. So relative to the TAM that you laid out Analyst Day. I think it's a 40% CAGR out through 2030. So -- and do you think you can outgrow that CAGR out through 2030. So is your plan to grow your revenue more than 40% during that period?

Yes. So as we look at the overall TAM, first of all, it's a very exciting time in the compute market. When you look at every part of our business, whether it's server CPU, data center AI or our embedded business and our PC business we sum all benefiting from the AI tailwinds. And beneath that, I think in each 1 of the segments, we see an opportunity to grow above the market. So from that standpoint, with our view of the TAM being greater than 40%. We are also saying that we will grow greater than that from an overall revenue standpoint for the company. .

And the next question comes from the line of Vivek Arya with Bank of America Securities.

Lisa, I'm curious with all the announcements that you have made, how many gigawatt of compute does AMD have line of sight into for 2027 and what is AMD's monetization per gigawatt. I think the current assumption from everything you've announced is about 3 gigawatts and I think from some of the warrant numbers that you had mentioned before, right? It seems to be at least $15 billion per gigawatt. But so I'm just curious if you agree with those views and then what is the upside or downside to either the gigawatt or the content assumption?

Yes. So Vivek, there are lots of pieces to that question. So maybe let me take them 1 at a time. From an overall data center standpoint, I think we have given a little bit of color on 2027, we believe that the server CPU portion of our data center business will grow by over 70% year-over-year off of a higher base. And we believe the overall segment will grow by over 100%, so we'll more than double due to the data center AI ramping. So we do see a very significant ramp into 2027.

We are very happy with our strategic anchor customers in OpenAI, Meta, and Anthropic -- they are all in the process of ramping compute demand is high, and this is very much aligning their data center build-outs and their work with their cloud service providers together with our ramp -- the way I would describe it is we have supply to more than meet the guidance that we've talked about and the upside potential is there.

The key thing is to work very closely with our customers as they're planning their build-outs overall. So I think from a Helios ramp standpoint, we're expecting a very significant ramp over the next couple of quarters. We've said Q3 is the very beginning of the ramp. Q4 is a step-up, and then Q1 will be a further step up and we'll ramp as we go through 2027. So hopefully, that gives you a little bit of color on how that data center AI will play out.

And in terms of the revenue per gigawatt, I think we've said double-digit billions. We're still in that range. I think there's -- that looks like that's pretty much where it will be. And the key here will be continuing to work with our partners as they ramp because there's a demand for a lot more compute, and we would like to satisfy that demand.

And for my follow-up, a little of a technical question. When we look at the specs of Helios, it has, I think, almost 50% more HBM than the competition. And I think you have always had more HBM in your products. So how much of the benefit is because of having 50% more HBM and then doesn't that expose AMD more to memory cost inflation. So I'm just curious how are you ensuring that you're able to maintain margins, right, and get this memory allocation? And I think Jean already suggested you are comfortable with the gross margin range. But I'm curious how much of the benefit is just because of the use of more HBM? And does that still enable you to meet your profitability targets and allocation requirements over time?

So look, Vivek, we're working very, very closely with our memory partners across the board on both our GPU, HBM memory as well as just the general memory for the systems across our data center business. And what I would say is we've worked with our memory partners for multiple years to ensure that we have a strong ramp in our business. There is very good visibility into HBM allocation for what we expect to deliver in 2027. And the other piece of it is in terms of the memory bandwidth and the memory capacity, it is 1 of the advantages of the AMD solutions, when you think about the larger model sizes, they really benefit from the larger memory footprint, and that goes into the total cost of ownership.

We do recognize in this memory environment though, every customer is looking at how to optimize their memory footprint. And we have the opportunity, if desired to also modify that memory footprint if the total cost of ownership is not as significant in certain workloads. So think about it as -- the benefit of the memory is workload dependent. And we know a bunch of our customers are very, very happy with what that returns in terms of performance. But there are some workloads, let's call it, medium-sized models that may not get as much of a benefit. And in that case, we would address the memory footprint as you might expect.

The next question comes from the line of Joshua Buchalter with TD Cowen.

Congrats on the very strong results. I wanted to also follow up on the server CPU assumptions for 2027, so it sounds like the supply for that is locked up, and you have ability to service higher demand if it continues to track that way. But I guess any help you can give us on sort of the unit and ASP assumptions that are baked in there? Like I think investors are struggling to model the CPU market overall. So like how should we think about, I guess, core count growth or whatever is the right proxy as we think about modeling this business .

Yes. Sure, Josh. So let me maybe start with a little bit of grounding on the growth that we have seen so far. There are certainly both unit and ASP growth. So if you just look at our Q2 performance, when we said we grew over 70% in both cloud and enterprise. We actually had double-digit growth in both units and ASPs, but it was actually more unit. The unit growth was higher. That is very much the nature of our business. I mean we're seeing just very strong demand from an overall market standpoint.

On the ASP growth, we have ASP growth as we go to higher core counts or certainly, but as we go forward, you should expect both unit and ASP growth. And what we have been working on very diligently over the past couple of quarters is as soon as we saw the significant inflection point in server demand. We've been working across our supply chain. That's wafers, that's back-end capacity, that substrates, that's all of the components to raise the overall capacity for servers.

And we're seeing that play out through this year. That's 1 of the reasons we can raise the second half guidance. And we're seeing much more capacity coming in line in 2027 that supports the growth that we've been talking about. So you should think about both units and ASP in this framework.

And Lisa, in your prepared remarks, you mentioned that Helios was ahead of your original forecast -- can we unpack that comment a little bit? Was that a comment on volumes? Was it yields? And then if it yields -- how should we think about sort of the first quarter or 2 of Helios gross margins compared to as it gets later into its ramp, should we expect gross margins to improve as it ramps?

Yes. So Josh, when I was talking about Helios was ahead of our initial forecast, it was as it relates to the overall volumes. So the -- let's call it, the amount of demand there is for Helios in 2027. It's like an outstanding product. So what we're seeing from every one of our customers who's had a chance to not only spend time with Helios, but also spend time in our overall ecosystem. There is a high confidence that Helios will be a great addition to the AI portfolio, particularly around inference, and that was my comment about higher than our initial expectations.

As it relates to your comments about yields and performance and what do we expect as we go through the ramp. One would expect that the overall yields will improve as we go through the next few quarters, the starting quarter this quarter here in Q3, and we will be ramping over the next couple of quarters and we always would expect that the yields will continue to improve as we go through the first few quarters, especially on a product like this, which is highly complex.

And the next question comes from the line of Aaron Rakers with Wells Fargo.

As you can imagine, I'll stick with the server piece of the business as well. At your recent event, you highlighted the server market, the new $220 billion TAM that you're throwing out there as kind of in 3 buckets, right? -- general purpose, I think you called it or defined in the sandbox servers and then kind of the AI front end. -- node. So as we think about the growth that you're talking about, I'm wondering if you could help us conceptualize the sizing of those buckets at all? And how big is that kind of middle category, that sandbox AI market as we think about growing 80% plus this back half of the year and 70% plus, how big of that opportunity do you see that becoming.

Yes. So Aaron, what I would say is -- maybe let me start from the end points. So when you think about the $220 billion TAM in 2030 -- we actually see this agentic AI or these agentic sandboxes being the largest piece of the TAM. It's the fastest-growing piece. It's also the smallest piece today. So in the near term, I would say that there is certainly growth in that area. But as we go out over the next 3, 4, 5 years, we think that's the largest growth in the server market.

And I think the thing that is very strong about our portfolios. We believe we can grow in every one of those segments. So when you think about the types of CPUs you need for each of these workloads, you need sort of different optimization points. That's one of the things that we've really tried to point out. It's not just 1 CPU. It's actually a real family of CPUs that you need -- from our standpoint, Venice is absolutely leadership in all of the categories.

So whether you're talking about per core performance or you're talking about overall socket performance and that gives us a very strong position across general purpose agentic-AI and also the head nodes for the AI accelerator business. So from our standpoint, when we look at our customer traction in our customer momentum going from turn into its turn was already extremely widely used, especially amongst the hyperscalers what we're seeing is when we go into Venice, the workloads actually expand. So there are more workloads that are going to be run on the next generation of EPYC than are run on the previous generation. And that's what gives us the confidence to say that we can grow substantially ahead of the market given the product positioning.

That's very helpful. And as a quick follow-up. I know it's probably lower on the radar for a lot of people. But I think in the past, you've talked about the client business even with some pressures in the PC market growing this year for you guys on a year-over-year basis. Do you still see that? And are you willing to any kind of give any thoughts on what you might think of 2027 on the client CPU side?

Yes. Well, let me say, first of all, the client CPU business, we have continued to I think it's a very important business. I particularly think that local AI will continue become a larger and larger piece of how people experience AI. So I think AI PCs will become more important. As it relates to our client business in 2026, I think our first half performance has been very strong. And although we are expecting that the market will decline in the second half, the market has actually held up better than most people would have thought.

So our view is we will -- we see the growth in 2026 on a year-over-year basis certainly there. And as we go into 2027, I think we have a strong product portfolio that is coming on board to address not just the traditional notebook and desktop markets, but also as you think about a more AI-centric PC experience like what we have been talking about with the Ryzen AI Halo. So I remain optimistic about the PC market as an important way for us to reach a broader set of users and our ability to grow ahead of the market. The market itself will depend a bit on some of the components. We're all watching the component costs and how that will how that will play out over the next couple of quarters. But I think our portfolio and our rising content in enterprise are very positive for our client mix.

And the next question comes from the line of Stacy Rasgon with Bernstein Research.

First thing, I wanted to pick on something in some of the press release. It tells you that data center accelerates in the second half. So it grew 107% year-over-year in Q2, you're guiding it up like in the '80s, though, like in Q3. So how do I interpret which is not acceleration, how do I interpret that statement? Is like the second half collectively higher than 107% that you see in Q2? Or were you just sort of -- like how should I be interpreting what you guys wrote in the press release around data center accelerating as having house in second half.

Yes. Stacy, this is Jean. Good question. I think when we talk about the acceleration, we are talking about the second half versus the first half. If you look at the first half, our data center business year-over-year growth versus the second half, we do think there's an acceleration. That's the reflection there.

And then I think for my follow-up, I want to dig in a little more into the data center targets for next year. So you said like more -- I think servers more than 70%. The total more than doubling. I guess what I'm asking is how much work is the more in that statement doing? Like is -- like you seem to be on a trajectory that would take you like well above those numbers. I guess, just how should I be interpreting the more in the statement? understanding that we're only in August of '26, I get it. what is on your line there.

Stacey, you bring a smile to my face. So I would say, we're trying to give you a way to think about 2027. And yes, the server CPU we're seeing is more than 70%, which we think at this point is a very strong statement just given where the business is. And we do expect the overall data center business to be well over 100%. And the well over 100% comes because the data center AI business is going to be well over 100% just given the strength of our strategic customers, the ramp of Helios and all of the things that we've talked about. So hopefully, that answers your more questions. .

And the next question comes from the line of Jim Schneider with Goldman Sachs.

As you think about the early days of your MI400 series ramp over the next few quarters, can you maybe talk about the customer diversity you expect in the early stages. I think you've noted 4 or 5 large customers that you've announced publicly. Can you maybe talk about how many of the customers will be contributing, let's say, in Q4 and how many in the first half of '27.

Let's see. So Jim, I guess, the way to say it is we've talked about sort of the large frontier model companies, OpenAI, Anthropic, Meta they will be consuming through a number of CSPs both hyperscalers and others -- so the diversity of the business -- and there are additional customers. There are lots of customers who are interested in Helios at, let's call it, a more regular scale than gigawatt scale. So I think we'll have a good diversity of customers as we go through the next couple of quarters. .

And what we're really doing is matching to when the larger data center build-outs are ready. And so we are working with each of our customers on their data center plans and ensuring that we are meeting their data center plans. But yes, I think there's a good diversity, especially as we get into the Q4, Q1 time frame.

And maybe as a follow-up, you mentioned data center readiness. And so I wanted to sort of test how you're seeing your various customers and their ability to accept your products into their data centers, whether that's land power shell or anything else? Any constraints that you see hanging into the first half of '27 or even the back half of '27 that would give you pause about hitting your targets.

Well, Jim, I don't think we see anything that would give us pause about hitting the targets. I mean we feel very good about the targets. Now when you ask me what the range could be there is a range, and the range will depend on the ability to bring on more capacity in a timely fashion. So the way I view it is we are building sort of the entire supply chain. So ensuring that we have the entire -- our silicon CPU, GPU networking components that go into it, all of the Helios components, I think we feel very, very good about that part of the supply chain. .

We are working very closely on the data center operators and ensuring that we have good visibility into what's going on there. And that gives us, let's call it, strong confidence in what we've guided for the data center AI business so far. And we're going to be continuing to look at how to accelerate some of those bills. And I think there's a clear desire on the part of everyone in the ecosystem to bring on more AI compute faster. We are seeing every day more opportunities with operators to accelerate some of that capacity, and that's much of the work that we're doing together with our customers and partners.

And the next question comes from the line of C.J. Muse with Cantor.

If I take your data center guide, Lisa, it sort of implies instinct revenues of $30 billion, give or take. And so curious, 2 parts on this. How do you see kind of the revenue cadence first half, second half and then if we isolate the Instinct only and start thinking about 455X and Helios, how should we think about the underlying gross margins for that business starting in Q1 and then exiting in Q4?

I think, C.J. you're seeing $30 billion. Are you talking about the server? Or are you talking about which one are you -- you're talking about '27, right?

'27 instinct yes.

Yes. 2027. Maybe let me help CJ, I think what you're hearing from us is that your data center AI number is probably too low. And maybe back to Stacy's question without going into exact numbers, I think this notion of over 100% should consider be well over 100%. And look, as we look at this, it's a progression over the next couple of quarters. And as we go through the next couple of quarters, we expect to continue growing Helios quarter-by-quarter. But yes, that's the best way of thinking about it. And I'm sure Jean can work on if you have further questions on that. But Jean, does that yes, yes, yes.

And on the gross margin side, if we isolate to instinct, how should we think about beginning 2027 and then exiting.

Yes. I think overall, when we think about the gross margin for 2027, as I said earlier, is really determined by both the pace of server CPU business ramp and the data center AI ramp. So I actually think overall, quarter-over-quarter, it could be a different mix. It could change differently. But overall, when we think about this, we are actually very optimistic about how we manage the ramp of the MI450. At the same time, with the server business continue to improve and grow significantly in 2027.

Overall, I think they have a good offset. In general overall. But I do think it's important to remember we have multiple other levers from the company's factive not only embedded business, but also client on the business, we continue to improve gross margin and operation team continued to do a great job. So we'll give you more color when we get there to guide the 2027.

And the next question comes from the line of Joe Moore with Morgan Stanley.

Great. In the server CPU business, are you supply constrained now? It seems like the market is very tight. And do you anticipate that tightness persisting? And when you sort of think about next year, just the general ability of the supply chain to support the level of growth that you're talking about specifically in CPU.

Sure, Joe. I would say the server CPU supply chain is tight right now, and it has been tight for the first half of the year because much of this demand was unforecasted. As we get into 2027, the demand is better forecasted. And so we would expect that the 2027 server supply situation should be better than '26. We feel very good about being able to satisfy what we just talked about, which was over 70% year-over-year growth. And I believe depending on how things play out, there may be opportunities for that growth to go higher as we get through the next few quarters.

Okay. And is it -- you guys are ramping a lot on 2 nanometer, and there's been a lot of focus on 3-nanometer being in short supply. Is that helping? Or is it still challenging to bring up new capacity on the new note that way?

Yes. I think, Joe, it's always challenging to bring up new capacity on new node. I think what makes our approach a little bit special and different is that because we're using the chiplet technology we actually ramp in fewer wafers in the new node. And so that gives us the opportunity to Again, we're working very hard on ensuring we get the supply necessary to meet the very strong customer demand.

So from my perspective, I think all of that is work that's being done -- we're certainly looking at the overall supply chain, not just wafers, so that includes back-end capacity, packing capacity, substrates, all of those things. but we feel good about where we are to satisfy both the strong ramp in servers as well as the strong ramp in the data center AI business.

John, I think we have time for 1 more caller before we close out the call, please.

Our final question comes from the line of Atif Malik with Citi.

Lisa, you guys announced a partnership with [indiscernible] that your advanced and on this aggregate compute. Can you talk about just qualitatively, how do you expect the sales to grow in the fast inference market this year into next year?

Yes, absolutely. So I think the inference market overall is growing very substantially over this year into next year. The fast inference in particular, is an area which is, let's call it, starting to become more and more relevant. And so with our partnership with Cerebos, I think they have great technology together with Helios Plus their first scale engine, we get a very good solution for customers.

And we would expect that solution to be -- start becoming available in Q4 in the Cerebros crop cloud and extend into 2027. But from our view, this is an important part of the market, and we continue to look at ways to, I would say, customize and optimize our technologies for the various workloads out there. So we view this as just more of what we do in an open ecosystem.

Great. And one for Jean. How should we think about OpEx growth relative to that overall 40% market growth that you talked about in the next few years? .

Yes. I think Lisa talked about the TAM growth at 40%, and we'll be growing faster than -- from an OpEx perspective, we'll continue to invest given the large opportunities we have had for us. But you should expect us to manage this OpEx increase less than the top line revenue growth. That's what our business model is designed for so we can drive in more operating leverage to deliver earnings per share -- that is also consistent with Lisa said, our EPS will be significantly higher than $20. We outlined at our Financial Analyst Day.

Thank you very much for all the analysts and investors that joined our call today. John, you can go ahead and wrap up the call. Thank you.

Thank you. Ladies and gentlemen, that does conclude the question-and-answer session, and that also concludes today's teleconference. We thank you for your participation. You may disconnect your lines at this time.

Good morning.

Good morning.

That's a pretty good, good morning. I'm going to try one more good morning. And welcome to Advancing AI 2026. It's so great to be back here in San Francisco and to see so many friends and partners and customers and especially all the developers that are here with us today. And I want to say a big welcome to everyone who's joining us online from around the world.

This is my absolute favorite event of the year. It's where we bring the entire AI ecosystem together to show what we've been building and where we're going next. And this year, this is our biggest show ever because we have so much to tell you. So let's go ahead and get started.

At AMD, our mission is to push the boundaries of high performance in AI computing to help solve the world's most important challenges. And I often say AI is the most important technology of the last 50 years. And frankly, the progress that we're seeing in the industry is just incredible. Like every month, every few months, we see something new that was far beyond our imagination. And you can already see the impact across every industry.

In healthcare, AI is helping researchers identify new drug candidates faster. In science, we're solving problems that we didn't think were possible a few years ago. And across every industry, AI is changing the way we get work done. And the most important thing is we're still in the very, very early innings of what's possible.

Now just take a look at some of these charts. We look at these every year, and you kind of see the rate and pace of AI growth. A couple of years ago, we were just starting. People were experimenting with chatbots, and it was really cool. But if you look at today, we're really seeing more than 35 quadrillion tokens consumed every single month. That's an increase of nearly 160x in just 2 years. And that curve is actually just getting steeper. You're going to hear a lot about that today.

So just looking at where is all that demand. I mean, obviously, training remains incredibly important and foundational. And over the last 4 or 5 years, we see the compute that's used for training advanced models continuing to increase by, let's call it, roughly 5x every year. And those models are getting better. Like we're all experiencing that. We're seeing better reasoning. We're seeing new capabilities. We're now seeing a growing number of specialized models. And one of the things that I really believe in is there is no one perfect model. I think we're all going to use a slew of different models depending on what task you're trying to solve and which industry you're in.

But a bigger shift is actually in inference. And we expected this. We expected that inference would grow faster than training, and we can say that in 2026, for the first time, the world is using more AI compute to run models than to train them. And we've seen that to a point where we think this year, roughly 60% of the global AI compute capacity will be used for inference. And the reason for that is actually pretty simple.

As billions of people every day use AI, the workload shifts from building models to putting them to work. And actually, we're going to talk a lot today about agentic AI, and that's accelerating the shift even faster. So if you think about all of this, we see agents as the next big step for AI, and the most interesting thing is like this has really just started accelerating over the last, I would say, 5 or 6 months this year.

So when we started with LLM, they were great to answer questions, but we found that to really get the power of AI, you want agents to be able to really answer full questions. So you give it a goal. It keeps working until it solves the problem. And what that means for us in the compute industry, it means that compute demand is growing at an incredible pace. We're actually seeing a step change in compute demand. Because when you ask the agent to do something, it actually has dozens of steps, and it has to reason and it has to call tools and then it has to access data, and it has to keep doing it over and over until it solves the problem. And so you need lots of GPUs to do all that reasoning. But importantly, you also need a lot of CPUs to orchestrate every step around it. So when you look at just what that means from a market standpoint, it's really hard to put this market data up because I feel like every few months, we're changing sort of the perspective based on what our customers are saying.

But what we're seeing is that the shift to generative AI is certainly growing the AI accelerator market very significantly. It was actually last year at this event that we called the AI accelerator market at about $500 billion by 2028. And at that time, it actually felt like a very big number. Didn't you think it was a big number? But honestly, you look at it today, and it looks conservative because the AI demand is just continuing to accelerate. And you could see it, right? Better models, more usage, more usage requires more compute and more compute actually builds better models. And so we're now expecting that by 2030, the AI accelerator market is going to reach about $1.4 trillion by 2030. So what do you think about that? Is that a big number?

I mean what that means is by the end of the decade, the AI accelerator market is going to approach the size of the entire semiconductor market today. And although there will be many different types of accelerators, I'm a big believer and there's no one size fits all as it comes to chips. We do expect that GPUs are going to make up the vast majority of that market because the algorithms are still very much in their infancy, and we're still continuing to see the workloads change and that favors programmability in the overall silicon ecosystem. Now perhaps the most interesting part of the last 6 months has been GPUs are only part of the story. What we're actually seeing is Agentic AI is creating an entirely new growth vector for server CPUs.

Now I've updated this number a lot also over the last 6 to 8 months. But look, we call it like we see it. And we were early. I mean we saw from our largest hyperscale customers, and you're going to see some of them here today, who said, look, as AI inference is going up, we need more server CPUs. And so we saw growth in the server CPU market. And we said last year, we thought it might grow, let's call it, 18% to 20% CAGR to an overall market size of $60 billion. But frankly, what we're seeing is that the rate and pace of the agentic AI adoption is much, much faster than any of us thought.

So we're talking about agents going from millions to billions, like we're able to increase all of our productivity, and that requires a tremendous amount of CPU infrastructure. And every customer conversation is telling us that this build-out is just beginning. So based on what we're seeing today, and we're going to talk more about how agentic AI is evolving with CPUs, we now expect the server CPU market to grow by over 50% to over $200 billion by 2030. So that's starting from today, a $25 billion market going to over $200 billion. So there's a lot of excitement about CPUs as well. Now probably one of the things that is differentiating AMD and how we think about the market is this is not just a data center opportunity.

As AI becomes a larger part of our daily lives, we want the intelligence to run right where the work happens. And that means cloud is super important, but the devices that we use every day are going to be extremely important. And that means that we're going to do a lot more at the edge as well with autonomous machines that sense and act in real time. And we really believe that you need AI to be infused everywhere. And that's exactly what we're focused on at AMD. So putting all of that together, we actually expect that the market for our high-performance and adaptive computing products to grow at roughly a 40% CAGR over the next several years, approaching a $2 trillion market by 2030. And frankly, the only way you're going to be able to service a market like that is for us to work together as an ecosystem. There's no one company that can solve it all. But this is the opportunity to bring the best and brightest together.

And we have never been in a better position to lead. We have the broadest product portfolio. We have the strongest road maps we've ever had. And the thing that I'm most proud of is we have the deepest partnerships with the companies that are building this future. So a little bit about our strategy. Look, I think we've been very consistent. We've talked about our multiyear strategy, and it's really built around 3 priorities. The first is just compute leadership. We're building the broadest set of compute engines in the industry so that we have the right compute for the right workload.

Second, it's going to be about open platforms. We want everyone to come together in an open ecosystem. We believe an open ecosystem is essential to the future of AI, and that's how we get the force multiplier of everyone coming together. And that's open on hardware in terms of hardware standards as well as open on software. And that's why we're investing so heavily in our ROCm software. And you're going to see that AI has been a tremendous multiplier in this rate and pace of progress that we're making in ROCm and in software. And it really means that if we put all of this together, we can have our customers deploy on AMD hardware faster than ever.

And the third piece is we're going to get a chance to talk to you about powering AI everywhere. So that's AI in enterprise, that's AI at the PC, that's AI in the physical world. And that's adding new capabilities across all of our products. So today, you're going to see all of that in action. We're going to start with compute for the agentic era. I'm going to show you a lot of hardware. And then we're going to talk about the software and then our overall products from a platform standpoint. And we're honored to have some really special guests really going to join us to really help bring the technology to life.

So let's start with the data center. Today, AMD EPYC runs on the most important workloads in the world. We power the largest cloud providers. We power the digital platforms that billions of people use every day and the most important critical systems that are used by the largest businesses, including more than 60% of the Fortune 100 are using AMD EPYC, and that momentum is just building. Last quarter, we reached a record 46% revenue share in the server market, and we're continuing to see every major customer move more workloads to EPYC.

And on the GPU side, our Instinct adoption has also accelerated. We have a broad set of customers across the largest AI labs, cloud providers, leading AI start-ups as well as some of the national labs and sovereign AI opportunities are being built on AMD Instinct, and we really appreciate that opportunity. Now AI compute is becoming a lot more complicated. And what we're seeing with Frontier AI is it's really raising the bar for what infrastructure needs to do. It takes more than a single chip or a single server. You actually have to design the entire rack as one system. And that requires leading CPUs that requires leading GPUs that requires high-speed networking that connects everything inside the rack and across the data center. And just as importantly, this is about making these systems super easy to use. And so they need to be easy to deploy, easy to service and very reliable. And that is exactly what we built with Helios.

So today, I'm super excited to launch Helios, the industry's highest performance AI rack. Now Helios is built to train and run the most demanding frontier models in the world at massive scale. And I have a lot of show and tell for you. So for those of you who know me, you know that I love holding up chips. I call myself sometimes the Dana White of chips. But it turns out with Helios, there are lots and lots of components. And so you can see these are the chips inside the Helios rack. These are the Instinct, thank you. So this is Instinct EPYC and then the Pensando chips for networking that are inside each rack and every one of these chips is part of the system. So MI455 is the engine. It's the highest performance GPU in the industry. Then, what drives it is the world's fastest CPU course. And Pensando DPUs actually and NICs actually connect them with leadership programmability and scale-out bandwidth.

So I have a few more props to show you. Let's kind of take a look at how these things fit into the overall system. So first, this is now the MI455 accelerator, which is mounted on what we call an enhanced accelerator module or EAM. It's a production module that goes into Helios, and it is much more than the GPU package. The EAM actually integrates the GPU, memory, power delivery, high-speed interfaces, system management, cold plates for liquid cooling all into one compact serviceable module. Pretty cool, right?

Now if I just give you some of the specs, we're talking about 320 billion transistors. This is built with TSMC's leading 2-nanometer and 3-nanometer process technology. And most importantly, it brings together nearly a decade of AMD chiplet innovation. So it allows us to combine 12 compute and I/O chiplets with 432 gigabytes of memory, all connected by leading industry 3D chip stack packaging. And I can say for sure, it's the highest performance AI accelerator in the industry. Okay.

So this is the CPU board that's inside the Helios compute train. And I'm going to talk a lot about CPUs later. But just to give you a high-level view, it's high-speed 96 core EPYC processor, DDR5 memory, all the I/O that's needed to feed the GPUs, all in a single motherboard. And we also have our Salina DPU. This is a critical part of the networking infrastructure, and it really allows us to deliver both front end as well as the scale-out and scale across bandwidth overall. And this is our Vulcano AI NIC. And you can see this actually allows us to really have up to 6 Vulcano NICs on a board, and it uses the open Ultra Ethernet standard. So each Helios compute tray includes 2 of these Vulcano boards, and that allows us to have the scale out and scale the scale-up networking overall.

So you put all that inside the rack, and we have an additional 6 dedicated networking trays that handle the scale-up networking. And those are all connecting 75 GPUs together with UALink over Ethernet with silicon from our partners. And that's what we mean by an open ecosystem. So when you bring this all together, I want to say Helios is simply the best AI rack in the world. Now let's take a look at some numbers. Clearly, it's a very competitive world out there. So when you compare Helios to the competition, we're delivering 15% more compute, 50% more HBM4 memory capacity and memory bandwidth and 50% more scale-out bandwidth. And what that means is that every Helios can deliver more performance for the largest models, more capacity for longer context and the bandwidth to scale across thousands of racks.

And now we can kind of come over here and take a look at another very large piece of hardware. This is now the production hardware that customers are deploying, and you can see more of that as you go through the exhibit center. And each one of these weighs more than 160 pounds and stands less than 2 inches tall. That's just an extraordinary amount of technology when you look at what's in one of these trays. And when you bring these things together, the numbers are just incredible. More than 18,000 CDNA 5 GPU compute units, over 4,600 Zen 6 CPU cores and 31 terabytes of HBM4 memory, all in a single rack. That's what it takes to run agentic AI at scale.

So today, I'm excited to announce that Helios is in full production. We have shipments on track to start at the end of the third quarter and ramping into the fourth quarter and the second half of the year. And I can tell you, customer demand for Helios is extremely strong. And we're extremely proud of the work that we have done across the leading AI labs to adopt Helios. And so let's start with some of our guests. Today, I'm excited to be joined by our newest Instinct partner, and one of the leaders in Frontier AI. To share more about what we're doing together, please welcome Anthropic Co-Founder and Chief Compute Officer, Tom Brown.

Hello, Tom. It's so great to have you here, and it is really such an honor to be working with Anthropic. We just -- we had a big week this week.

We did. Yes.

So look, Anthropic has just had incredible success over the last few years. Claude has made such a big impact on the overall industry. Compute is such an important factor in the progress. Can you just talk a little bit about your compute strategy and just how things have been evolving?

Yes. Yes. First, thank you so much. t's a hard act to follow with you Lisa. So with Anthropic, as you mentioned, the scale that we're growing just as an industry is enormous. And so we have been working to make sure that we have the best like different chips, and we can use the best chips for the best workloads.

Yes. Look, I think you've been really leading in that area. And I'm happy to say that I've always wanted to be a smarter part of your infrastructure. I remember the early times we talked as you were starting Anthropic. But we have big announcements this week. We're honored that you will be deploying up to 2 gigawatts of Helios. Why now? What was the key reason?

Yes. So I think that -- like the first thing was just Helios is an amazing machine. It's like it's absolutely fantastic and then I think the thing that we were thinking about originally was whenever we're bringing out a new hardware platform, it's a big effort. It's like a huge thing. And so as part of the -- as we were thinking about this, we started doing our own evaluation of MI355, you guys generously got us a rack to start working. And we expected this to be kind of a big process.

Our actual experience was we had one engineer who started doing it. They spun up -- they connected it to Claude, asked it, hey, bring up this machine, left it going over the weekend. And we ended up with a graph of the actual performance of our leading model on it just going up and up and up over the weekend. And so yes I think that that's a testament to the open platform that you guys have made where like anyone like human or AI can now build real models on your platform.

Tom, I kid you not, when my team told me that you had an engineer working on MI355, I said, "Hey, like are we like making sure we help them?" And they're like, they say they don't need any help. They got it all covered. And I was like, hey, that's a wonderful story. Thank you for that. Now we're also super excited about the work we're doing together because one of the most valuable parts of this partnership is Claude and using Claude within AMD, like we're big believers in the fact that AI is a force multiplier. And the more we're able to have the leading foundational models really familiar with the AMD architecture, the better we're going to be able to service the overall market. So can you just talk a little bit about how you're seeing Claude really evolve in these types of highly engineering-specific workloads?

Yes. So I think this is a great place for collaboration where, in turn -- like the bread and butter of Claude is doing software engineering. And now more and more, we're seeing it help out with the type of workloads that you guys are doing also like design, layout, which is not quite the normal software engineering, but is adjacent. And I do think that, that is the place where we can work together to make the next generation of chips even better.

Yes. Tom, No, I think -- and also on the software side, I think the work that we've seen with Claude and the kernel development has been just incredible. So Tom, one of the things for our audience here, we like to talk about the present, but it's actually -- most of us are working on the future. So this is really the beginning of what we believe is a very strong multiyear partnership. Can you talk a little bit about, number one, what are you most excited about in the industry? And then number two, like what can we do more together to really make sure that we're satisfying those key opportunities?

Yes. Those are great questions. So I think the -- working together for the scale-up seems like it's probably the the biggest thing that I'm most excited for, where it does seem like we see consistently using more better compute results in better models that then add more value to all the folks using it downstream. So I think that's the biggest thing. And then I think one thing now that we're investing more and more with you guys to, and I think we can continue to is security, where that's a place where now as machines are getting more complex, we can work together to secure things at the chip layer, the server level, the rack, the whole network, and that will help out like not just us, but the entire industry stay safe.

No, I think you're completely right. We are super focused on the idea of how can we use AI to improve every aspect of our product whether it's performance, power, software and security, as you said. So really, Tom, huge thank you for the opportunity to work very closely with you and your team. Our team loves working with you guys, and we look forward to everything that we're going to do together.

Thank you so much.

All right. So look, Tom talked about how important performance is. And we have been laser-focused on ensuring that the real-world performance of MI455 and Helios really comes true when we look at system results. So let's take a look at some of those results.

So here as an example, we're running DeepSeek-V4-Flash, one of the latest reasoning models, and we are actually comparing MI455 to MI355. The key thing is, with every generation, we want to make huge leads in performance. So we're seeing that at lower concurrency that MI455 delivers 4x the throughput of MI355. But clearly, what you see with large users with -- across the board that you want to increase the number of simultaneous users, you want to increase the amount of work that you can do at the same time.

And so at the highest concurrency, we're seeing Helios deliver up to 34x more throughput than our prior generation. And what that means is more users, much faster response and much better efficiency at scale. And performance is just one aspect of it, right? The other aspect and what I should have mentioned when Tom was on stage is every one of us in the enterprise is seeing our AI budgets go up every single month. Is that right? Are you seeing some of that? And so customers really care about cost per token.

And with every generation of Instincts, we're driving the cost per token down. And that's with more memory, more bandwidth and much more compute. And we're taking another major step with MI455, delivering up to 18x more tokens per dollar, so customers can serve far more users with the same investment. And as you think about how that comes together in a rack, many of our data centers right now are really limited by power. So power is kind of the maximum limiter. And so we tested Helios over a full range of workloads where you have a fixed rack power compared to the competition. And what we're seeing is across the highest throughput workloads, the most interactive applications, all of these leading inference workloads, we're seeing Helios an average of 10% to 15% more performance than the competition. And that comes with all of the system work that we've done in the overall system. And how that translates to a customer is cost.

What we expect is the Helios rack delivers more performance, and it also delivers up to 30% more tokens per dollar than the competition. So I think that's a pretty good overall value proposition. Now I'm happy to say that the customer demand for Helios is extremely strong. From the largest AI labs to hyperscale and NeoCloud providers, we are working across the entire ecosystem to enable Helios together with our OEM and ODM partners. And one of our deepest and earliest partners deploying Helios is OpenAI.

To talk about where AI is headed and the work we're doing together, please welcome to the stage, OpenAI's Head of Infrastructure, Sachin Katti. Thank you, Sachin. It's so great to have you here. And most importantly, we are so thankful for the partnership with OpenAI. We've been through so much on this journey together.

You're really operating on the frontier of AI. Can you just tell us a little bit of what you're seeing and just what compute looks like for you?

Thanks. Great to be here. I was thinking about this event and over the last 2 years. It feels like the attendance at this event is following scaling laws I think every year over the last 2 years. But more seriously to your question, we've always predicted that models will evolve where they will start at chatbots, then they become reasoners, then they become agents, then they become interns. And the trajectory has kept up with that, right? So we are all seeing that models are becoming a lot more capable, more agentic and able to do very long-running tasks, let them go and they just take care of it. And so all that really means is we need to keep scaling compute, right?

The more compute we scale and throw at training, the more capabilities emerge. The more compute we scale and provide it to distributing intelligence to the world, the more people want to do with it. And this is what we see with exploding different budgets. We've seen internally, for example, in OpenAI that not just engineering, but every aspect of the enterprise is beginning to use agents for every part of the work. This is why we need platforms like yours that can scale with how quickly the capabilities of these models are scaling as well as how quickly the whole world is beginning to embrace what these things can do.

Well, Sachin, I don't think I've ever spoken to you where you haven't asked me for more compute. So he's actually very, very consistent. Look, our teams have been working really closely together. I think it's been a true journey when we think about just all of the technology that we've been doing. Last year, we announced a really landmark partnership where you're going to deploy up to 6 gigawatts of AMD infrastructure. We were one of the first -- I think they were the first to actually have MI455 racks a few months ago. So can you talk a little bit about the journey?

No, it's been a phenomenal partnership. As we mentioned, we were first to bet on AMD, and we are really thrilled with how that bed is turning out for us. Stack got started on MI300 and expanded very quickly to 355. And it's been really fun for our team to see how it was -- it was possible to optimize the software stack, the networking and everything and deploy our models running on AMD infrastructure.

So as Lisa mentioned, so we just got hands on the Helios racks 3 months ago, and the collaboration has gone to the next level. Our engineers are working side-by-side with AMD engineers to optimize the software stack and run GPT-Plus workloads on Helios already. Phil will probably be on stage later talking about how productive and fast that collaboration is. So we are really excited about the capabilities Helios is already showing us -- and we expect that we'll be deploying Helios at massive scale starting towards the end of this year and then accelerating throughout 2027.

Like I said, I know, more faster.

We need it earlier.

So Sachin, the other thing is one of the most exciting parts of our collaboration is actually some of the research work. So very thankful for the opportunity to spend really bringing our best engineers together with your researchers to talk about how AI should be really implemented and used for future both silicon and software. So can you talk a little bit about that work?

Yes. I think one of our key bets, and I think for a lot of us in AI is about recursion, right? AI recursively improving the systems it needs to run on and eventually, obviously, AI doing the research itself, right? Today, expert engineers spend an enormous amount of time optimizing kernels, figuring out how to translate new models to run them efficiently on different hardware, dealing with all the usual things you think about compilers, communications, libraries, system configurations. What we've been super excited and encouraged by the work we are doing with you is now AI is showing the potential and demonstrating it in production to automate a significant portion of the work when anyone programs to an AMD GPU.

So it dramatically has sharpened the path from a new model coming up to an efficient production deployment and something that we can much more easily tune to changing the workload requirements, right? And that's a big deal. So one of the other things that's super exciting for us is because AMD is built around an open software ecosystem, all these things that we are now inventing internally for our own use, we believe we can also bring it to the whole world and allow the whole world to use it for all of the AI models out there, not just our models.

I think super excited about that. I think the idea is that it's really the tide that lifts all boats, right? As we work -- it helps our work together, but it really helps the overall ecosystem. So a little bit of your crystal ball. Looking forward, I think we all know that compute is super critical. What do you need from the next-generation AI infrastructure? And what can we do as really a key partner to help you accomplish that mission?

If you haven't heard already, I need more compute more quicker. I think the thing that the whole industry is realizing is it's a systems problem at a data center scale. So I think we obviously very quickly internalized that AI was a rack level problem. But now it is very clear that it's not just a rack level problem, it's a rack in a data center. So the whole data center is a system. And we have to think about how we design the future based on what we know about our workloads, what's coming down the line, but codesign it with you from CPUs to GPUs to memory, networking, storage, power distribution in a data center and the cooling systems that go with it. These are all intertwined and they all have to be codesigned.

And one of the really fun parts about what we do with you and your team is how easy and productive it is for us to work with you and give you insights into how we expect the world to change in the future from what our workload needs are, what our models are and how quickly your team responds in turning those insights into better chips, better systems, better software. So that's worked well, really well with MI400 and left shifted a lot of the engineering work that we needed to do to bring this up online quickly. And that's why we are very confident we can productionize those systems very quickly. And so really looking forward to accelerating that with MI500 and beyond and increasingly using AI to actually do that work for us rather than being limited by humans.

That's fantastic, Sachin. And look, we really do recognize that input is so, so valuable. So I have to say thank you again for just the extraordinary partnership, the amount of work that our joint teams are doing together, and we couldn't be more excited about the opportunities in front of us. Thank you, Sachin.

Thank you very much.

All right. So now let's turn to the world of CPUs. There's been a lot of talk about CPUs recently. And I can say that this has been the foundation of our AMD data center strategy for the longest time. We launched EPYC in 2017. And we've been on a very, very clear mission. Every generation, more performance, more efficiency more capability for a wider range of workloads. Now Naples was our first generation. It established the foundation. But Rome and Milan really changed the economics in the data center. We did more than what people expected in terms of adding core count and adding throughput. And then Turin set -- Genoa expanded our performance and then Turin set a new bar for density, throughput and total cost of ownership. That's really been the EPYC formula, a clear road map and very consistent execution and leadership that actually grows with every generation.

This is why if you look today, 5th Gen EPYC Turin is the best server CPU in the world. With up to 192 cores and 384 threads, Turin delivers leadership performance across a broad range of cloud, enterprise and HPC workloads. But the thing about the data center is compute never stands still. With the agentic AI, it's actually creating a whole new class of infrastructure and the CPU matters now more than ever. And that's exactly what we built Venice for. So there's been a lot of talk about agentic AI, but it's actually a very new field. And what we see is that server computing is actually splitting across a couple of different workloads.

So the first section is GPU servers, and that's fairly well known. That's where the GPUs job -- the CPU's job is basically to drive the GPUs. And so the job there is all about speed. We want the highest frequency cores. We want the fastest I/O, and we want to keep the GPUs fully fed. Now the biggest middle part, this is the largest growth is actually in agent servers or what we call agent sandboxes. And this is actually a whole new class of workload. Here, you execute code, you actually call a bunch of tools, you actually query data that's outside the model. You have a whole bunch of things around it. And in this place, the priority is actually density. So we want the highest performing cores per watt to run thousands of agents at once. And then we have our traditional general purpose servers that run the enterprise.

And here, you're going to have a diversity of workloads. It's all about efficiency. It's running the applications, it's running databases, it's running data services. And all of that is what we would traditionally say is general purpose server. And so with these different workloads, you actually need the right CPU for the right workload. And EPYC is the only CPU portfolio that leads across all 3. So Venice is our newest and most advanced server CPU family, and it's actually designed for the agentic era. We're extending Turin's leadership across every single metric. That is performance, that's efficiency, that's TCO for cloud, that's enterprise efficiency, that's HPC workloads. And then it starts with our all-new Zen 6 core. We have higher IPC, higher frequency, and that delivers up to 1.8x more performance than Turin.

That is one of the largest generational gains. I mean we have been doing this for 6 generations, but this is one of the largest generational games in the history of EPYC. It's 203 billion transistors built on TSMC's newest 2-nanometer process, and it uses a next-generation chiplet that supports up to 512 threads per socket. And this is actually a really important point. I'm going to show you why in a few minutes. This gives the highest compute density, but it also allows us to double both the memory and the I/O bandwidth. So I have a few more chips to show you, if that's all right. And what's clear is Venice is not one chip. Venice is actually an entire family of chips. And this is where our chiplet architecture becomes a huge, huge advantage. It lets us take the Zen 6 architecture and build a whole portfolio around it.

So let's start with this guy. This is Venice HF. This is the highest-performing CPU that drives AI host nodes. It has 8 compute chiplets, each with 12 cores running at up to 5 gigahertz with the I/O and memory bandwidth to keep the GPUs completely fed. And this is a CPU that I showed you before shipping inside Helios.

This is its bigger brother. It's then Venice 256 core. This is the chip that's built for agentic sandboxes, and it also has 8 compute chiplets, but each of the chiplets now has 32 cores. So it scales all the way to 512 threads, and it is the highest compute density in the industry. Customers get more agent capacity per watt, per dollar, per rack bar none. Okay.

And then one more member of the family. This one is Venice 128 core, and this guy is optimized for enterprise and general purpose servers. This is the same leadership performance per core, but it's in a lower cost, lower power design for enterprise applications. And this one is really tuned to give customers the best performance per dollar across a wide range of configurations. So you can see really the power of the family as they come together. And in addition with these guys, I can tell you that next year, we're adding a few other chips to the family. We're adding Verano, which is the next generation for AI host nodes. And what that adds is a more power-efficient low-power memory as well as faster interconnect between the CPU and the GPU.

And then we're also adding for our high-performance computing or technical computing, Venice X, which is using our 3D V-Cache stacking technology. And we're really the first to stack memory chiplets right below the compute chiplets to accelerate performance even for those types of workloads. So you can see one architecture, a CPU is not a CPU. There are many different types of CPUs. And for us, one architecture spans dozens of different chips. And that breadth is someone that no other CPU vendor can match. And that's what makes 6th Gen EPYC the best CPU in the data center.

Now you're going to bear with me because I'm going to take you through some performance because the numbers are just incredible. So looking at AI workloads, right? Again, you have different use cases and different workloads, starting first with CPU servers. Many of the largest frontier models require you to move data between the CPU and the GPU. And in that scenario, Venice moves the data faster than the best competitive x86 processor, delivering up to 1.8x more tokens per second. And when you look at the agentic CPU servers and sandboxes, because of our density, Venice delivers more than twice the agents per watt for agent sandboxes.

And for general purpose applications, in this case, we're standardizing on 100-kilowatt rack, Venice delivers more than twice the performance per watt. So pretty incredible results. And the gap is even wider if you look at ARM processors. So if you look at the leading ARM processors, when we're comparing at the chip level, Venice supports up to 2.8x more agents per watt. And when you look at the rack level, Venice is delivering up to 3.3x more performance per watt. And what that means is at the data center scale, you can just get a lot more agents in the same power envelope.

Now there's been a lot of talk about what matters most for agentic AI with some saying that per core performance under load is the only thing that counts. But that's really only part of the story because agentic AI actually runs as a distributed platform, and that is across databases and vector stores and orchestration. And all of that has to happen at once. And that requires not just raw speed, but you absolutely need density, efficiency and per core performance all together. And what I'm happy to say is that Venice leads across every one of them.

So when you compare Venice against the highest performing ARM CPU from our competition, EPYC delivers 20% higher per core performance. And when we I like that number. I like that number. And when we pair that with leadership core density, we're delivering 2.2x more performance at the socket level. And so the main point is whether a customer needs per core or per socket performance, Venice is the leader. And there's one more point that's a benchmark doesn't really capture because when you think about what agentic AI is, it's really like you're adding thousands of employees to your enterprise, to your sort of traditional workflow. And that software enterprises have been running on for years. The truth is that software runs on x86.

So Venice runs all of it. And this is where x86 has an advantage. You must have the best performance, you must have the best density, but having that software compatibility is a big, big plus. So customers scale up their agents on the stack that they already have. And now let me just show you what that means at the rack level. So at the rack level, what we find is density is super important. Everyone is trying to optimize their data center, everyone is trying to optimize their power envelope. And with Venice, every major server OEM is offering a broad spectrum of racks for agentic AI.

So what that means is we have the full spectrum from 25,000 cores to 50,000 cores, and that's more agents per rack, and that gives customers the choice. The key is every data center is different. And with that, the flexibility allows you to pick the right cooling, the right space requirements for what you're trying to do in your data center. So there you have it. Venice, the best server CPU in the industry. And I'm very happy to say also today that Venice is in full production.

Customer demand is incredible. It's the strongest we've ever seen for a new EPYC generation. We're seeing every major server OEM, every major cloud provider on track to begin rolling out in the fourth quarter as we start with making the broadest EPYC launch we've ever had. So with that, I want to turn to my next guest, who runs some of the largest and most advanced compute infrastructure in the world. And it's really been our privilege to partner with them as they've deployed multiple generations of both EPYC and Instinct.

To share more about our work together, please welcome to the stage, Meta's Head of Infrastructure, Santosh Janardhan.

It's so great to have you here. Thank you for being here. Santosh is a true friend. I have to say it's been a tremendous opportunity over the last few years. My story about Santosh is the first time I sat in his office, he said to me, Lisa, we're going to deploy a lot of CPUs. Please make sure they work. And I said, I will. I will.

And you lived it up.

Meta operates -- you operate some of the most advanced infrastructure in the world. What's happening in the data center world? What's happening in your world? Tell us a little bit about the architecture and what you're working on.

Sure. First of all, thank you. It's awesome. This is like a hardware head's paradise, right? You talk about transistors, you're talking about hardware, you're talking about cooling. This is sort of the jam. I like it, right? And I think this is an audience that's receptive to it. I usually talk to audiences, they don't -- they have usually either not been to a data center or not really seen a chip ever. So this is refreshing, I have to say.

The thing I'll say about, listen, we are seeing sort of demand go through the roof. If you look at inference, training, recommendation system, that's how a News Feed works. Content creation, the demand for that is just exponential right now, right? And when I think about sort of the overall, Mark has this vision of delivering personal super intelligence to billions of people. So if you go to sort of Facebook or Instagram or WhatsApp or whatever surface you go to, the idea is that we'll meet you there and we'll deliver super personalized intelligence, right, wherever you are. And that's why we established this meta compute initiative and because it's a top level thing Mark himself oversees.

Now what it means is that we are now moving away from an area where we used to take servers, optimize it, right? This is what the conversations we were having many years ago that, hey, I'm deploying a CPU. I really need to maximize performance out of it. It's different now. You talked about how we are -- we should be thinking about the whole system end-to-end. We are at a point that we have to look about the whole data center and think about it as one integrated system, servers, hardware, networking, cooling, power, all of that is not optional.

All of that has to work together, right? And this is where I think there's a huge opportunity for collaboration with partners like AMD because you have to now codesign, co-create systems, not just deploy sort of something that comes off the shelf, right? The other thing is, by the way, flexibility. The thing I really like about what you said is that this is such a big opportunity in the industry right now. All of us need to lean in and sort of work on this together. So it needs to be open. It needs to be heterogeneous. And I think there's no one company, one partner that works for any one of us. We need to sort of work on this together. And that's why I think you're one of our most important strategic partners.

Thank you, Santosh. And look, we completely agree with your philosophy. Now you are actually one of our broadest partners because if you think about our work on CPUs, GPUs, we developed the rack-scale OCP systems together. You've deployed lots and lots of CPUs, so millions of CPUs and you're a lead partner on Venice. Can you just talk a little bit about sort of your CPU infrastructure and some of our work together?

Sure. So many years. And the story that Lisa was saying was many years ago. I think we went from Milan to Bergamo to Turin, now to Venice. So it's -- at least the fourth generation. So long-time partnership, obviously. I actually think the world is changing in the sense that it used -- the world is -- every year is the new world these days.

Like every month.

Months exactly. So what ends up happening now is that it's not just a GPU game anymore. It's CPUs and GPUs. If anything, I think CPUs have become at least as important, if not more. You're looking at a world where sort of the workforce -- the workload is changing because while you have your workforces and GPUs, at the end of the day, there's agentic workloads. You have to run your tools, you have to run your systems. And all the code that's been generated still needs to run somewhere, right?

So I'm super excited to work on. I think there's a lot to come there. And you have to think, again, this is a theme that I'm assuming we'll go throughout the presentation that you have to start thinking about CPUs and GPUs as conjoint things, you hand off workloads depending on the workload, you employ the hardware.

Yes. So absolutely right. And we've gotten a lot of feedback from your technical team. I think that's what I really appreciate about the partnership with Meta. Now clearly, they're deploying lots of GPUs, too, a lot of accelerators. And you're also one of our deepest partners on the accelerator side, starting with MI300, and then we had a very large strategic partnership announced around up to 6 gigawatts starting with MI450. And what we're doing is quite unique with Meta. So can you talk a little bit about sort of the evolution and the MI450 plans?

Again, multiyear collaboration. I think we started with MI300. We did a bunch -- it has really nice memory capacity. I remember talking to you about that, right? And then there's a bunch of pre-training we did. It is important to sort of deploy that at production scale, sort of get hands-on experience on both sides. 350 is the first time we deployed this on our ranking and recommendation systems.

So that's now getting into sort of some good numbers. And 450 is, I think I'm super excited about it. We just got some of the racks. And we are going to deploy it across the board because it gives us the opportunity to go and sort of really collaborate. I think the difference -- big difference between sort of the 300, 350 and 450 is that we have had engineers sitting in the same room talking to each other, co-designing, co-creating, like power cooling. It's like I say, it's not just a single thing. And we are not shy in feedback like...

You are not shy, for sure.

But you're very receptive. So this is a true partnership, right? This is the point about co-designing and co-creating that, I think. And like I say, that's a huge sort of deployment coming in. The more we deploy, the earlier we codesign, the better we are.

Yes, absolutely. Really appreciate the effort together on MI450. Now I want to ask you also about your crystal ball. So you look out into the future and you see what the next few years really means as you push forward in the AI industry. What are the biggest challenges and really opportunities for us? This is the industry ecosystem here. So what should we be focused on as an industry?

Listen, all of you know about the bottlenecks. You know about the power, the data centers, the silicon, like all of these, I think, are choke points that I think the industry is waiting on. But those are things I'm pretty confident will sort of work our way through. At the end of the day, I think the demand is immense, people are responding. But the thing that I really want to make sure all of us realize is that we should not think about systems in isolation.

Like I was saying, CPUs and GPUs are one conjoined system. You should start thinking about it that way. You should -- we really need to codesign things early. See, data centers takes years to build, okay? Like one of my favorite stories Mark comes to me and says, hey, I want a gigawatt worth of data centers. You should have talked to me 2 years ago. It takes time. It's the same with silicon, right? It just takes time.

So we need to be starting and sitting down in the room, codesigning today for what we need to deploy in '27 and '28. That, I think, is when we truly unlock sort of the power of the system. The systems are amazing. The specs are amazing, right? But think about how much more you can get out of it if you sit in a room and just go design it today. And then in '28, we'll be having much better graphs out there.

Fantastic. Santosh, I completely agree with you. I want to say again, thank you. We are so, so happy with the amazing partnership that we have across the board and most importantly, with your engineering teams, and we truly are excited about what we're going to do in the future. Thank you so much.

Thanks.

So you heard Santosh talk about just the diversity of workloads and really the massive scale that you require with AI. Now I want to turn to another part of the inference market where the requirements are actually very different. So lots and lots of workloads. And as inference is moving into more products and services, we're actually seeing it segment into different workloads depending on what you're trying to do. Some of these are, let's call it, less interactive. So they're really tuned for, let's call it, maximum throughput or the lowest cost.

Other of these applications require more balanced throughput, so you have to balance throughput and responsiveness. And then there's this new class of applications that actually want very, very fast results. You have extremely smart engineers and they don't like to wait. And that is where ultra-low latency or every millisecond actually matters. Each one of these types of workloads requires a different type of compute. And one of the best ways to reach ultra-low latency today is disaggregated inference. So if you think about the different pieces, both prefill and decode are 2 different jobs.

So prefill tends to need more compute, decode tends to need more memory bandwidth. And with Instinct, we have a very balanced machine so that we deliver both great compute and great memory bandwidth. But you can actually take this a step further. If you know what workloads you're trying to run, you can actually let customers tune each of these pieces independently. And to address this market opportunity, we've been working with Cerebras.

So to talk more about what we're building together, please welcome to the stage, Cerebras Co-Founder and CEO, Andrew Feldman.

Andrew, it's great to have you here. It's been an exciting few months for you. I know that for those who don't know exactly what you've been working on, can you talk a little bit about Cerebras and what problems you've been trying to solve?

Sure. Great to be here and great to be sort of among people who love hardware. It's nice. Look, I'm thrilled to be here with you and to announce our cool new partnership. At Cerebras, we build the world's largest and fastest chip. It's a full wafer. And we package this into a system and [indiscernible] systems, right, into racks and then racks into clusters. And we deliver them both on-premise and via the cloud. And you've had some of our customers up here already to date. On the frontier labs, we serve customers like OpenAI and the hyperscalers like AWS, in the small agent in the coding space, leaders like Cognition. And they do this because we're blisteringly fast.

There's no question, Andrew, that you have some tremendous innovation with what you've been working on. And our teams have been collaborating really closely over the last few years to really have Helios together with the Wafer-Scale Engine for ultra-low latency inference. Can you just to educate the audience a little bit? What are we trying to do? And what does that mean for customers?

Sure. I think what's happened, and you described it previously is that sort of AI has moved from being a novelty to being useful and then in some domains a necessity. And when something is a necessity, people want to use it and they want to use it quickly. And to serve this market, this segment of ultra-low latency, we sought a partnership that could extend our capabilities and our footprint, and there was no better answer than AMD. We were already an AMD customer as we use AMD CPUs to surround our systems. And we were so excited at the opportunity to build a disaggregated solution that combined AMD CPUs the cool new Helios Rack and the Cerebras wafer scale engine.

Yes. Look, the technology is pretty cool. So just talk a little bit about how it comes together.

Sure. So as Lisa described, what you have with Instinct and the Helios rack is you have the real leader in performance and memory capacity. And you marry that with our Wafer-Scale Engine, which is the leader in sort of in SRAM and memory bandwidth. And that combination allows us to deliver a solution that is unmatched in the industry.

Very, very exciting. I know that customers are excited about what we're going to do together as well. So when are we going to have it in market?

Later this year, and it will be first in the Cerebras Cloud and it will be later at a store near you.

I think the key thing is we're going to give customers the choice to really put together what is the solution that they want. And I think we're very, very happy to be able to do this together with you. I think it's a huge step up in terms of what we can do for this ultra-low latency segment.

I think that's right. I think until recently, customers sort of -- they could have sort of high throughput or they could have extraordinary speed. And by bringing together the Helios rack with the Cerebras wafer scale engine, we give you 5x the throughput while continuing to deliver this extraordinary speed. It's really something amazing.

Fantastic. Well, look, Andrew, huge congratulations to what you and the Cerebras team have done. I mean I think it's been incredible. I think you've been clear on what you're trying to accomplish. And it's really nice to see not only it come together, but us come together to offer a solution that will be very, very compelling to customers. So we can't wait to get this into the customers' hands.

Soon. Thank you.

Thank you, Andrew. All right. So look, I think we've shown you a lot of hardware this morning. But as we all know, hardware is only part of the story. It's actually software that turns all of this into a platform that developers can really build on. So to share the progress we're making with ROCm, please welcome AMD's Senior Vice President of AI, Vamsi Boppana to the stage.

Thank you, Lisa. Good morning, everyone. It's great to be back to talk about software. Just a few years ago, programming AMD GPUs required deep expertise and significant engineering effort. We've come a long, long way in a remarkably short period of time. Through our open source community collaboration and sustained investment in ROCm, our software stack, AMD Instinct GPUs are powering some of the most important and consequential workloads on the planet. But the biggest change is still ahead of us. AI is transforming how software is getting built. And we are bringing that transformation to ROCm. And that's what I'm excited to share with you guys today. Our software teams have been moving fast with relentless focus on developers. We've invested at every level of the stack and build and test infrastructure, so we ship faster. ROCm releases now go out every 6 weeks, not every 4 months. We've continued to expand the work we do with our AI ecosystem partners and the result, more features, faster performance and a significantly better out-of-the-box experience. Our strategy that has gotten us here has been remarkably consistent. It's built on 2 core principles: partner deeply with the open source ecosystem, and build with the right layers of extraction to enable developer productivity. Open Source gives us velocity and scale and abstraction makes developers more productive. Our deep commitment to Open Source has resonated with the community that has truly embraced AMD, from frameworks and compiler stack to inference engines and model hubs. AMD is now becoming part of default enablement for the most important AI communities, such as Hugging Face, PyTorch, JAX, vLLM and SGLang. This is why new models get day 0 support and why the world's most important AI workloads run on AMD today. That is a big shift. Inspired by engineers seeking productivity, abstractions have evolved to enable work at the right level of detail, hiding complexity when they want speed and exposing control when they need performance from low-level programming in assembly and C to block programming abstractions like Triton to Pythonic frameworks, we've invested in key extractions.

AITER, our kernel library and ATOM, our serving engine get your peak performance without writing the kernels yourself. FLYDSL is a brand-new Pythonic domain-specific language that gives you low level control with the performance of hand-tuned C++.

And Mochi, our communications library helps deliver leadership performance on the latest models like MiniMax, but something very, very exciting is happening. Over the past year, I've seen something remarkable inside AMD. Our engineers are using AI models to generate GPU kernels, optimize code, debug issues and improve performance. And in some cases, these AI-generated kernels are shockingly good, better than what we expected, sometimes better than the most manually tuned versions. The first time you see this, you're a little bit skeptical. You run more tests, you try to break it, you look for what went wrong, and then you realize this is real. What has happened in general software development is coming to significantly transform GPU programming. It's going to reduce the time it takes to bring up workloads, make optimization automated and it will remove any remaining barriers to broad adoption. We want to put that capability in the hands of every developer. So today, I am excited to introduce ROCm.AI. It's an agentic AI platform that brings the capabilities of AI assisted GPU programming to developers. It starts with a simple idea of giving developers access to the power of an AMD ecosystem through the AI coding agents, they already use, whether you're using Cursor, Claude, Codex, ROCm.AI helps those agents understand AMD platforms, understand ROCm and helps you build and optimize your workloads. In other words, we are making those popular coding agents into ROCm super users. You should be able to describe the workload you want to run, the performance target you want to hit and then let the agents help you get there.

For years, we have been building the open software foundation. Now we are adding an AI-assisted layer that helps developers use that foundation faster and more effectively. So let me show you what is inside ROCm.Ai. It's built on the foundational capabilities of ROCm, our core software stack, run time, libraries, compilers, tools, framework integration, a lot of the core software that makes the stack work. On top of that, we built a layer of AI-assisted optimization called Hyperloom. Using Hyperloom, the system can analyze the workload, tune configurations, select and tune kernels, adjust parallelism strategy and iterate towards performance goals. To give you an example of the power of Hyperloom, I was talking to the lead developer last week. He told me they just pushed through a suite of 14,000 models through Hyperloom, optimizing them and creating valuable insights. This would have been impossible to imagine even with a large team of engineers before, that's the power of Hyperloom.

And finally, we provide an AI-native interface for our developers, AI skills that allow coding agents to understand ROCm natively. I'm incredibly excited by what we have built with ROCm.AI. It's going to profoundly transform the way developers interact with our platforms. So let me make this concrete for you with an example. Imagine you are a developer who wants to optimize MiniMax M3 with vLLM on MI355. Today, there are many things you need to get right. You need the right vLLM optimizations, the right model recipes, the right environment variables. You may need to write new kernels or optimize kernels. And then you still need to tune all of this. With ROCm.AI, this interaction becomes much simpler. So let's see it. You can say, optimize MiniMax M3 with Hyperloom. Enter. And behind the scenes, the agent does what an expert would do, pulls on a known-good recipe, runs and profiles the workload. It creates and test a bunch of kernel and runtime configurations, and it keeps iterating towards target. You can actually watch on screen here, ROCm.AI write a GPU kernel. It sees an opportunity to write a more optimized MOE grouped GEMM. And the result ROCm.AI delivers a 38% more tokens per second improvement on this specific example. This is the experience we want for developers. Simple to start -- that's pretty cool, yes. Simple, simple to start and powerful to optimize. We have continued our relentless focus on performance with every release, we have delivered significant performance gains on leading models like DeepSeek, ROCm.AI Delivers 3.3x speed up over ROCm 7. This has been made possible by innovations at every level of the stack, techniques such as the use of block-sparse, fused MOE kernels, KV cache quantization, expert parallelism. But look, the bigger story is how AI agents actually profile the workload, propose new kernels. tested configurations and actually validated all of the results. That exact same focus is also on delivering performance gains for training. On the same hardware, ROCm.AI improved training performance by an average of 2.4x through optimized kernels such as fused FlashAttention, more efficient checkpointing and advanced parallelism strategies.

Our engineers have done amazing work to unlock the capabilities of MI455 through ROCm.AI. So I'm delighted to share some incredible results on hardware. On running real workloads, we are able to see the system deliver 20 terabytes of memory bandwidth and 20 petaflops have delivered compute. These are the highest demonstrated compute plus memory capabilities of any accelerated platform in the industry today. It's not theoretical, there's measure.

Super exciting that we were able to share this on day 1. I'm also thrilled to share that the leading AI ecosystem partners who had their hands on MI455 have had a delightful experience. You can see here what they've had to say from end-to-end PyTorch testing to ensuring Hugging Face models are getting validated to vLLM and SGLang running well on MI455, we are incredibly encouraged and excited by what the community is about to unlock with MI455. This matters because day zero readiness is one of the most important objectives we strive for, and that's what ROCm.AI enables.

So now let's look at ROCm.AI in action on Helios. This time, we are going to use a Codex front-end. Let's deploy DeepSeek-V4 Pro, a 1.6 trillion parameter frontier model on Helios. Now watch what happens, ROCm.AI pulls the right recipes. It configures the run time. It implements the model graph on the hardware and finally, serves up the model. There you go. The model is actually up and running. And let's see the model write a poem point now about advancing AI. Let's hit Enter, is that a pretty cool poem? This is the experience we want to give our users. The developer should not have to manually reason through every layer of model architecture, kernel selection, comp strategy and ROCm-level deployment. AI is transforming this experience and that is a very, very big shift for GPU hardware and software.

Now there are a few organizations shaping the frontier of AI as much as OpenAI. You heard Lisa and Sachin talk about our partnership, it's been truly wonderful, collaborating closely with the OpenAI technical team to make rapid progress with ROCm. To talk about how they are advancing the frontier of software development and our collaboration, I'm delighted to welcome to stage Philippe Tillet from OpenAI.

Thank you for joining us here, Philippe. So just so you all know, Philippe is the creator of Triton, One of the most important software innovations in modern AI infrastructure. It's a huge privilege having you here.

OpenAI has always pushed the limits of infrastructure. So can you tell the team a little bit here today as to how your team actually develops GPU software and how have technologies like Triton evolved?

For sure. Well, first of all, thanks so much for having me. I was here 2 years ago. It's my great pleasure to be back now. Look, our way of looking at things is there's no one way to fit all applications, and we really try to meet users where they are. So for that reason, we've developed multiple solutions. So far, workflows where speed of iteration is more important. We've developed Triton, as you mentioned. So this is typically what researchers will use. But more and more, we've been in the case where we've had to hyper optimize performance and Triton didn't give us the level of control that we wanted.

So we've developed Gluon for that, which is a lower level language. And this we've been really able to optimize our kernels for AMD, among other things. And yes, like our agents have become extremely good at both of them, and we've worked with AMD on both of them as well.

Yes. And more and more, what we're seeing is agents slowly taking over like, yes, they're getting just very good, as you just mentioned, at writing this kind of code.

Amazing. So look, I think the work you have done with Triton and now Gluon, right? It's been really, really moving the ball forward for the entire industry. Now we began our work together with 300 and extended at 350, but now we're collaborating at a different scale with 450 and Helios. So would love to hear your thoughts on how that collaboration has been going and how has that experience been?

Yes. Super well. It seems like ages, I think we started years ago and hardware codesign when MI450 was like in the very early prototyping stages, I think we're able to work really well together, and it's so nice to see the outcome of that. And the chip -- and more recently, we've been working very, very deeply on software enablement for MI450, both for our models, but also for the broader ecosystem. So a lot of development has been done Open Source. And yes, we are collaborating on the full stack now down to the very low level LLVM code generation. so that like all the nice hardware advances that came with MI450 can be properly leveraged in our end-to-end applications.

Yes. And you got your Helios racks. So tell us a little about that.

Yes, it works like we got the chip -- yes, we got the chip and very quickly like a single engineer in a few days, was able to confirm that in everything work. Obviously, like we have to optimize performance. But we're very confident that we can get in a very, very good place.

That's awesome to hear. So look, one of the more amazing things that we have done recently, right, is our teams have collaborated even more closely in terms of using AI to program AMD GPUs. So would love to hear your thoughts and anything you can share with the audience?

Yes. No. I mean, everyone in this room knows how much agents are taking over our jobs in a sense as kernel engineers. And yes, it's been amazing just to see like how much better they've got over just the past 6 months. And just now they're extremely good at generating those high-quality GPU kernels in a way that were about before. And it's been very great collaborating with AMD on just like making sure that agents not only are capable enough but also have all the right context they need to be able to make the decisions and good optimizations.

That's great to hear. Yes. I think the quality of code that's getting produced by AI-assisted kernel generation has been truly remarkable through our collaboration. I'm very excited. Now as you look ahead, right, what crystal ball do you have for us for the future of AI software ecosystem and collaborations like ours?

Yes. So there's 2 things that really come to mind. One as we keep talking is just how good agents are getting. And the other one is, I think how much the open approach to software of AMD is enabling not only for OPenAI but I think for the industry as a whole. I talked a little bit earlier how we've been collaborating on LLVM code generation, LLVM is an open source project of the entire compiler stack from AMD's Open Source. And that has allowed us to make our agents really good at very, very low level cogeneration down to scheduling instructions in assembly. And this has led to like very, very significant performance gain that I don't think we have been able to achieve in a fully closed source stack.

Yes, that's one of the real sort of big value propositions that we have to offer, right? Because everything that we do is in the Open and the fact that AI can actually pick that up is a big one.

Thank you, Philippe. We truly appreciate the partnership, and we're really excited about what we're going to build together.

My pleasure. Thank you so much for having me.

What you just heard is extremely important when we combine better abstractions with AI-assisted development and the close hardware software codesign, we can create some incredibly powerful capabilities, and that's exactly aligned with our strategy. Now everything I've shown you so far has been about AI. But ROCm is also a great stack for developing high-performance and scientific computing applications from frontier models to climate simulation. We use the same ROCm, the same library, same tools, same AI assisted development, one open stack for every single workload. And in fact, when you look at our road map, we have retained a very strong focus on HPC and scientific computing. Just like MI455 leads in AI compute with FP4 and memory components, MI430X leads HPC compute with FP64 and memory performance. 2 platforms purpose-built for 2 very different classes of workload. The MI430X is built by leveraging our modular chiplet architecture to deliver a purposeful variant with native FP64 hardware. Giving customers leadership performance across AI and HPC in a single platform. And this is hardware, double precision, not emulation. For the scientific community, that distinction is everything. This product delivers 288 teraflops of FP64 compute, nearly 9x better than competition and the same leadership, memory capacity and bandwidth of MI455, the MI430X ships in the first half of 2027. This is why leaders in Sovereign AI love this product. The first sovereign exascale AI factories in both U.S. and Europe are being built on the MI430X, discovery at Oak Ridge National Labs and [ Alister cope ] with [ CEA and GENCI ] in Europe.

So look, as we look ahead, let me leave you with this start. Every so often, our industry goes through inflection points. We've all felt it where neural networks first started recognizing cats and dogs, when transformers changed what was possible with AI. I believe we are experiencing another one right now, a real inflection in the ability of AI to program complex hardware systems building on a surface area that's been exposed from Open Source software and leveraging the power of abstracted interfaces, AI is making it dramatically easier to program AI systems. We cannot wait what you will build with it. Thank you.

Now AI is getting pervasively infused into all forms of computing. Enterprises are experiencing the same shift with their own unique requirements. So to tell us more about that transformation, please welcome to stage Senior Vice President and General Manager of Compute and Enterprise AI, Dan McNamara.

Thank you, Vamsi, and good morning, everyone. It is a great pleasure for me to bring to you the third pillar of our strategy, which is powering AI everywhere. And the vision behind this has remained the same for several years for us to deliver the right compute engine to the right workload and they continuously drive optimization points across our entire portfolio from CPUs, GPUs, networking and FPGAs.

So building on what we've already shared, I'll cover how our strategy comes together across our enterprise portfolio and how the foundation we built with the CPU franchise has truly positioned us for the next era of AI across multiple deployment models.

So Lisa talked about the frontier model and the hyperscale landscape and just the rapid and massive advances we've seen over the last 6 to 12 months. And those advances are driving the adoption of AI well beyond the cloud. And as AI continues to grow, it will demand compute beyond the massive purpose-built data centers that we all know and love. Extending into enterprise, personal and physical AI. And each of these brings a new -- unique set of requirements. Enterprises are usually constrained by power and cooling. Personal AIs usually must operate within a fairly limited compute memory footprint and physically AI must perform reliably in real-world demanding environments. And AMD is the only company delivering a complete portfolio that spans all of these models -- development models. So I just mentioned that our journey began and with -- in the data center with our CPU franchise and I want to touch on it a bit. With each generation of EPYC, our strategy has been to listen to our customers, address their evolving needs and deliver the highest performance, lowest total cost of ownership and the fastest time to value. As Lisa mentioned earlier, we've introduced many industry firsts across our generations that our enterprise customers has actually asked for. And that customer focus has brought us to where we are today with EPYC as the leader in enterprise computing. And over the last 3 years, there's been very, very strong adoption across the enterprise. With the world's largest companies moving more and more workloads to AMD. And it's reflected in a number of places. First, public cloud, where our VM consumption has grown at 76% CAGR over the last 3 years. And on-prem deployments with our OEM and ODM partners is growing quite aggressively across key industries. But just as important, this position allows us to understand truly the diversity of the workloads that our enterprise customers run every day. And each workload places new demands on the CPU, some require maximum thread count, some depend on per core performance, while others depend on memory bandwidth cash efficiency or I/O performance.

The really important point is that there is no single SKU that solves every enterprise application. And that's why our EPYC portfolio with Venice spans 800 to 256 cores with a range of power, frequencies and I/O capabilities, giving customers the flexibility to choose the right solution for their workload and most importantly, at the most optimized cost. So Lisa mentioned this earlier, and now we all know, we covered Venice and she showed how very -- we have strong leadership across AI host codes and the agentic workflows. But I wanted to show you how Venice actually leads across the general purpose workloads of the enterprise and these workloads that are running -- our customers rely on every day. As you can see, we have a minimum of 2.5x the performance across our competition, across these very, very important enterprise workloads. So now as we all know, agentic AI is really the next major transition for the enterprise. And it truly is a force multiplier for enterprises, and adoption is accelerating. In fact, the latest data here shows that nearly every enterprise next year will have some form of agentic deployments. And as enterprises move from pilots to production deployments, we've talked to CIOs all the time. And they are looking at 3 major things: First, predictable infrastructure cost and token costs; secondly, making sure the data is secure; and then lastly, and sometimes this is overlooked, right, integrating AI into their existing infrastructure. It's not all greenfield out there for the enterprise customer. So to solve these problems, we truly believe that it's going to be a distributed deployment model that leverages frontier models, cloud services, on-prem infrastructure and of course, AI clients. And as AI has moved from chatbots, responding to prompts to agents doing real work, writing code, summarizing documents and running complex flows. The requirements have evolved also. And those are exactly the requirements we've been designing for multiple years. So while AI will be deployed across all these environments, the on-prem deployments in our enterprise customers have fairly unique constraints. First and foremost, power cooling, really big challenge in space. And of course, costs remain top of the list for CIOs that are trying to drive things forward within -- in the AI world. And then lastly, the full solutions they need to need to be easy to deploy. That's been common for many, many years for the enterprise, but it has to -- that's something we're very, very focused on. So to address these constraints, I'm very happy to announce the launch of the Instinct MI350P.

So today, every enterprise wants AI in the data center, but the challenge is how do you add those capabilities without starting over. That's exactly what the 350P was designed to solve. It's an air-cooled GPU designed to fit within the power and cooling of enterprise servers today without a facilities upgrade. Making it easy to bring LLM scale in, for instance, to today's enterprise data center. And at the same time, there's no compromise in capability. A single MI350P can support up to 260 billion parameters. Allowing customers to run the majority of enterprise AI work that was on a single GPU. And the economics are actually even better and more compelling. The 350P delivers more than 4x the tokens per second per dollar than the competition. It's pretty cool. And it actually turns customers' existing data center into an AI data center. Now, we all know that benchmarks are one part of the equation, right? The other part is really evaluating, does the advantage hold up across the actual workload you run in the business and how your business runs. So we tested a set of workloads enterprise use cases using production models. And the 350P delivers much higher productivity than the competition, delivering up to 2 to 5x tokens -- faster tokens per second than the competition. So if you think about that, that's more tokens per second, more work completed, more users supported all at a lower cost for the business.

So another area that I wanted to talk about is our earliest customers, our own AMD IT team. And they're a tough group to work with, even for us. So -- but they put us through our paces and really this has been our model from day 1 is, we want to put our technology into our own data center first. So like many companies, many enterprises were looking at how to deliver AI service to our employee base, but also manage costs. And protect our data, manage and control the data. So we focused on 2 use cases that we wanted to try out. And first was autonomous threat detection. And that is probably one of the faster-growing agenetic applications that we're seeing today. And the second was a personalized AI assistant running on Open cloud. So both of these run route requests through the router, and the gateway sort of test the complexity, the sensitivity of the information and the service level requirements. So some request or sent to the frontier model, while there's a route to open-weight models running on EPYC and an MI350P. So with intelligent routing, we reduced our token costs by 43%, while delivering up to 3x faster response times for the workloads running locally. So we believe this is how enterprise AI will be deployed. But we also know that from our history here with EPYC that enabling enterprise takes a whole lot more than great Silicon. It requires a complete ecosystem of software and solutions. And over the last several years, we've truly grown significantly and expanded our enterprise AI ecosystem. Today, we support numerous AI ISVs, provide zero day support for all of the industry's leading open models. And deliver robust frameworks that help customers bring AI into production as fast as possible. And it's all built on trusted platform software and OEM partnerships that enterprises rely on every single day. Now everything I just shared comes down to 1 goal for us, and that's helping customers deploy AI where it creates the most value for them, whether it's in the cloud, on-premise or in the edge -- at the edge.

So to bring this to life, I'd like to welcome Jeremy Legg, Chief Technology Officer at AT&T.

Yes. Thanks for having. Yes, this is great. It's kind of a homecoming I grew up in the East Bay. So it's good to be back.

Yes, fantastic. We really appreciate you joining us today. So look, I think everyone knows AT&T is the largest communications company, one of the largest in the world, and I'm sure all these things happen out there probably have AT&T network right now, but...

Hopefully, more than half. But yes.

So AI is touching way more than the network. And you and your team are really doing a fair amount of work driving it. So can you share with us sort of the opportunities you approach, you're working on and your experience to date?

Sure, happy to. At AT&T, we're now burning about 1 trillion-plus tokens per month, and that number has gone up pretty dramatically. It's moving up double digits. And so this is now becoming very pervasive inside of the enterprise. We've got over 100 GenAI models in production across our enterprise as well. And those are stretching from things that I think a lot of folks are doing in customer care, things that we're doing in fraud, but also things that you wouldn't necessarily think about like where is the best place to place a cell tower or RAN that is the most effective use of it inside of a network. So all of these things are happening across our enterprise, but they're happening at a scale that is not necessarily normal. We get over 300,000 calls a day. The transcription of those calls is an enormous workload and then driving the insights out of those calls that we then bring back to the business is another set of insights. So it's really an interesting thing and now we're beginning to actually rebuild entire workflows inside of the company beyond just the point use of a use case. So as we think about HR, we think about finance, and we think about different kinds of things that each of those organizations do, cash forecasting, things that we do on onboarding employees as an examples, we're now agentifying those entire workloads.

Yes. So it is -- we work closely with your team, and we've had many conversations, it's truly amazing, the number of actual opportunities you're addressing with AI. So can you -- it's super broad range of use. So can you just explain sort of how -- what are the key learnings in moving these to AI at enterprise scale and then maybe a bit on what we've done together?

Sure. I mean, I'll start at a place that I don't think a everybody starts from here, which is the human. You need world-class teams in order to do this and you need people that think in workflows. There's an enormous amount of what I call racing from stoplight to stoplight. Folks are going from 0 to 100 and then they hit the next stage of the workflow and then there's a red light, and they haven't worked their way all the way through that. So I would start with people and the quality of the people that you really have across your enterprise in order to do these things. The second point is we are big believers in data sovereignty. And data sovereignty for us means not being tied to a specific chipset, not being tied to a specific model or specific set of development tools you need as an enterprise to manage your data, your data is your fuel in terms of how you implement AI across that broader enterprise.

What we've done with AMD, and AMD has been such a great partner in this, is leveraging open source models, leveraging different sets of chipsets, building our own models, training other models. We're one of the few companies in the world that's actually post trained model on AMD and then been able to drive equivalency in terms of performance and accuracy with other more expensive chipsets in the marketplace. So as we have done that, we've been able to drive down token costs essentially through moneyballing across all of these different models and leveraging that human capital in a way that's made a big difference to our company. And so increasingly, even though token consumption is going up, we're able to manage that underlying token cost at enterprise scale, which is not a small thing.

Yes, It's just a -- it's a bunch of great work. Another key area that's sort of groundbreaking, which is exciting also. You're the first telecom company to actually trained a model on AMD. And also, you're driving fully open source models in telecom. So tell us about that experience and why you felt like that was needed.

We felt like the industry was moving down a path where folks were only going to use close source models to solve solutions. And while closed source models are an important ingredient in that, we didn't want to lose sight of the fact that open source models were going to be a very significant part of the solution. So we announced what we call OTel 1.0. That's the Open Telco AI model been over a year ago or so now. And we had over 18 million downloads of that model since then. We're excited to announce today the launch of OTel 2.0 which is a more advanced set of models that we have trained on AMD, and we're now making that available via open source across the broader industry. And I think is going to give folks an opportunity to see that not only do you need to think about leveraging models, but you also need to think about training those models and how they work across your enterprise.

That's just great. We certainly appreciate the partnership. And we really feel like this is germane to the conversation. It's exactly what we've been talking about today. And I just want to thank you again for joining us and the partnership.

Thank you. You're a big part of it.

Okay. So let me close with one last point. We've established ourselves as a leader in enterprise computing by listening to our customers and delivering the right solutions for their workloads. Today, we're extending that leadership to help customers accelerate their AI journey. This is our vision for powering AI everywhere. And to expand a little bit on the data center into physical and client AI. It's my pleasure to introduce Mr. Jack Huynh, who leads our Computing and Graphics group. Thank you.

It's great to see so many friends, partners and developers here in San Francisco. Today, we have seen how AMD is going to fund the compute foundation for AI in a data center. But the full AI transformation requires 2 more ingredients. Intelligence that becomes personal and intelligence that can understand and act in the physical world. Those are 2 frontiers I want to explore with you here today. Let's begin with personal AI. Enterprise AI is already reshaping the data center for agentic AI will extend far beyond it. Agents are the next great productivity multiplier, the jump from software that responds to software that completes work, a personal agent can double what one person can do. A team agents can multiply that by 10. And as agents gain more capabilities, adoption can grow exponentially, but that comes as a compute consequence. Unlike a single AI query an agent reasons call on tools, and [ goes ] to other agents and run nonstop, agents become more capable and pervasive, so will token consumption mean that demand will require every layer of the computing ecosystem to scale together. Data centers remain the foundation for frontier AI in the most demanding workloads. But edge and client systems can extend that capacity putting more intelligence closer to where data is created in where the work is done. The PC is already one of the most powerful compute engines ever placed in the hands of individual across billions of devices, enormous computing capability is really deployed every single day, yet its potential remains largely untapped. Putting that capacity toward changes the economics of computing. More workloads can run on systems that are already in place, reducing overall infrastructure costs and allowing data centers to focus on the largest, most demanding tasks. And local computing brings inherent advantages, your data stays with you. The applications remain responsive and your system keeps working even when your network does not. And thanks to a new generation of smaller, more capable AI models we're beginning to unlock that potential. And the unlock is happening faster than anyone expected.

Last August, GPT-OSS needed 120 billion parameters to score 80 on GPQA. Just 7 months later, Qwen 3.5 score even higher with only 9 billion better performance with 13x fewer parameters. This is not an incremental improvement, is a dramatic shift in the compute required to deliver advanced intelligence. And this is a trend, not an outlier. Smaller open models are closing the [ frontier gap ] at an extraordinary speed. Qwen 3.6, 27 billion parameters models now outperform leading front tier models from previous generation on graduate level reasoning and a fraction of the size and cost. That changes what is possible. Workloads once reserved for the cloud can now run on top compute engines designed for personal AI compute.

Let's map these nonperformance breakthroughs to AMD platforms. Ryzen AI 400 supports models up to 24 billion parameters more than enough for Qwen 3.5, 9 billion with substantial headroom. Ryzen AI MAX extends that capacity to model as large as 200 billion parameters, running natively on a personal system. This is not reduced AI, compressed the fit on a PC. There's a new class of personal computing built to run powerful models where the work actually happens. And this is exactly what led us to build Ryzen AI Halo. We wanted to give developers a platform powerful enough to run these models locally, yet simple enough to use every day with 120 gigabytes of unified memory and support for models up to 200 billion parameters developers can build, test and iterate locally directly on their desk.

Ryzen AI Halo embodies the future of personal AI for everyone, a future that shortens the distance between an idea and a working model. For developers, that means a dramatically more efficient and optimized path from experimentation to deployment. And as Vamsi showed earlier, ROCm.AI is making a hardware programming radically more accessible. Developers can understand modify and optimize code for ROCm even without deep ROCm expertise and better yet, the same open software foundation extends across all of AMD's hardware, developers can use the frameworks and tools they already know. Start locally on Ryzen AI Halo and scale across the full AMD compute platform without rebuilding their work. And we're not stopping there. Today, we're expanding our partnership with Hugging Face to bring the open AI ecosystem directly on Ryzen Halo. Powerful compute only creates value when developers can quickly access the right models, optimize them for the platform and turn them into new applications. Together, the AMD and Hugging Face will deliver Halo optimized performance for open models and agenetic workflows. How cool is that? And along with co-engineered libraries and tool kits designed to help developers spend less time configuring infrastructure and more time building. In later this year, every Ryzen AI Halo box will include a full year of Hugging Face Pro, putting the power of partnership in developers' hands from day 1. Freedom and more power to the user. This is how we accelerate innovation in the AI era, and we are already pushing that vision further with Gorgon Halo. And seeing is believing, look at how small and beautiful this box is. We increased unified memory from 128 to 192 gigabytes, the largest unified memory pool in its class and increased model support from 200 billion to 300 billion parameters. And this is no longer limited to a single platform form factor, together with the world's leading OEMs, AMD now powers the broad portfolio of personal AI systems from laptops and workstations to many PCs and developer platforms that meet developers, enterprises and creators can choose the right system for the way they want to build and work. Personal AI is not a concept, it's a category and is available right now on AMD. We now have all the pieces. Models are efficient enough. Hardware is powerful enough, software stack is open and ready. It open harnesses like [ Ollama ] Server can route workloads to the right model and right compute. The technology is here. The next challenge is taking it from one developer test to thousands of users across an enterprise. Cisco sees and believes in that future the same way we do, to take us deeper into how we turn that vision to enterprise scale reality, please welcome to the stage, Jeetu Patel, President and Chief Product Officer of Cisco.

Jeetu, how are you doing?

How are you?

You look great.

You too.

Thank you for joining me here today.

Congratulations.

Thank you. Thank you. Now Jeetu you and I talk. We're at an inflection point, and for the past few years, we just spent time, just building intelligence. The next phase is how we deploy this intelligence at scale enterprises. You spend so much at your time talking to CIO technology leaders. What are you telling you? What's changing?

Well, firstly, it's a really exciting time to be alive in tech. And if you just take a step back right now and see what's happening, there's actually a fundamental shift that's happening in the patterns of inferencing. So if you think about the patterns of inferencing that existed during the chatbot era that was human-led, it was very, very spiky. You asked a question, you get an answer. What you're starting to see with agents is they're working 7 by 24. They're very consumptive on bandwidth and infrastructure. So you're starting to see a much more persistent pattern of kind of demand signal for infrastructure. We usually joke around internally. Humans click, but agents swarm. And so as you start to see this, what you're going to have is a tremendous amount of -- and there's 2 big concerns that are also there, which is one is every customer is concerned about security and every customer is concerned about cost. So what you're starting to see happen is inferencing is not just going to be limited to the data center. Inferencing is going to be distributed everywhere. And there is a whole new class of computing that's emerging with this desk side computing that you just announced, where there's going to be inferencing that you're going to have a laptop or a desktop or it's human's work, and they might have a desk side computer where agents are kind of going out and running their jobs on behalf of humans. So that's what we're seeing essentially.

And we're fully aligned Jeetu. It's exactly the shift we're seeing. Now AI is moving closer to employees, compute has moved closer as well. right? The PC is not only just the productivity device because of the intelligence node in the enterprise but deploying intelligence to everybody creates a new type of CIO that you and I talked about so much. How do you operate thousands of these systems with security, governance and control enterprises require? Jeetu how is AMD and Cisco solving it? And what are we building together?

Yes. So this is an exciting time because what you're starting to see is there's a few things that need to really be done. You can't just go out and deploy our desk side computers and hope that everything works out in the enterprise. You need to make sure that it's governed, it's managed. And so there's a few things that are really important. Number one, you have to have an appropriate amount of network bandwidth to satiate the needs of the agents that you're going to run locally. So this notion of network infrastructure that can get modernized to keep up with the needs that you're going to go out and generate is going to be very important. That's number one. Number two, like I said, every customer is really worried about token costs. And so we need to make sure that we can actually contain the cost of an agent's consumption of tokens when it actually starts to go away.

Up and down where needed.

Exactly. Number three, then is how do we monitor agent behavior for safety and security. So that if it does start to do things that we don't want it to do, we can in run time, provide enforcement guardrails. And then fourthly, it's just the overall safety and security that you need to have from an apparatus perspective to get this entire thing humming in the way that you wanted to.

Jeetu, see, how you and I think are very alike. So I mean just -- the opportunity don't just bring at every device as you now talk. How do we enable intros to deploy, manage and feel like with confidence. Why don't we take a look at what we've been building in the past few weeks and months?

Yes. It's exciting. So basically, what we've done is there is a full stack that you can start to think about. So what you folks have done with Halo is you've got an isolated secure agent sandbox. You've got intelligent routing, you've got an MCP kind of core set of integrations. And then what we've done above that is we've made sure that you have the right level of security policy enforcement. We've also got the right level of observability so that it's an end-to-end resilient infrastructure stack. So it's a observability of how is your apparatus and the infrastructure working? Is the agent behavior being done in the right way? Can you observe the agent behavior? And are the tokenomics within the guidelines that you want them to be. And then what we have is a single unified management playing, a control plane that can look at every single dimension that you have and be managed within one environment. And so that's kind of the full stack, and we are really, really excited about the partner too.

Our teams love working with your team, Jeetu. How does it all manage? How does the CIO deploy this at scale?

Yes. So let's take a look at what this looks like, right? Because what we would have is this is Cisco Cloud Control, which is the overall management plane. So if you happen to have Halo devices on every desktop, which it might actually get to. What you want to make sure that you do is you are able to go out and manage your entire estate of not just your Halo fleet but also what do you have running on the data centers as well as what you have running in the cloud. So you can provide full visibility for the entirety of the estate. And then once you've got that, what we also have is this notion of how do we go out and monitor tokenomics. And so is the agent behaving the way that it needs to be behave? Or is the agent actually going out and being overly consumptive? At which point you should be able to isolate any individual Halo device so that you can make sure that you can quarantine it?

Love this. Then we go to see the return on invested capital. It could pay for itself in 3 or 6 months, even quicker.

Exactly. So what you see over here basically is you can tell how every single agent is performing and whether or not they're consuming way too much kind of token costs. And if they are, how do you actually contain the cost as you're going through it.

No, I love this, Jeetu. Now everyone else is wondering, when can we buy this? When is this open for everyone?

Well the good news is you can buy the Halo already today. And then what we're doing is we're making sure that we provide this entire management apparatus that's currently in early availability for select set of customers, but we'll have it in general availability in the U.S. in early fall. So we're excited to make sure that we can provide not just the compute but all of the apparels that's needed to go out and manage this so that you can have inference running in the cloud. You can have inference running a new data center privately or you can have inference your desk side.

Jeetu, we're still greatful for a partner with you and Cisco. We're so excited.

As are we. Thank you so much. Congratulations.

Thank you. What you just saw brings the vision, scale AI seamlessly, securely for personal devices all the way to the cloud. But in the age of AI, the [ intelligent ] should be pervasive and it will be everywhere. The next frontier is where AI leaves the screen, enters the world and acts. That frontier is physical AI. The alpha application of agentic frameworks. In the digital world, an agent, plans and complete a task. In the physical world, you must also sense motion, navigate uncertainty and protect people. Here, intelligence does not just produce an answer, it puts an action. When answers become actions, latency becomes safety and reliability becomes trust. That is why the promise of physical AI is so incredibly powerful. It is not about replacing human potential, it's about amplifying it, giving surgeons greater precision, keeping people out of harm's way making farms more productive, supply chains more resilient and factories more efficient. The best technology does not make people smaller, it gives them much greater reach. For more than 20 years, AMD has helped power the core capabilities that robotics depend on, from sensing and functional safety to deterministic control and precision motion. That foundation is really trusted by many of the companies defining the future robotics. And now that foundation is coming alive in extraordinary ways. These are not machines following a script. They are beginning to perceive the world, reason through uncertainty and act with purpose to promise a physical AI already in motion. The next area of robotics is not programmed, it's autonomous. While traditional robot execute instructions, autonomous robots are given an objective and from there, they determine how to achieve it, and they do this all at once in all in real time. The robot must orchestrate intelligence, motion and safety continuously and every part must be spun at the speed of the world around it. That demands an entirely new kind of robotics brain. And to solve that, we rethought this architecture from the ground up. Today, I'm so excited to introduce the AMD Kria AI system on module powered by Ryzen AI Embedded X100. It brings CPU, GPU, NPU and unified memory together in one compact open standard architecture. This allows robots to process perception. AI reasoning and do it all simultaneously and in real time. The architecture is different, but the results are decisive. In third-party testing against NVIDIA Jetson Thor, Kria delivers 3.4x better real-time results, 2.3x more concurrent agents and 1.6x more CPU capacity. That means faster reactions and greater headroom for the robot to take on more complex work. And all of this without compromising control or safety. But to find the next area of robotics takes more than a powerful brain. Developers need a complete path from idea to a machine operating in the real world today. We are introducing AMD Kria AI Robotics Developer Platform. The industry's first turnkey open, wholly integrated robot system. At the center is the Kria AI system on module paired with a robotics carrier card designed for sensing and connectivity built on ROCm and RoCEv2, this platform gives developers everything they need to move from concept to prototype in days. And the same module carries forward to production with no migration or redesign needed at the finish line. You can build faster and move imagination to deployment with absolute confidence.

This is the breadth only AMD can bring to autonomous robotics for the first signal to the final motion. AMD Kria for the brain, Versal for the spine, Zynq for the joints and Spartan for the sensors, each layer of purpose-built for its role, working together in one intelligent system, connecting perception in high-level reasoning to real-time control. The future will not be defined by one model, one machine or one company. It will be built in the open across silicon, software, systems in an ecosystem moving forward together from the cloud to the PC and now into the physical world, AMD continues to build a compute foundation for intelligence everywhere. We're laser focused on the future, so our solution augment and expand what people are capable of achieving. The next era of AI will not just unfold behind the screen. It will unfold the world around us. And now to bring it all together, you take us to what comes next. Please welcome back Lisa to the stage.

All right. Thank you, Jack. And it's been a big day. We've covered a lot from our data center products to enterprise AI to personal and digital AI and the software that ties it all together. This is the strongest product portfolio in our history. And we believe it's the strongest portfolio in the industry. But it's really just at the beginning because what our customers count on is not just today, they really count on us looking ahead and pushing the bleeding edge of technology. So I want to spend the last few minutes giving you a preview of what we're working on. Starting with CPUs. In 2028, we're going to introduce Florence. Florence brings the next Zen 7 cores, it's leading-edge process technology. It's a new set of AI compute extensions to really ensure that we have all of the AI capability and it supports the latest memory technologies. And we're not stopping there. We're already deep in development of Ravenna our 8th generation EPYC family built on Zen 8, and that family is already well under development for 2030. Now to give you a little bit of a flavor, Florence is again going to extend our biggest advantage with EPYC. It's all about product breadth and having the right foundation and also the right family for each workload. So you have Florence, you have Ferrara, you have Fidenza. And it's really built on a full family of CPUs that is optimized for these different workloads. That's getting the customers the right CPU for every workload, and that is our commitment with EPYC.

Now moving to Instinct. We are continuing our cadence of delivering a new generation every single year. MI500 brings next-generation HBM. It brings a larger scale-up domain because scaling is everything. And it's introduced new copper and optical interconnects. And with MI600 is powered by our CDNA-Next architecture, and it's already deep in development for 2028.

Now to give you an idea of what we see from a performance standpoint, every generation of Instinct has delivered a major big step-up in performance. And with MI455, we've talked a lot about it, delivering 35x more inference throughput compared to MI355. But we're really, really excited about MI500. What we see is another opportunity to take a major step up and actually bend the curve again. So MI500 will deliver the largest generational leap in the history of Instinct, putting us on track to deliver more than 2,000x higher inference throughput in just 4 years.

Now what I can tell you is we are working with a number of our customers already on this design point. MI500 is super exciting. And the feedback we're getting from customers is just fantastic. Now put the CPU, GPU and the networking road maps together. And you can now see a complete cadence of our rack scale systems. So you should expect from AMD, a new Helios system every year delivering more performance, more capacity so that we can run larger models, more agents at better total cost of ownership. And that gives our customers the complete predictable road map to plan and scale their infrastructure for years to come. So lots of exciting things in store over the next few years.

So with that, let me wrap this up for this morning. today, I hope you saw a little bit about how -- we're really looking at that compute from every angle. So every aspect I compute, whether you're talking about Helios or MI455 or Venice for the world's largest AI systems or you're talking about MI430 and MI350 for sovereign and enterprise AI. You saw the passion in [indiscernible] with ROCm.AI. It's so good to see how much has come through that platform in terms of making developers -- make it much, much easier for developers to access the AMD platform.

And then Jack talked about Gordon Halo and our new Kria AI platforms, really bringing together the leadership of the end-to-end AI story, including PCs and physical AI. So we'll have some new information. But I want to say a very, very special thank you to all of our partners who joined us today because it is really through those partnerships that we're able to do the most amazing things together.

So if I just leave you with the final thought. When I think about where AI is today, the biggest change that we see is we're no longer talking about what might be possible.

We're actually seeing how AI can have real and significant impact across every industry and every part of our lives, our personal lives. And I have to say that I've spent my entire career in tech, believing that high-performance computing can make the world an incredibly better place. and I have never ever believed that more than I do today. At AMD, what we're focused on is building the technology, the road maps and the partnerships. And there's never been a more exciting moment than today to our 30,000-plus engineers. This is really the next phase of AI. This is where we make AI, give you the opportunity with all the tools and all the capabilities to really bring meaningful real impact. And I can tell you, I could not be more excited to build all of that together with you, our ecosystem.

Thank you so much for joining us today.

All right. Good afternoon, everyone. I hope you are enjoying your time at our Advancing AI 2026 Conference. My name is Matt Ramsay. I lead the Financial Strategy, Investor Relations folks at AMD, and we're delighted that you're here, and I'm delighted to be joined on stage by many of the folks that you saw speaking in the keynote earlier.

For those on the webcast, I'm just going to -- it's an audio-only webcast. So I'll do a little bit of introduction and then provide some ground rules for the conversation that we're going to have here with the investment community. First of all, joining me on stage here, these folks probably need no introduction are our Chair and CEO, Dr. Lisa Su; Vamsi Boppana, who runs our AI business; Dan McNamara, who runs our server business. And next to me is Forrest Norrod, who leads our overall data center business.

We're going to take some Q&A here for the next 45 minutes or so. A couple of ground rules from me. You probably all know that we report our Q2 earnings in about 10 days' time. So if you could ask your questions on today's event and the contents of the conference and the keynote and the related press releases, I'd appreciate it. And if you ask any questions about our near-term financials, we will -- you have wasted your question because I will instruct the folks not to answer it.

Secondly, if you could maybe ask one question so we can get to as many questioners as we can. And the third point I'd like to make is I think we're all aware that there's another company in the ecosystem that reports earnings this afternoon, and those numbers may come out during this session. So if you ask any questions related to that, I'm going to step in on those as well. So let's just make sure we have a productive session and everybody is on the rails. So Liz and Prab from my team are going to be running around to make sure that we get to your questions with microphones. But I just want to turn the floor over to Lisa to make a few opening comments. Thank you.

Okay. Great. Thank you, Matt. Thank you all for being here. I think I've done a lot of talking this morning already. So I'm probably not going to have much in terms of opening commentary other than to say it's just incredible how every -- I feel like every time I talk to a group like this, we have -- so much has happened since -- just in the last few months. So we talked a lot today about sort of the large and growing opportunity in AI. We talked about some new TAMs for the accelerator business as well as the server CPU business. I think we are tremendously excited about the launch of Helios and the launch of Venice and all that it entails. But why don't we just get -- jump right into questions?

Liz, since you're standing next to Tom.

Thank you. Sitting in the front, sometimes pays off. Appreciate it. It's been a great day. I guess I'll start with the CPU TAM. So $220 billion in 2030. You guys had previously talked about 50% of that market. With the bigger number, is that still the expectation? And maybe update us on what you think the share could be?

Yes, absolutely. So we continue to be more and more excited about the CPU market, the agentic AI workloads, the more we talk to customers, the more we understand what's happening, I think we see tremendous growth in CPUs. So yes, we've updated our TAM to over 50% over the next 3 or 4 years, reaching over $200 billion. We're still very much focused on being over 50% of that market. I think we've had really tremendous progress over the last few quarters.

But with the Venice launch, what we're hearing from customers, what we're seeing in terms of the interest is actually an expanding set of workloads. So Venice truly is optimized for AI workloads, head nodes, agentic AI. We're making tremendous progress in enterprise. So all of those things give us confidence that we can continue to grow significantly ahead of the market.

Liz, why don't you just go with Mark there since you're right next to Mark.

Great. Thanks for the great presentation today. Really appreciate it. I had a 2-part on the same topic, if you don't mind, Matt. Part 1 would be the $220 billion TAM for CPU, server CPUs. How does that split between agentic versus standard server versus the head node? And then the second question is -- or the second part of the question is, how do you estimate the TAM here? Because, I mean, if I have -- if I use one agent or if I use 10 agents or I could use 100 agents, it seem like the TAM could grow exponentially. So like how do you get to a number?

Yes. Well, don't -- Dan, why don't I let you start and Forrest and I can add.

Yes, that's a great question. So look, let's start about how we did the TAM, right? I mean, it's obviously talking to customers, but it's also analyzing our own workflows, right? I talked earlier about some of the work we're doing. So we've looked at that very, very closely. Now in terms of the breakdown, in the outer years, we believe that the agentic part of it, which is more sandbox type application will be probably like 50% of it. And then kind of do a random split.

You could argue -- because there's a bit of an argument in terms of general purpose because general purpose gets floated up too for -- due to some of the calls and things like that. So that would -- that's probably the number that I would give you in terms of thinking about the agentic, in particular. And then like Lisa said, is we do believe we're extremely well positioned for that with Venice and as she showed today with Florence. You guys maybe add too.

Maybe the only thing I would add to that, Mark, is we've talked about CPU to GPU ratios and how to really think about those CPU-to-GPU ratios. So if you think about the various categories, if today, in the head node type configuration, the CPU-to-GPU ratio may be 4:1, that is 4 GPUs to 1 CPU. We certainly expect that to tighten as we go into future generations. And then when we add agents, we -- certainly, for the new TAM, we're expecting that the CPU ratio will actually be greater than 1. So maybe we get to the point where it's 2 CPUs for 1 GPU. But it's hard to call exactly, but we're certainly seeing from a workload standpoint, the migration to needing a lot more orchestration around the full end-to-end workload.

I think we can go to Stacy over here on the side, Prab.

I had a question on the Helios ramp. So Lisa, just from your comments, I just wanted to clarify, it really sounded like it was getting -- going in Q4 rather than Q3. And it doesn't fundamentally matter to me which side of the line it lands. I just want to make sure that I understand that properly. And then I just wanted to ask about Anthropic. So you talked about 2 gigawatts. The start of the first gigawatt, I guess, ramping in the first half. Do you guys expect -- of '27, but do you expect to get that first full gigawatt in '27? Or does it stretch out farther? And is that $15 billion to $20 billion kind of content still that you've talked about in the past still kind of the right number for that?

Yes. So Stacy, you have successfully asked 3 or 4 questions.

It's around the same question.

So let me make sure I get through each of them. Starting with where the Helios ramp is. Actually, we will start first shipments here in the third quarter. So you should expect first shipments of Helios to start in September. It will ramp into the fourth quarter, and it will continue to ramp into the first half of next year. We've actually built the ramp this way because it is a complex system.

We want to make sure that we're letting our ODMs really get a chance to really get the manufacturing process fully tuned out. It also corresponds very well to the data center buildup for our largest customers. So we know which data centers these Helios systems are going into. So that's one.

And on Anthropic, we're very, very excited about Anthropic. I mean, I think having really Anthropic, OpenAI, Meta all leaning into Helios is a big deal for AMD. In terms of the Anthropic timeline, as we said, we will start the first gigawatt shipments in the first half of '27. I don't know if I will say exactly all in '27, but we would expect to be fairly aggressive on the ramp of the first gigawatt.

So our plans are to get as much of that into '27 as possible, and it's more just aligning with the data centers and when they are ready for production. Was there another question in there? Content. Again, not talking about any specific customer, but in the same ZIP code.

Liz, maybe Chris Caso is there next to you.

Chris Caso from Wolfe. One of the things that came up during the presentation was comparison between Venice and the AMD -- I'm sorry, the AMD, the ARM ecosystem and kind of it attempted -- it seemed like attempting to put to rest some of the performance and performance per watt characteristics there. Could you speak to that a little bit more and maybe give some indication of where you think your market share may be relative to some of the ARM solutions in the market?

Sure. I don't know. Forrest, do you want to take that and I can add.

Sure. Yes. So first off, we're very pleased with what the team has done on Venice. The whole family of parts, we think, is exceptional. And we've really tuned Venice, as you heard, for a number of different workloads, a number of different deployment scenarios. In that, we think we have achieved the highest performance any way you want to measure it, highest performance per core, highest performance per socket, highest overall throughput performance for just about any workload. But we also are demonstrating, we believe, outstanding performance -- power performance efficiency at each one of those operating points.

And so from our perspective, what we're trying to do is provide the best CPU for any workload regardless of architecture. And I think that the teams have absolutely achieved that. And so for us, it's less about the ISA. It's less about an x86 versus ARM discussions. It's about how do you produce the best CPU for any given workload to deliver the best TCO, and we're highly confident that we've done that.

The only thing I would add there, completely correct. But if you think about the ARM solutions out there, they're very uniquely optimized for like one point. And that covers even some of the cloud ARM solutions, right? They're very optimized. And like Forrest just said, look, we're optimizing -- we're building a complete portfolio to solve multiple problems and hitting those different optimization points also. So I think that's the thing that never comes out in this conversation. And that's why we believe we're extremely well positioned to continue to gain share.

Yes. And maybe just to finish off on the market share point. Look, we're very proud of the progress that we've made certainly across all of the largest clouds are deploying Turin, and that has gone really well. The important point on Venice is we think our share grows. And that's not just because the market is larger. Of course, the market is larger. But we think our share grows because we're seeing the breadth of workloads that people are wanting to put AMD on. And I think that says a lot about the strength of the portfolio. So we're excited about the market, but we're more -- even more excited about just our place in really being the CPU partner across a broad set of workloads.

[indiscernible] overall market?

It's share within x86 and it's share within the overall market.

All right. Prab, I'm going to go to Josh next and Liz, I'll do Joe next. So just to give you a time to walk around with the mic.

Congratulations on the informative day. Maybe following up on Stacy's question. The language in the release for the Anthropic deal was very specific, I think, the 2 gigawatts for MI450. Could you speak to, one, how the deal, I guess, came together from a background perspective, but also, is it -- should we assume that it's multigenerational because it was different language than your OpenAI and Meta deals?

Yes. Well, I think, Josh, what you should expect is that every customer is a little bit different. Every deal is a little bit different. Every one of these things, which we're doing these large strategic engagements is different. With Anthropic, in particular, what we announced was the MI450 engagement. And that was a choice. I think up to 2 gigawatts, very large scale, really ensuring that the first gigawatt gets delivered as soon as possible. So to Stacy's question, I think the vast majority of that will be in 2027, if not all of it.

And to the framing of where we go from here, I think what you should expect is nobody wants to choose an accelerator for a single generation. Like it's just too much work, no matter how good Claude or Codex is. It's a lot of work to get the teams fully integrated. So we are actively talking with every one of our largest customers, including Anthropic about what's beyond MI450. A lot of excitement about MI500. I mean we're getting more and more positive feedback about how that design point is put together.

And then a lot of discussion about where workloads are going in the future and starting with MI600. So you should assume that we view -- just like we did with EPYC, I mean, it's very, very similar where you start with a deep relationship, but you expand into more and more workloads over time. You heard that from Santosh and Meta. That was exactly what we've done is start with an initial installment and move forward. I think the difference today with the foundational model companies is there are no small deployments. There are no pilot deployments. These are at-scale, large deployments for the sake of ensuring that you're amortizing all of the engineering work appropriately.

Josh, before maybe moving to the next question with Joe, Vamsi, since Josh brought up like the agreement and some of the things there with Anthropic, maybe you could spend a little bit of time talking about the Claude collaboration between the 2 companies because I think that's quite important, and it would be good for you to expand on that.

Yes. I think there's 2 aspects that I would bring up, right? Part of it, you saw in my keynote, right? We made some choices in terms of the strategy for how we make it easier to access our platforms, relying on open source and abstractions. What has really helped is because AI now has surface area across all the things that we put out in the open, unlike some of our competition, whether it's instruction sets or compilers or toolchains, they actually learn all that pretty readily right off the bat, and they're productive even now.

But what makes it even more uniquely special is we've been doing work with them to further tune and extend Claude's capabilities to be able to target high-performance optimization. So start with what's out there, which is already pretty good because of our strategy and then further optimize it.

Joe, go ahead.

Joe Moore, Morgan Stanley. I wonder if you could talk about the Cerebras partnership. And how tight do you see that integration going? You talked about disaggregation. How closely do you need to work together to be able to handle those sort of disaggregated workflows?

Yes, I can take that. It's gone really well to date. What I can say is that with our 350s, we started -- we wouldn't be here if we didn't do work already on 350s. So we have things up and running in our lab infrastructure between 350s and their Wafer-Scale Engines and the disaggregation software stack is all serving well. We see excellent performance, which gave us confidence to jump forward to what we would do with Helios. And early work on Helios in terms of analysis and simulation is also progressing well. We expect these deployments for their initial version, which is basically token service under Cerebras Claude to happen by the end of this year.

Prab, maybe to make your life easier, you can go with Ben, and then we can move to Simon and Aaron since they're all sitting next to each other.

Ben Reitzes with Melius. It's great to be here. This is probably for you, Vamsi. ROCm AI, the software, how are you looking at that in terms of disrupting or impacting the ability to run apps versus CUDA? Do you think it's revolutionary? How should we think about -- is that a game changer? Is it evolutionary? And does that help level the playing field, you think, with developers?

It's a great question. Obviously, I have enormous passion for this, right? We truly believe it's the biggest leap that we have made, maybe since the early days where we laid out our strategy. We've made excellent progress every year. But if you have to point to is there a moment in time where we say, okay, this is actually the biggest leap and spring forward, I would point to this time.

Now I don't mean that by August 14, everything is different, right? But the inflection that's happening now, what is likely to happen that we build over the next many months, together with the biggest labs, right, our collaboration with OpenAI, where Codex gets better, our collaboration with Anthropic, where Claude gets better, it's going to be a significant differentiation in terms of like how easily accessible the platform would be relative to any time in the past. So over the coming months, you can expect the productivity of people to access platforms to be quite, quite different.

Simon Leopold with Raymond James. When we think about the new TAM outlook, I wanted to see how you're thinking about the biggest risk to that, in particular, the ability of your customers to get power to their data centers or your ability to get manufacturing capacity, wafers, et cetera. How is that factored into your view on that? And what's the consideration?

Sure. So when we think about TAM and especially with the accelerator TAM being as large as it is, I think we look at all of those components. So not just broad demand, but we also look at what is the rate and pace that power is coming up as well as what is the rate and pace that our suppliers are adding capacity.

So from that standpoint, I think in the near term, I think we have very much planned sort of capacity for significant growth in '27 as well as '28. And then over the longer term, as you're thinking about '29 and '30, I think it's a rate and pace of growth that would require the entire ecosystem to be building at the same pace and have the same vision.

Probably the largest change that we've seen is everyone has been thinking about the accelerator TAM growing very fast. So that has sort of been in the numbers. The fact that the CPU TAM has accelerated as much as it has, has required some adjustments to overall capacity. But we are very happy with the supply chain relationships that we have, and we do see significantly more capacity coming online to satisfy those larger TAMs.

Aaron, go ahead.

Perfect. Aaron Rakers at Wells Fargo. Congrats on all the announcements today. I guess I want to build on that question. Maybe it's not the supply chain, it's your ability to actually stand up these massive rack configurations. So Lisa or Forrest or anybody, if you were to conceptualize like you've got 6 gigawatts here signed up for Meta, OpenAI, 2 gigawatts at Anthropic, how do we think about the pace of your ability to ramp from a gigawatt per quarter basis? Or how do you -- how quickly can you stand up that much capacity? Any kind of color would be helpful.

Well, it's a great question. So if we start answering it from the sort of the rack level on up, because I think Lisa has already addressed it from the rest of the supply chain. First off, we're working very, very closely with our key OEM and ODM partners. So Sanmina, Wiwynn, et cetera, as well as the OEMs to ensure that we've got the manufacturing capacity in place to build the racks, to build, integrate, test and validate the racks at the right pace. And that is an important part of it because the better you can get at that the easier it is to actually support the deployment in the data centers.

Shipping a very high-quality rack is an important part of making sure that you can turn them on very quickly in the data center. Beyond that, we see the next choke point is in actually deploying both logical as well as -- physical as well as logical deployment of the racks. And that's something that we're working again very closely with, with our manufacturing and OEM partners.

One of the things that we acquired as part of the ZT acquisition was we acquired a large services arm, which we have retained as part of AMD. And we are actually using that team right now, not just to support some legacy customers, but also do all of our internal deployments within AMD and then to help our customers deploy very rapidly, both MI350s as well as MI455 systems in their data center. So with that set of capabilities and training our partners, we're pretty confident that we'll be able to stand up to build at the pace required and then to stand up and provision and get turned on the systems in the customers' data centers.

And maybe the only thing I would add to that is now when you talk about overall data center power, we're also very active in that process with our customers so that as they're planning power, we're planning the GPUs and the Helios systems that go along with that. So it's much, much more involved than it was in the past where somebody just places an order. I think there is easily 12 to 18-month visibility.

Maybe we'll go to Srini here. And if I've missed people out there is that the quality of these spotlights is spectacular. So I'm not doing intentionally.

We are kind of on the bright side. Yes.

Srini from RBC. Lisa, I have a question on your road map that you talked about, in particular, the scale-up networking. I think you mentioned optical and copper with MI500. I'm just curious if you think the market and the ecosystem is ready for optical -- or will be ready for optical next year? And if so, do you have all the pieces of the puzzle to be able to support that? And also, as part of that, I saw ESUN highlighted a bit more than UALink. I just want to hear your thoughts on which, I guess, scale-up you will be supporting going forward?

Yes. You want to start that?

Sure. Yes, let me start. So first off, I'll take both of those pieces in order. So we do see the MI500 generation as the one where we begin transitioning from a purely optical -- sorry, purely electrical interconnect for scale-up networking to start to see optical play a role as well. It's going to be a transition. We don't view this as a light switch. We don't view this as, "Hey, we're going to hit a generation, whatever, MI500, MI600, and everything is going to flip to optical. Instead, we see MI500 starting the transition.

We're working very closely with a number of partners across the ecosystem as well as we've been investing in optics for quite some time. And we're highly confident of our ability to begin that transition that the terminus of that generations out in the future is co-packaged optics on all the major components and optical really being the backbone of many connections within the rack as well as between. But that's going to take a little bit of time to get there.

And we think, again, doing it in this phased approach allows us, our customers and our suppliers -- our partners in the supply chain, all to gain experience and to make sure that we're moving at the appropriate pace and not taking any operational disruptions.

On the scale-up protocol, look, we -- on MI450, we support UALink transported over Ethernet. ESUN is a set of extensions to Ethernet, which is helpful with that, and it actually is going to continue to evolve. And so we do expect to see that protocol brought forward and again, UALink over Ethernet brought forward and being available in MI500 as well. But that's not the only thing we're doing. And so we'll unpack more around scale up as we get closer to the MI500 time frame.

Atif, maybe you want to -- Lisa, did you want to expand on that at all? Or I kind of jumped the gun there. So Atif, go ahead.

Yes. Atif Malik, Citigroup. I have a question on the MI500 ramp as well. HBM content is a very important part of your performance and token economics. And a couple of your peers have cut their content for HBM memory in the future because of the availability of the memory. And my question is if your thinking has changed or evolved maybe in the last 6 months or so on how you're thinking about the content increase for MI500?

I think -- so we obviously study the workload characteristics and how capacity impacts, right? The first order, right, we separate out bandwidth and capacity. Bandwidth has a tendency to lift more boats in terms of more workloads directly getting impacted. So that's one order of consideration, and then you look at capacity after that. One advantage is because of the way we have our chiplet architecture, it actually gives us more flexibility and options in terms of how we are able to optimize capacity while preserving top order bandwidth constraints. So that's the uniqueness of our architecture.

We've leveraged that with our existing products, and we do expect to leverage that with future products as well. We're not sharing the exact configurations of what 500 would have or the road map, but that's one unique piece that actually we believe will play to our advantage.

Maybe if I just add to that. I think the way to think about it is we absolutely think our chiplet architecture gives us the ability to be very flexible in terms of memory bandwidth, as Vamsi mentioned, but memory capacity is useful. I mean, our customers have told us. The fact that we have more memory on MI450 is one of the reasons that we're getting better inferencing performance. I think the key is as we're going forward and all of our customers are doing this. I mean, this is an ecosystem discussion that we need to make sure that the memory that is there is really being used because it is such a larger piece of the TCO. And so we are doing some memory optimization along the way. And that's true on both sort of the CPU systems as well as the integrated Helios type systems. But memory is definitely super important. We will just make sure that every amount of memory that we're using is valued by the customer appropriately.

Blayne, you want to go ahead? I think Liz is on your other side there.

Blayne Curtis at Jefferies. I just want to expand on Ben's question on ROCm AI. So just kind of curious where you are on this AI journey, internal use of AI. I know Jensen throw out like half a person salary, which is a big number. But just kind of curious, are you tracking that? And if you could talk about where you are in terms of like day 0 support and automating that with AI and then where else you're using AI?

Yes. I'll comment specifically on ROCm AI and then maybe there's also a broader sort of corporate usage comment in here. So as far as ROCm AI goes, there's actually both internal acceleration of existing features and capabilities, but then externally, what we can put in the hands of developers that come with the platform. So what I mean by that is, imagine you have a profiler or a debugger feature that needs to be built.

Our engineers in the past used to say, okay, this is going to be a team of 20 people, 6 months, right? And now that's actually dramatically cut down because those profilers, debugger features can get out much faster because of the ability of AI Core. And that all comes part of the ROCm accelerated release.

And the piece where it actually helps significantly from an external perspective is the platform now becomes native in terms of AI agents being able to access it, and that's what we are going to start shipping starting August, both from just general out-of-the-box usability, but performance optimization and running these models through it becomes much easier.

Almost everybody on the ROCm team, they're all AI native, more or less because of the group they're in, are pretty much using AI Assist to be able to accelerate their plans to just give you a sense for like how fast or how extensive that is going within AMD.

And maybe to the broader point, we are seeing AI usage ramp up across AMD extremely quickly, I would say, every single month, we're seeing token usage, the amount of -- it's not just the number of tokens, but it's the quality of what we're able to get from AI. Dan mentioned what we're doing in terms of serving different models across our AI stack.

So I would say it is a very much a part of our development process across hardware and software, and I see it continuing to ramp. And these relationships that we are -- these deep relationships with Anthropic and OpenAI as well as a number of the other model companies are helping accelerate that rate and pace.

I think there's a question over here, probably to your left.

This is Bhavtosh, CLSA. Lisa, you started your presentation with 35 quadrillion tokens number, which is already all over the media because it's a shockingly high number for today's environment. My question is, how is AMD projecting demand beyond your conversations with your partners in the ecosystem? Do you have a fundamental way of thinking about where token use will go given current cost of compute? A lot of investors worry about the cyclicality of this industry, and that's where this question is coming from.

Yes. No, look, I think the way we project demand is really quite holistically. So we start with customers, we look at workloads. We look at adoption rates. We certainly look at the free cash flow of our customers to make sure there's the capital behind it. But when you put that all together, every projection that we've put out seemed like it was really high and then the market has actually gone faster.

So we continue to see just very significant demand across virtually every part of the portfolio. And I think that gives us a lot of near-term confidence in these higher market projections. That being said, I mean, we have to see how things develop over time. So I wouldn't say that our crystal ball is perfect. But I can say that it's self-consistent. So it's self-consistent that assumes all of the aspects of is power available, is supply available, is capital available and is productivity going to be able to close that loop when we're looking at TAMs.

All right, folks. I think I'm going to try my best here to wrap this session up and keep my executives here on time because they have a lot of other commitments. I thank you very much for coming out. Lisa and the whole team, it was a great day and a great conference. I think it's really, really exciting to be in a place where there's so much diverse demand for high-performance computing across what's now what, a $2 trillion TAM. So Lisa, if you have any closing remarks, I think we'll close the session if you do.

Yes. No, I'll just say thank you for spending the time with us. It's been a really exciting day. It's a culmination of a lot of work from across the company. What I would like to say is we really think about AI as a complete compute picture. So we talk a lot about CPU TAMs, GPU TAMs, Helios systems, all of that. But we really think about AI as every aspect of compute. And this is a place where we can be quite differentiated in the end-to-end story.

So hopefully, you heard a little bit of the comments from Jeremy at AT&T, the work that we're doing with Cisco, the work that we're doing in physical AI. This is like we're on this 5-year super cycle of just tremendous compute demand and having great partners to work on to unlock all that. So thanks again. We will talk to you soon.

Thanks.

Welcome back to this BofA Global Tech Conference. I'm Vivek Arya from BofA's Semiconductor, Semi-Cap Equipment Research team.

I'm really delighted to have the team from Advanced Micro Devices join us this morning: Jean Hu, Chief Financial Officer; and Matt Ramsay, Head of Investor Relations. And I'll go through my questions, but please feel free to raise your hand if you would like me to bring up something.

But really warm welcome to you, Jean and Matt. I really appreciate you joining us...

Yes. Thank you.

At the conference. And I guess, exciting times in AI and the semiconductor industry.

So maybe, Jean, as a start, just walk us through your state of the union, right? The growth has been very strong, right, since the start of the year. What has surprised you? Which are the areas that you think you are seeing kind of the best growth right now? And which are the areas which have been kind of pushed back because of this?

Yes. First, thank you for having us. Thank you all for joining us.

I think the biggest change during the last few months is really the rise and inflection of agentic AI. You can see continued momentum from training to inference, from AI adoption, experimentation, to more adoption at scale. So when you look at that, agentic AI, it's not about answering questions anymore. It's about orchestration, it's about database access and a lot of tool execution. And all of those require significant CPU performance. And what we are seeing is very significant and incremental demand for our CPU platforms. That has been really exciting.

At the same time, we also are seeing the economics of AI keeps changing, right? With the token generation going up so quickly, all the customers are really focusing on performance, TCO, and they're really trying to figure out how to use different computes to address different application and workloads. So we see demand for GPU going up, CPU and also a lot of other ASIC LPUs.

I think from AMD's perspective, you know our Q1 performance. We had record CPU performance. CPU business grew more than 50%. We guided Q2 CPU performance is going to go up year-over-year 70%. That has been really exciting. So all those things really benefit us because we have been investing in GPU, CPU, adaptive compute, to also address literally end-to-end applications, from data center to clients, to gaming. That feels like right now all the engines are really driving the business revenue growth very significantly.

Got it. So I think the growth in CPU has been kind of the biggest positive driver this year and we are kind of tracking this 2030 or end of the decade type of addressable opportunity, which I think you were the first to point out, right, would be $60 billion. Then said that doesn't sound that high enough, so they took it up to $100 billion. Then you took it up to $120 billion. And Jensen said, well, could be even more than that, right, up to $200 billion.

What do you think is the difference between all these forecasts? This is one of the most frequently asked questions, right, from the investment community. What is the right number? How does one kind of get to it? Is there like a simple unit and are we just all using different ASPs to get to a different number and everyone has the same unit growth?

You're asking a lot of questions here. But I think maybe let's take a step back. You're absolutely right, AMD was the first one. Literally last year, we started to talk about how AI drive the demand of CPUs. So last November, when we had our Financial Analyst Day, we actually outlined how we think about this opportunity under the TAM.

We basically said, okay, we are thinking 3 segments. One is the traditional general compute, which we are all very familiar with, has been driving the traditional enterprise application across the board. The second category is the head node, which is really the communication with the GPUs. You want speed to make sure you have the communication. We also outlined agentic AI last November, which is still very early, but our team already see the early signs of agentic AI. That has been the fastest-growing market in our view. That's when we said, okay, we believe the TAM is going to grow 18% CAGR to $60 billion in 2030.

Remember, the general purpose CPU used to be just a single-digit growth. So that market expansion, we saw it earlier, but we never estimated that it's going to be so fast. The pace and the speed of agentic AI adoption has been tremendous. So when you look at January, February, March, all the enterprise adoption started. You see Anthropic, OpenAI, their revenue has been going up very significantly. The demand for CPU continued to go up across all our customers. That's when we had our earnings call, we actually updated our market forecast to more than $120 billion.

Frankly, it's still very early, right, when you think about agentic AI adoption. What we are seeing is with agentic AI, it's quite complex. When you have millions of operations you have to execute, when you have a lot of agents concurrently working on different tasks, you actually need a very high core count CPUs.

We don't know what other companies, how they think about the TAM and the SAM. For us, we go through the bottom-up and the top-down approach. We know at this particular point, we feel pretty good about more than $120 billion. But it's evolving so fast and more and more complex, so you should expect for the agentic AI portion of the market, because you have so many diverse workloads, so complex, the ASP will continue to increase because, increasingly, you need very high core count CPUs, high-performance CPUs.

Matt, I don't know if you have other you can add here.

Sure. First off, thank you, Vivek, for hosting us and for you guys all coming. I think there's a couple of things, Vivek, that I would maybe add to Jean's comments. One is, I think for this whole audience, and us included, we were trying to figure out when was the primary driver of AI CapEx going to go from almost predominantly large model training and start the shift towards inference being a primary driver of CapEx.

And we're seeing that happen in real time. And I think the more powerful thing for our business and maybe the more transformative thing architecturally is during that transition, you're seeing chatbot inference become agentic inference. And what agentic inference does as a workload, right, you daisy-chain or can catenate all these automated agentic flows together. And between each inference task, there's a lot of CPU diverse work. So post-processing and data, figuring out what to tell the AI to do next, often based on the result of the prior inference. Where do I get the data from for the next step? Is it in a cloud? Is it in an ERP system, a payment system, a CRM system, wherever? Get all of that data, come back. Post-process the data from that step and feed it into the next agentic step for the AI.

And that workload is quite diverse. And that's what we're seeing driving the demand for Turin now in real time. And as the order book fills in for the 256 core 2-nanometer Venice parts that are going to launch in a couple of months and be the primary workhorse for next year, that's where we're seeing the order book really expand.

Got it. Is there a way to segment that market, whatever it happens to be, right, $120 billion, right, plus, in those 3 segments, Jean, that you described, kind of the traditional applications, the part of, let's call it, part of the AI cluster, and then agentic kind of stand-alone CPUs?

Yes. I think the traditional segment, it's very clear, I think 2025, different third party, I would say that's a $25 billion to $30 billion market opportunity. It will continue to grow. As Matt has mentioned, you actually have agents doing more work on the traditional database. So that traditional general purpose CPU, it's probably steady growth, but not as high.

The head nodes which is working with the GPUs, you can see that's when people talk about the ratios. Traditionally, it's 1 head node CPU manage your 8 GPUs. And then become 2 to -- yes, it's getting more and more. So over time that ratio is changing. So we do think the head node segment CPU will also grow very fast, much faster than the traditional general CPU.

But the most exciting portion of the market is actually agentic AI. So agentic AI, you actually -- you are seeing agentic AI server rack sit in between the traditional servers and the GPUs. And those racks are handling all those different workloads to really make sure all the agentic agents work. That market, we think, whatever, the $120 billion or $200 billion market opportunity, is the majority of that large market.

Over 50% of that market you think is that...

Agentic, yes. It is still very early right now. But if you just think about what Matt said, how complex, how dynamic those workloads will be. And we do see significant productivity improvement, especially software engineering side. Those are very complex workloads. You really need different core counts, you need really high performance, high core counts. You also need a simultaneous multi-thread, all those high-performance, you can manage millions of agents potentially.

Right. And the final question there is -- actually 2 last questions. One is x86 versus ARM. So if let's say that agentic opportunity does capture the dominant share of whatever that TAM number happens to be, what advantage, and if I dare ask, what disadvantage does x86 have versus ARM in capturing that?

Yes. I'll start. Matt can talk about the more technical side, is the way AMD has been thinking about it, there has never been, "Oh, x86 architecture or ARM architecture." It's more about how we provide the best TCO for customers. So for us, the performance is number one.

And secondly, what we have been trying to build is the breadth and the depth of platform portfolio. So if you look at our CPU platform, we have gone through 5 generations at the Turin, and the Venice is next generation. We had the breadth from the core count to 8, to 16 core count addressing enterprise applications to the Venice will have 256 core count.

So for us, we have all different configurations, design points to meet all the different compute to give them best TCOs. I don't know technically you can provide them.

Yes, Vivek, I think they're, as Jean talked about, the 3 different buckets, right? The traditional server workloads for enterprise, I think there is an affinity towards x86 because of the codebase. And many of those servers are deployed on-prem, some of them are deployed in third-party cloud. But it's basically the same workload depending on where you want to deploy it. And I think as these agents generate more traffic to all those traditional systems, then the x86 ecosystem is well positioned there.

Then you have the head nodes. I think that what you really need in head nodes is really high single-thread performance, really high bandwidth to -- I don't know, it's not a technical term, but feed the beast, of these really expensive accelerators that are in the box, right?

We will do our own head nodes for Helios. I think our large competitor in their NVL racks will basically do their own head nodes. And then there's a bunch of different XPU programs that are going to be launched by the industry. And I think when we have those conversations with customers, it's much more about what is the performance of your CPU such that I get the best utilization out of this really expensive XPU that I'm buying? It's not an Intel or AMD or x86 versus ARM conversation. It's a, what's the best server part?

And then when you talk about these agentic racks, they're -- it's really about how many concurrent agents can you run per rack or per megawatt of compute. And as Jean mentioned, really, really high-thread and core-count products, I think our road map is differentiated there.

And another thing that doesn't get mentioned but I think is important, is if you think about the dollars of CapEx that are going to come into the server market, the mission-critical nature of these systems, both head node and agentic racks, some of the RAS features and the security features of the x86 ecosystem, it's not an x86 architecture thing, it's just we and our x86 competitor have been put through the paces by every hyperscaler and most enterprises on all of these security features for a decade. And you might imagine, you get into these automated flows where those agents have access to mission-critical data, and you want to really have servers that have robust security features there. So we feel we get brought into essentially every RFQ that's in the industry for these servers, and I think we're positioned to win a very high percentage of them.

Got it. Now your x86 competitor has their own fab, right? They are adding up -- or they're planning to add a lot more capacity on the server CPU side. I guess this year, every CPU is being sold, right? But do you think that there is a scope for share shifts given that your competitor has incremental capacity to devote to this? Or do you think you will be adequately served by your foundry ecosystem?

Yes. We have been very pleased with our share gain trajectory. I think the last Q1 we just announced, we actually, from a value share perspective, we got to like 46%. When you think about it, is our success is really tied together with TSMC. From day 1, we have been using TSMC. We worked together with our chiplet design, with the packaging technology to come up with the best performance server platforms. That continues to benefit us going forward.

We have been planning for this ramp since last year. So if you look at the Q1 performance and the Q2 guidance, when we guide year-over-year 70% increase, and a lot of you actually know is right, the wafer started 6 or 9 months ago. So it is the early planning, how we work with TSMC to make sure we get support to continue to drive the ramp of the supply.

The demand is tremendous, supply is still tight. So I think for us, for this year, next year, every supply we can get, we can provide to our customers. We're going to continue to do that. It's the planning beyond the 2027 right now, for the longer term, I think we'll continue to work with TSMC and prioritize data center as how we drive the business growth going forward.

Vivek, one little thing that I would add to what Jean mentioned there is when we have conversations with the investor community on this particular topic, I think one thing that I've noticed is supply is tight. 3-nanometer is tight. There's a lot of other areas that are tight. And we're -- I think we're very well positioned with the relationships Lisa has personally with the executives in that space to get maybe more than our fair share of the incremental. But it is tight.

I think what I've noticed is maybe the investor scoping of what Lisa and the team initially asked for in '26 and '27 maybe didn't imagine how much growth we had already planned for in supply. So things are tight, but we've already asked for and already been sort of "granted" from TSMC, I think is a different -- I think we're having the right conversation there, but maybe not at the right starting point from what we were sort of allocated initially.

Got it. And then final one on the CPU side, on price. So you mentioned 50% plus growth in Q1 and then 70% plus. I think there was a sense that one could sustain this, right, for the rest of the year. How much of this, Jean, is pricing versus unit growth?

And then let me just add quickly Part B to that. As that agentic, right, that third segment becomes a bigger portion, do you think that can command much higher ASPs? When I talk with the ARM camp, they are talking about ASPs in the $3,000, $4,000, $5,000 range, which is well above the ASPs, right, that AMD is able to get. So just unit versus ASP and just the trend of ASP.

Yes. I think as you mentioned, both the more than 50% growth in Q1 and that we guided Q2 70% year-over-year, more than 70% year-over-year increase, 2/3 of that actually are unit growth. So majority of our revenue growth actually comes from unit expansion, versus ASP increase.

But to your point on the ASP going forward, especially with agentic AI, is a great point, is what we're seeing is we're at the early innings of agentic AI adoption. You can already see, with the software engineering part of agentic work -- autonomous agentic workflow, you really need very high performance, high core count CPUs. That tends to have a higher ASP.

So I think generation over generation, when we go to 2-nanometer with Venice and the future generation to handle complex workload to provide the performance, you tend to see the CPU price to continue to increase.

The other thing I would say is because we have really large platform, the CPU ASP is really determined by the mix and by the different configurations for different workloads. So it's not necessarily -- for enterprise, the per core ASP is very high, but because the core counts are relatively low, so ASP is relatively low. But the gross margin is great. So it's very nuanced from our business because of how big our platform is and how complex or different design point we are helping our customers.

Got it. We spent 2/3 of the time on CPU, so let's talk about GPU now. So you are at the start of launch of a major rack scale system to your portfolio. Maybe give us a status update of how you're feeling. When your competitor launched their first large scale-up product, it took a while, right, for the industry to get ready, liquid cooling, right, and all those things to get ready. So how do you think about your progress so far? How excited are you about the second half?

Yes. We are on track. We sampled MI450. We expect to be launching in second half or Q3 starting point, and Q4. As you know, we have been preparing for this launch for a long time. We acquired ZT Systems in 2024. The whole reason we're acquiring ZT Systems is to help us to add capabilities in system-level design. And we also have worked on the networking side.

So the whole team has been working to prepare this launch. We know it's very hard. It's very complex, very hard. But from an execution perspective, our team has been on track, meticulously preparing for all different aspect launch, not only technology, software side, but more importantly, the supply chain side, how you make sure the design redundancy with all the components, but also supplies, even very small components, that you have to make sure you have supplies. That has been this whole team's execution track record. We feel pretty good right now.

Yes. Vivek, I think we're anticipating, as Jean said, the -- we already have sampled and we have a number of customers that have full Helios racks in their own data center running their full production workload now and testing and doing all those things that you would expect us to be doing at this point in time.

I mean we learned -- the customers learned a lot, we learned a lot from some folks that blazed the trail, as you mentioned. And we're going to really focus on a concentrated set of ODMs as we launch for the first number of months until we get up to scale. And not like a few systems, but billions of dollars of scale. And then we'll be expanding the ODM ecosystem pretty significantly as we move through next year and get towards the MI500 series and whatnot.

So we feel good about where we are. The initial work that the customer base is doing in their own labs with sampled systems has gone very, very well. And so now it's about let's get -- we'll see a pretty significant jump in the fourth quarter of revenue and a fairly significant jump in Q1. And then we'll kind of go from there. So it is wood, so -- touch wood, everything is going exactly as we hoped it goes.

I mean when you're doing a thing like this, there's -- day to day, there's always something, right? But our job is to keep the duck calm on the top as we kick on the bottom, and get ready to ramp this thing starting in Q3.

Got it. When you've been asked, Jean, or Lisa has been asked, you've always mentioned the opportunity for multi-gigawatt scale deployments. I think we have heard about 2 of them, OpenAI and Meta. You think there is still the scope for us to be pleasantly surprised that more such announcements even for '27? Or do you think that just given how tight the industry is, that more incremental announcements could be further spaced out on that?

Yes. We talked about we're really pleased to lead our partnership, very long-term partnership, with both OpenAI and Meta. Last year, we established those relationships and we actually see the forecast from those customers are actually above our original plan for 2027.

I think one of the key milestones we're really focusing on this year is to continue to expand and deepen the customer adoption of MI450. We definitely expect there will be other mega, gigawatt customers. We have both across the board the hyperscale customers, the model companies and the even AI-native companies who we have been engaging, working with. I think it's important for us to continue to expand this customer base, and with the large scale deployment.

Got it. Makes sense. And then how do you expect the 2 customers you have announced to allocate market share? Because they have a number of supplier options. I know you have a somewhat unique engagement, right, with them through the granting of warrants and whatnot. But just if you set that aside, how is your visibility into getting, I think they're 2-gigawatt each, right, from those customers for next year?

Yes. Vivek, as you know, the lead time for MI450 is quite long. So we really have to start to plan and have the orders from customers for next year for the second half launch in the next year.

I think the supply chain overall is tight next year. But since we have been preparing to support our customers, we do want to ensure we get supplies for what they need. I think both are coming higher than our original plan, but we are continuing to increase the supply. Those are 2 of our most important anchor customers.

Of course, we also have a warrant, which we're going to be able to share the upside if we get more revenue. Our [ start ] price is higher, they will share upside too. That is absolutely the common interest for both companies that want to drive the adoption of MI450. We feel pretty good with the both of them. We think we can get the right supplies to support them.

Got it. Do you think having someone like a TSMC is almost adding a level of discipline, right, in the ecosystem? Because it's just harder to double count if everyone has to kind of go to the same source, right, to get -- I mean I know it sounds obvious, but do you see that also?

They're quite disciplined. I think they'll continue to be. I mean these are sensitive topics for us to talk about in these forums. But like we really, really, really appreciate the support that TSMC is giving us as a partner and the way that they're working with us to bring more capacity online to support not just the Helios and 450 growth, but the significant server TAM expansion that we're seeing and planning for, not just for this year, but for all of next year and conversations going into 2028 on wafer capacity to support demand.

So we really appreciate the work that they're doing on our behalf. And I think they're being, as they've always been, disciplined in the way that they bring capacity online and do their diligence 2 or 3 levels deep in the customer base to understand where the organic demand is coming from to support the capacity that they're bringing on.

Makes sense. And then the final one, Jean, is on just memory cost inflation, right? I guess when you wake up and see DRAM -- I'm sure DRAM price is not what you see when you wake up, but when you see that the cost inflation every day, is that a good thing because it gives you more pricing power? Or is that a bad thing because now you have to absorb that cost and hunt for more supply? What are kind of the pros and cons of that memory cost inflation for you?

It's really a great question. First, for us, right, we really want to make sure if there's a longer-term agreement with our suppliers, we can secure the capacity. That is the number one thing. When you have a large business at our scale, you have to plan ahead.

And secondly, we think more strategically. If there's memory cost increase, we do need to figure out a way to share with our customers to absorb that increase together. I think the most important thing for us is we actually don't want to use the shortage of control of memory to just increase the price. We want to make sure customers help us to absorb the cost increase, but we are more strategic with the long-term relationship.

I think from an industry perspective, the memory cost increase at this kind of level we have never seen, right? The price is really high. I think it does increase the CapEx spending for end customers where they spend the CapEx. I think in the longer term, just like all the industry, you always have a new capacity add up. You also have a new innovation on how you utilize memory more efficiently. It will figure out itself in the longer term. That's what we believe.

Got it. But do you have long-term agreements in place for the majority of your memory needs? Or do you think that this is a cyclical high, right, and maybe there will be kind of more normalization of that input cost?

Yes. From a planning perspective, because our manufacturing cycle is so long, if we plan for the MI450 ramp this year, we had to start to talk to our suppliers last year or year before. Similar situation is right now, not only with a very tight supply chain environment, not only we're planning 2027, we are planning 2028. So for us to continue to make sure we can support our ambitious plan to grow the company, we have to plan ahead.

Vivek, in the server market and in the PC market, the majority of -- the vast majority of the DRAM is actually acquired by our OEM and ODM partners and the hyperscalers themselves, right? That DRAM doesn't run through our P&L. And we work really, really hard with all of the customer base to try to make sure that we're matching like the memory supply that they're able to get and the CPU supply -- or the GPU supply that we're able to provide together to where there's not mismatching of supply in the industry or hoarding of anything. And so we've been working really hard on that.

But there's certainly increased input costs to end-devices in certain parts of the consumer and gaming market where there have been some anticipated market impacts. And I think we've given you guys our current view of that in the guidance that we've talked about. But we're working really hard on behalf of the whole customer base to try to make sure that the memory pricing and the matching of components is relatively seamless in a time like this.

Got you. And just final question before we close. When you look across your customer base, Jean, how far is the visibility like that they are willing to provide to you now? And when you get that, what are you doing differently in your business operationally to kind of deliver to those?

Yes. I think because the supply chain environment is really tight, customers really plan ahead. So we have a very good visibility into 2027 and beyond. I think our customers and ourselves, especially when you think about the data center investment, very large CapEx planning, very long planning cycle, they actually plan way ahead. 2027, we feel really good about both the visibility of the demand side and the visibility of the supply side. And of course, 2028 and beyond, that everybody is working on.

Okay. On that positive note, thank you so much, Jean.

Thank you.

Thanks, Vivek.

Appreciate your time.

Greetings, and welcome to the AMD First Quarter 2026 Conference Call. [Operator Instructions] And please note that this conference is being recorded. I will now turn the conference over to Matt Ramsay, Vice President of Financial Strategy and IR. Thank you, Matt. You may begin.

Thank you, and welcome to AMD's First Quarter 2026 Financial Results Conference Call. By now, you should have had the opportunity to review a copy of our earnings press release and the accompanying slides. If you have not had a chance to review these materials, they can be found on the Investor Relations page of amd.com.

We will refer primarily to non-GAAP financial measures during today's call. The full non-GAAP to GAAP reconciliations are available in today's press release and slides posted on our website.

Participants on today's conference call are Dr. Lisa Su, our Chair and CEO; and Jean Hu, Executive Vice President, CFO and Treasurer. This is a live call and will be replayed via webcast on our website.

Before we begin the call, I would like to note that Jean Hu will present at the Bank of America Global TMT Conference on Tuesday, June 2 in San Francisco.

Today's discussion contains forward-looking statements based on current beliefs, assumptions and expectations, speak only as of today and as such, involve risks and uncertainties that could cause actual results to differ materially from our current expectations. Please refer to our cautionary statement in our press release for more information on factors that could cause actual results to differ materially.

With that, I will hand the call over to Lisa.

Thank you, Matt, and good afternoon to all those listening in today. We delivered an outstanding start to the year driven by accelerating demand for AI infrastructure across our portfolio. Growth was broad-based with every segment increasing year-over-year, led by 57% data center revenue growth. First quarter revenue increased 38% year-over-year to $10.3 billion, earnings grew more than 40%, and free cash flow more than tripled to a record $2.6 billion, driven by significantly higher sales of EPYC CPUs, Instinct GPUs and Ryzen processors. These results mark a clear inflection in our growth trajectory and a structural shift in our business. Data center is now the primary driver of our revenue and earnings growth. And as AI adoption scales, demand is increasing not only for accelerators, but also for the high-performance CPUs that power and orchestrate those workloads.

Turning to our segments. Data center revenue increased 57% year-over-year to a record $5.8 billion, led by strong demand for our EPYC CPUs and Instinct GPUs. In server, we delivered our fourth consecutive quarter of record server CPU revenue. Revenue increased more than 50% year-over-year with sales to both cloud and enterprise customers each growing more than 50%. Share gains accelerated year-over-year, reflecting the ramp of fifth-gen EPYC Turin CPUs and continued strength of fourth gen EPYC processors across a wide range of workloads. In cloud, AI was the primary driver of growth in the quarter as every major cloud provider expanded their EPYC footprint to support a broad range of AI workloads from general purpose compute and data processing to head nodes for accelerators and emerging agentic applications. EPYC-powered cloud instances increased nearly 50% year-over-year to more than 1,600 with instances optimized for virtually every enterprise workload and expanded availability across the largest global cloud providers. In enterprise, demand accelerated with record revenue and record sell-through in the quarter. We expanded our customer base with new wins across financial services, health care, industrial and digital infrastructure companies while also building momentum with mid-market and SMB customers. We are well positioned to continue gaining share as more enterprises standardize on EPYC across on-prem and hybrid environments based on our leadership performance and TCO.

Looking ahead, our sixth-gen EPYC Venice processor built on our Zen6 architecture and 2-nanometer process technology is designed to extend our leadership across cloud, enterprise and AI workloads. The Venice family spans a broad set of CPUs optimized for throughput, performance per watt and performance per dollar, including Verano, our first EPYC CPU purpose built for AI infrastructure. Across the portfolio, Venice widens our competitive advantage, delivering substantially higher performance per socket and per watt versus competitive x86 offerings and more than 2x throughput per socket versus leading ARM-based AI solutions. Customer demand is very strong with more customers validating and ramping platforms at this stage than with any prior EPYC generation, and we remain on track to launch Venice later this year.

Looking more broadly, we are seeing a meaningful acceleration in customer demand driven by the rapid scaling of AI workloads across both cloud and enterprise. Inferencing and agentic AI are increasing the need for server CPU compute as these workloads require additional CPU processing for orchestration, data movement and parallel execution in addition to serving as the head notes for GPUs and accelerators. As a result, we are seeing both stronger near-term demand and deeper engagement with customers on long-term capacity planning.

At our Financial Analyst Day in November, we outlined the server CPU market growing at approximately 18% annually over the next 3 to 5 years. Based on the demand signals we are seeing today and the structural increase in CPU compute requirements driven by agentic AI, we now expect the server CPU TAM to grow at greater than 35% annually, reaching over $120 billion by 2030. In response to this demand, we are working closely with our supply chain partners to meaningfully increase our wafer and back-end capacities to support this growth. As a result, we now expect server CPU revenue to grow by more than 70% year-over-year in the second quarter, with robust growth continuing through the second half of 2026 and into 2027 as we ramp our next-generation EPYC processors.

Now turning to our data center AI business. Revenue grew by a significant double-digit percentage year-over-year as adoption of Instinct accelerates across cloud, enterprise, sovereign and supercomputing customers. We're seeing strong momentum as customers move from pilots to large-scale production deployments, particularly in inference where our leadership memory capacity and bandwidth are key advantages. This momentum is driving deeper long-term customer engagements, including large-scale multi-generation deployments.

A key example is our expanded strategic partnership with Meta to deploy up to 6 gigawatts of AMD Instinct GPUs spanning several product generations. Our agreement includes a custom GPU accelerator based on our MI450 architecture, co-designed to support Meta's next-generation AI workloads. Shipments are on track to begin in the second half of the year, leveraging our Helios rack scale architecture, which integrates Instinct GPUs with EPYC Venice CPUs to deliver fully optimized high-performance AI infrastructure.

Together with our previously announced OpenAI partnership, these engagements position AMD as a core partner to the world's largest AI infrastructure builders with deep co-engineering relationships and multiyear visibility into large-scale deployments. More broadly, Instinct adoption continues to expand across AI native and enterprise customers for both training and inference workloads. Existing partners are expanding Instinct across a broader set of workloads, while a growing number of new partners are deploying production AI workloads on Instinct, highlighting the maturity of our hardware and software stack.

On the software front, we continue to make strong progress with ROCm, improving performance, scalability and enabling customers to reach production faster. In our latest MLPerf results, MI355X delivered strong competitive performance across the full suite with leadership results in multiple categories. We also expanded day 0 support for the leading open models, including the latest Google Gemma 4 family, Qwen, Kimi and others, enabling customers to deploy new models quickly with optimized performance.

To build on this momentum, we have significantly accelerated our ROCm development cadence through increased software investments and agent-based coding workflows, enabling faster performance improvements and more rapid deployment of new capabilities. Looking ahead, customer pull for Helios is very strong, driven by our leadership performance, memory bandwidth and scale out capacity. Helios development is progressing well with strong execution across silicon software and systems as we advance through key milestones. We have begun sampling MI450 series GPUs to lead customers and remain on track to ramp Helios production shipments in the second half of the year. As we approach production, demand for MI450 series GPUs continues to strengthen, with lead customer forecasts now exceeding our initial plans and a growing number of new customers engaging on large-scale deployments, including additional multi-gigawatt opportunities. With this expanded visibility, we have strong and increasing confidence in our ability to deliver tens of billions of dollars in annual data center AI revenue in 2027 and to exceed our long-term growth target of greater than 80% in the coming years.

I look forward to sharing more on our next-generation Instinct GPUs, EPYC processors, Helios rack scale platform and our growing customer engagements at our Advancing AI event in July.

Turning to client and gaming. Segment revenue increased 23% year-over-year to $3.6 billion. In client, revenue grew 26% year-over-year to $2.9 billion, led by strong sales of our latest Ryzen processors and continued share gains across consumer and commercial markets. In desktop, we strengthened our Ryzen lineup, including our latest X3D processors that deliver leadership performance across gaming, content creation and professional workloads. We also introduced the Ryzen AI 400 series and Ryzen AI Pro 400 series desktop CPUs -- our AITC offerings across both consumer and commercial systems. In mobile, we delivered strong growth driven by a richer product mix as Ryzen 400 mobile PC shipments ramped and commercial adoption increased. Commercial was a key highlight in the quarter with sell-through of Ryzen Pro PCs increasing more than 50% year-over-year as Dell, HP and Lenovo broadened their AMD offerings. We also closed new enterprise wins across large technology, financial services, health care and aerospace customers. Looking ahead, we expect demand for our Ryzen CPUs to remain solid in the second quarter. However, we are planning for second half PC shipments to be lower due to higher memory and component costs. Against this backdrop, we still expect our client revenue to grow year-over-year and outperform the market, driven by the strength of our Ryzen portfolio and expanding commercial adoption.

In gaming, revenue increased 11% year-over-year to $720 million. Semi-custom revenue declined year-over-year as expected at this stage of the console cycle, while engagements with customers on next-generation platforms remain strong. In graphics, revenue increased year-over-year led by demand for our latest generation Radeon 9000 series GPUs. We also strengthened our Radeon portfolio with updates to our FSR software that improved performance and digital quality across a broad set of gaming workloads. Similar to the PC market, we believe that second half demand in gaming will be impacted by higher memory and component costs, and we are planning the business accordingly.

Turning to our Embedded segment. Revenue increased 6% year-over-year to $873 million, driven by strength in test, measurement and emulation, aerospace and defense and communications as well as increased adoption of our embedded x86 products. Design win momentum grew by a double-digit percentage year-over-year with billions of dollars in new wins across markets, reflecting the continued expansion of our embedded business from a primarily FPGA-focused portfolio to a broader set of adaptive embedded X86 and semi-custom solutions significantly expanding our TAM. Our semi-custom engagements also expanded in the quarter as data center, communications and other embedded customers leverage our broad IP portfolio and high-performance expertise to build differentiated solutions.

In summary, our first quarter results mark a clear step-up in our growth trajectory with accelerating momentum across the business. Our client business continues to outperform the market, driven by rise in adoption and share gains, while an embedded design win momentum and demand are strengthening across our expanded adaptive and x86 portfolio. At the same time, our data center business is inflecting with strong demand for both EPYC and Instinct products significant growth.

While we are still in the early stages of the AI infrastructure cycle, the pace and scale of deployments we are seeing today reinforce both the magnitude and durability of the opportunity ahead. As inferencing and agenetic AI deployment scale, they are fundamentally increasing compute requirements, driving both larger scale accelerator deployments and significantly more CPU compute. AMD is uniquely positioned to lead in this next phase of AI with leadership products across high-performance service CPUs and AI accelerators and the ability to optimize them together as fully integrated rack scale solution. We have a world-class supply chain and are making significant investments to expand capacity and execute at scale. With the momentum we are seeing across the business and the expanding market opportunity, we see a clear path to exceed our long-term financial targets, including delivering more than $20 in EPS over the strategic time frame.

Now I will turn the call over to Jean to provide additional color on our first quarter results. Jean?

Thank you, Lisa, and good afternoon, everyone. I'll start with a review of our first quarter financial results and then provide our current outlook for the second quarter of fiscal 2026. We are pleased with our outstanding first quarter results delivering accelerated revenue growth and earnings expansion driven by strong execution and operating leverage. First quarter revenue was $10.3 billion, exceeding the high end of our guidance, growing 38% year-over-year, driven by strong growth in the data center and client and gaming segments and the return to growth in the embedded segment. Revenue was flat sequentially with continued growth in the data center segment, offset by seasonality in the client and the gaming segment and the embedded segment. Gross margin was 55%, up 170 basis points versus a year ago, driven by a favorable product mix, including a higher data center revenue contribution. Operating expenses were $3.1 billion, an increase of 42% year-over-year as we continue to invest in R&D to support our AI road map and the long-term growth opportunities and go-to-market activities.

As the business scales, operating income grew faster than top line revenue. Operating income was $2.5 billion, representing a 25% operating margin. Taxes, interest and other result in a net expense of approximately $275 million. For the quarter, diluted earnings per share was $1.37, up 43% year-over-year, underscoring the significant operating leverage in our model as we scale.

Now turning to our reportable segment starting with the data center segment. Revenue was a record $5.8 billion, up 57% year-over-year and 7% sequentially, driven by strong demand for EPYC processors and the continued ramp of Instinct GPUs. Data center segment operating income was $1.6 billion or 28% of revenue compared to $932 million or 25% a year ago. Client and gaming segment revenue was $3.6 billion, up 23% year-over-year. On a sequential basis, revenue was down 9%, consistent with seasonality. The client business revenue was $2.9 billion, up 26% year-over-year, driven by strong demand for our latest -- processors, favorable product mix and continued share gains across consumer and commercial markets. Sequentially, client revenue was down 7% due to seasonality. The gaming business revenue was $720 million, up 11% year-over-year, primarily driven by higher demand for Radeon GPUs, partially offset by lower semi customer revenue. Sequentially, gaming revenue was down 15%, consistent with our expectations.

In addition, as Lisa mentioned earlier, we expect second half demand in gaming to be impacted by higher memory and component costs. We now expect second half gaming revenue to decline more than 20% compared to the first half.

Client and gaming segment operating income was $575 million or 16% of revenue compared to $496 million or 17% a year ago. Embedded segment revenue was $873 million, up 6% year-over-year as demand strengthened across several end markets. Sequentially, embedded revenue was seasonally down 8%. Embedded segment operating income was $338 million or 39% of revenue compared to $328 million or 40% a year ago.

Turning to the balance sheet and the cash flow. During the quarter, we generated $3 billion in cash from continuing operations and a record $2.6 billion in free cash flow or 25% of revenue, demonstrating the cash-generating power of our business model. Inventory was roughly flat at $8 billion. At the end of the quarter, cash, cash equivalents and short-term investment was $12.3 billion.

In the quarter, we repurchased 1.1 million shares and returned $221 million to shareholders. We ended the quarter with $9.2 billion authorization remaining under our share repurchase program.

Now turning to our second quarter 2026 outlook. We expect revenue to be approximately $11.2 billion, plus or minus $300 million. At the midpoint of our guidance, revenue is expected to be up 46% year-over-year driven by a very strong growth in our data center segment, growth in our client and the gaming segment and a double-digit growth in our embedded segment. Sequentially, we expect revenue to be up approximately 9% driven by double-digit growth in both our data center and embedded segments and modest growth in our client and gaming segment.

In addition, we expect second quarter non-GAAP gross margin to be approximately 56%, non-GAAP operating expenses to be approximately $3.3 billion, non-GAAP other income and expense to be a gain of approximately $60 million. Non-GAAP effective tax rate to be 13%, and the diluted share count is expected to be approximately 1.66 billion shares.

In closing, the first quarter of 2026 was an outstanding quarter for AMD, reflecting strong momentum across the business with accelerated revenue and earnings expansion. We are very well positioned to build on the momentum as we scale our data center business, expand margins, drive continued earnings growth and the long-term shareholder value creation.

With that, I'll turn it back to Matt for the Q&A session.

Thank you, Jean. Operator, we're ready to start the Q&A session now. [Operator Instructions] But please go ahead and poll for questions. Thank you.

Thank you, Matt. We will now be conducting a question-and-answer session. [Operator Instructions] And the first question comes from the line of Joshua Buchalter with TD Cowen.

Congrats on the results. Actually, I'm going to start with CPUs, which hasn't happened in a bit. It hasn't been that long since you announced the $60 billion server CPU TAM for 2030 at the Analyst Day, and it's very quickly doubled. Agentic AI has obviously gotten a lot of attention in recent months, but it would be helpful to hear your thoughts on how this TAM is inflecting and changing so meaningfully in such a short amount of time. And maybe you could also speak to your confidence in hitting that greater than 50% share target from the Analyst Day as your x86 competitor seems to be improving its supply and also there seems to be more momentum on the merchant and custom arm CPU side.

Yes. Sure, Josh. Thanks for the question. So first of all, back to the -- when we think about CPU TAM, I mean we've always said that CPUs are very critical part of data center infrastructure, and that's been where we've invested. And we saw the first signs of, let's call it, AI demand really pulling CPU demand last year, and that was the reason we updated the TAM to, let's call it, the 18% CAGR or approximately $60 billion. And what we've seen is all of the things that we believed in terms of agentic AI and inferencing and all the CPU compute that is required, is just happening, and it's happening at a much faster pace. So over the last, let's call it, the last few months, as we've talked to our customers and we've seen how AI adoption is really unfolding, we're seeing significant more CPU demand from really every major cloud provider as well as enterprise customers. And the way that comes across is as AI adoption scales, you need more inferencing. As inferencing scales and you do more -- you have more agents and agentic AI, they all require CPUs for all of the orchestration and the data processing and these other tasks. So with that, we've looked at it both bottoms up in terms of talking to customers and having them give us longer-term forecasts as well as just doing some clear workload analysis. And yes, I mean, it's a very exciting TAM. I think it's exciting to see CPUs growing greater than 35% to over $120 billion.

And then when you think about AMD in the context of that, I mean, CPUs are critical for so many tasks that you are seeing a lot more discussion about CPUs in the market. But we actually view it in 3 categories, right? There's general purpose compute. There's the head nodes that really support the AI accelerators. And then there are CPUs just for all of the agentic AI work. And to do all of this, our belief is you need a broad portfolio of CPUs, and that's really what we have been focused on is building not just one type, but really broader in terms of throughput optimized, power optimized, cost optimized, AI infrastructure optimized as we've done in the Venice family. So when you put all that together, we're very excited about the larger TAM, and we're also very happy with the traction that we're getting. We're clearly feeling like we're seeing significant share gain as we're going into our Turin portfolio that has ramped very nicely. Venice is extremely well positioned, and we're working with customers right now on -- beyond Venice and what we're doing in those architectures. So we feel really good about the market as well as our opportunity to grow to greater than 50% share of that market.

I wanted to ask about the Instinct side. So in the press release, you mentioned that MI450 and Helios engagements are strengthening with customer forecast exceeding the expectations and the pipeline growing. You certainly have the big public OpenAI and Meta deals. Was this comment referring to those engagements upsizing versus the announced initial deployments? Or was it other customers and maybe the increase on the MI450 time line? Or is it MI500 and beyond?

Sure, Josh. So we are very excited about MI450 on Helios. We're seeing significant customer interest in those products as well. We have certainly talked about our large partnerships with OpenAI and Meta, and those are going really well. We appreciate the deep co-engineering that is not on there.

When we look at the totality of, let's call it, based on our current visibility, how those forecasts are coming in with all of our customers, we're actually seeing it above our initial plans that we had planned for 2027. And I think the encouraging thing is we're seeing a breadth of customers who are now very interested in deploying at significant scale MI450 series. And those are for both training and inference workloads, although the largest deployments are for inference. And based on all of that and the scale of new customer interest, we see a path to really get to exceed our original targets of greater than 80% CAGR.

And these are really 2027 time frame. Obviously, when we talk to customers, we're talking to them about MI355. There's a lot of good traction we're seeing there. MI450 and Helios, I think for significant large-scale deployments; and then many customers are also very engaged with us on the MI500 series and all of the opportunities there. So we feel like very, very good progress. And the key is that we're continuing to broaden and widen the scope of both customers as well as workloads.

And the next question comes from the line of Thomas O'Malley with Barclays.

Lisa, if I get your numbers correct here in the March quarter, it sounds like the server processor side of the CPU side grew over 50%. If you take it just at the word, it looks like maybe the data center GPU side actually grew in Q1. So I was curious around the cadence of this year kind of previously, you had talked about really a back half weighted and then kind of more so Q4 weighted here. Could you talk about if that's changed at all?

And then the second part of the question is, as you go into 2027, clearly, you're pointing out a lot of upside from the larger customers and then kind of the ecosystem around them with new customers as well. But when you look at supply, that's a major issue in the ecosystem today, could you talk about where you're concerned on supply, if you are? And then any gating factors as you look into next year, whether that be power, data center build-outs, et cetera? Or do you feel really good about the ability to grow?

Yes. Okay. A lot of pieces of that question, Tom. So let me try to get through it. So first of all, on the data center segment in Q1, the server business was greater than 50% year-over-year as we said in the prepared remarks. The data center AI was actually down modestly because of the China transition. We had more China revenue -- I'm sorry, sequentially more China revenue in Q4, and it was less in Q1. But as we go forward, I think we see strong growth in both segments. So we guided data center Q2 up sequentially double digits, and that's double digits in both server as well as data center AI.

And the progression as we go forward. So first, on the server CPU side, we talked about growing to over 70% year-over-year in Q2, and that continuing into the second half of the year. And on the data center AI side, we will be ramping Helios in the second half of the year, so let's call it, starting with initial volume in Q3 with a significant ramp in Q4 and then continuing to ramp in Q1. So that's kind of a little bit of progression.

And then to your questions about customers and supply, I think I answered, Josh, the customer question. I think we have very good visibility now into the deployments that are on track for 2027. And when I say good visibility, it's visibility down to which data centers are the GPU is going to be installed in. And so that's necessary just given all of the constraints out there. We feel that there is tightness in the supply chain, there's certainly tightness in sort of data center build-outs, but we are confident in our ability to supply to the levels of growth that we're talking about and to exceed the levels of growth that we're talking about. And we're also working very closely with our customers and our partners to ensure that we have good visibility to data center power. And there is much more power that's coming online in 2027. And so with all those things in mind, I think, again, lots of things to manage. It's a complex ramp, but we're very pleased with the progress on the ramp.

All right, Tom, I think you shotgun approached the multiple questions there. So operator, maybe we can go on to the next call, please. Thank you.

The next question comes from the line of Ross Seymore with Deutsche Bank.

The first one is just on the EPYC competition. Lisa, you went through some of the statistics of you versus x86 and you versus ARM, but I wanted to dive a little bit deeper into that. How do you see AMD truly differentiating, especially when you're signing -- well, you see some of your competition signing up the same customers from the ARM side and the x86 competition having more supply. So I just I wanted to see if you could dig a little bit deeper into how you think the market share is going to trend over time?

Ross, look, we're very engaged with every major hyperscaler and in terms of understanding their needs on the CPU side. I think we have very much wanted to, let's call it, optimize our CPU road map for the various workloads. I think we were early to call this AI component of CPUs. And so we've been actually optimizing very closely with those customers.

The way to think about this, Ross, is that you're going to need a broad portfolio of CPUs, like not all CPUs are the same. Frankly, you're going to need different CPUs for whether you're talking about general purpose operations or you're talking about head nodes or you're talking about agentic AI tasks, they're going to be optimized differently. And we thought through that, and we are absolutely optimizing across the various workloads. So from a competitive standpoint, we feel very good about where things are. And from a deep relationship with the customer set, I think we feel very good about that. So from our current standpoint, I think the depth of our road map just expands as we go forward. And you shouldn't think about it as people are going to do one or the other. I think you're going to see people actually use x86 and ARM for many of the large hyperscalers. And even for those who are developing their own, they're still buying lots of CPUs in the merchant market for the reason that I just stated, which is unique different CPUs for the different types of workloads, and there's very high demand at the moment.

I guess for my follow-up, maybe more for Jean on the gross margin side of things. It's nice to see the gross margin popping up in the second quarter again. But I just wanted to get some trends longer term, maybe not specific numbers, but how should we think about when Helios and the Instinct side really ramps in the fourth quarter and more so next year. I could see some offsets with that carrying a below corporate average gross margin, but then everything that Lisa talked about with the EPYC side of things being significantly stronger might be more of an offset than it was in the past. So just walk us through the puts and takes of that and maybe directionally where you think gross margin goes over the next year or 2?

Yes, Ross, thanks for the question. We are very pleased with how our gross margin is trending. It came in really strong in Q1. And also, as you mentioned, we guided Q2 higher at 56%. I think as we think about the second half quarter-over-quarter, as you know, there are some puts and takes, right? I would just say, from a tailwind perspective, we actually have multiple tailwinds really are going to help our gross margin. First is the server CPU. Lisa talked about the service CPU expected to grow more than 70% in Q2 and continue to be really strong in second half. That really helps our gross margin.

Secondly, in the second half for gaming actually is going to come down, and our client business actually continued to go up the stack. So from a client and gaming segment, the gross margin actually is going to be also very helpful. Embedded actually is very accretive to our gross margin. Its momentum actually is continuing in the second half. So we are really pleased with all the tailwinds we have. On the other side, MI450 will start to ramp in Q3 and then ramp significantly in Q4. That is below corporate average. So that will have different puts and takes in Q4 in the gross margin side.

But when we sit here, when we look at all the positive trends we have to really offset some of the gross margin dilution from MI450 side, we actually feel really good about the setup of the gross margin for 2026. And into next year, I think some of the tailwinds I talked about that will actually continue. That's why we feel confident about continue to drive the gross margin. We actually, during our financial Analyst Day, we outlined the long-term gross margin in the range of 55% to 58%. We think for the first year, we are making good progress there.

And the next question comes from the line of Timothy Arcuri with UBS.

I wanted to ask about units versus ASP for server CPU. If I look at the June guidance, it sort of implies up 25% to 30% for server CPU. And Lisa, you had mentioned second half of the year. It sort of implies that server CPU could grow like 70%, maybe a little more this year. And so I guess my question is, how much of that growth either in June or for the year, like units versus pricing? Is the -- are these price increases sort of mostly captured in June? Or is that also helping in the back half of the year?

Yes. Tim, the way I would say it is, maybe let me bring you back to Q1 for a moment. So if you look at our significant growth in the server business, it was actually -- although we were up on a year-over-year basis for both ASPs and units, it was actually much more unit driven. So we are shipping more CPUs across not just the high-end Turin family, but we're actually shipping a lot of generalist sort of the Zen 4 family as well.

As we go forward for Q2 and into the second half, we are guiding for a significant amount of growth. I think there's a little bit of ASP in there, but the way we're thinking about pricing to be fair, is we are in a range where the supply chain is tight. And so there are some inflationary pressures. Costs have gone up a bit, and we are sharing some of that with our customers. But we are also being very thoughtful in -- look, this is -- we're playing out for the long term, and that means that we are -- our goal is to ship more units and a lot more units. And so from that standpoint, you should imagine that the majority of the growth is unit driven, and the ASPs are just really to help cover some of the inflationary pressures.

And just to add to what Lisa said, our ASP is increasing because of the mix where actually each new generation, the call counts, those are increasing, that actually drives the ASP up.

And then I guess, Lisa, also -- so there's a lot of new architectures that are being used from multi-tenancy all the way to low latency. And your competitor has talked about the low latency part of the market being 20% plus and they, of course, added to their portfolio there. Can you talk about how you see that part of the market? I mean, obviously, you have enough business now you don't need to worry about that probably for now. But can you talk about that?

Yes, sure. So look, I think what we're seeing is what we expected in the sense that as you go -- as the AI adoption continues and the volumes continue to go up and the overall market goes up, you are going to see, let's call it, different compute architecture is being used because you want to get more cost optimization from that. So we expect that even in that situation, obviously, the vast majority of the TAM is still going to be, let's call it, data center GPUs as the primary accelerator. But you may choose to do optimization around inference, around lowly, around certain parts of the stack, whether it's decode versus prefill, I think that's very natural.

The way we look at it is we're developing a full compute portfolio. So that CPUs, that GPUs, that's the ability to connect to all accelerators as well as the ability to customization for certain customers, and we've also talked about our semi-custom capabilities. And with all of those sort of compute capabilities in our tool chest, I think we will be able to address, very effectively, a large portion of this market, including the low latency portion of the market. So from our standpoint, this is kind of a natural evolution.

Now how fast it goes depends a bit on the technology in terms of what share of the TAM these things become, but we should expect that there will be different variants, and we're well prepared to address those different variants.

And the next question comes from the line of Vivek Arya with Bank of America.

Lisa, do you think agentic CPU growth is incremental? Or is it coming at the expense of GPUs conceptually? So if you're raising server CPU TAM, are you also implicitly kind of raising AI TAM? So just I'm interested in your perspective on what did you think server CPU was as a percentage of AI TAM before? And what is it now with this $120 billion number?

Sure, Vivek. So the way we're thinking about is it's largely additive to the TAM. So you should think about we need all of the accelerators to run these foundational models, and then as these agents do work, they spawn more CPU tasks. So I would say largely incremental. The key is to make sure what we're seeing is in these deployments, the key is to make sure the ratio of CPUs to GPUs are the right ratio. So if you're installing a gigawatt of compute, the ratio -- there's a percentage of CPU as part of that gigawatt will increase. Some of the conversation in the industry has been about CPU to GPU ratios. And it's very hard to call exactly, but we certainly see the movement towards where in the past, the CPU to GPU ratio was primarily just as a host node in like a 1:4 or 1:8 configuration node, now changing and getting closer to a 1:1 configuration or even -- you can even imagine if you get lots and lots of agents that you could have more CPUs and GPUs. So -- but all -- all in all, to answer your question, I think it's largely additive to the TAM. And the key is that everyone is now planning and thinking about CPUs at the same time that they're thinking about their accelerator deployments, which is a good thing.

All right. And from my follow-up, Lisa, we continue to see memory prices go up. I imagine that is both kind of a cost inflation for you but perhaps an opportunity to take price as well. I'm curious, how is that dynamic playing out for AMD? And especially for your customers because a greater part of their CapEx increase is really kind of this memory inflation tax, right, that they have to pay. So how is this dynamic playing out for you and for your customers? And the part that I'm very interested in is that have you secured enough supply versus your other larger competitor who has disclosed a lot of prepayments and other things? So just how is this memory inflation dynamic playing out? And are you kind of adequately supplied from that perspective?

Sure. So Vivek, let me answer the second one first. I think from a supply standpoint, we are very happy with our partnerships with the memory vendors, and we have secured enough supply to certainly meet and exceed our targets. So it is a tight memory environment. Let me be clear. But I think we are very deep partnerships with the memory providers.

And then back to your comments on the inflationary pressures. I mean, look, this is something that everyone in the industry is working with in the time of tight supply, we are seeing some cost increases on the memory side. I think we are all working through that. The way we're seeing it unfold in the market is actually on the data center side, because of the, let's call it, the demand for AI compute, I mean people are largely focused on supply and ensuring that the supply assurance is there. The corollary of that, the larger impact that we're watching is the impact on the consumer markets. And as we said in the prepared remarks, we are expecting that there could be some demand -- sort of the demand impact as a result of the memory price increases on things like the PC business in the second half of the year as well as the gaming business. So we're taking that into account in our overall model. And we continue to work closely with the memory providers as well as our customers to ensure that every time we ship a CPU or GPU, then it's paired with the memory on the other side so that we don't have compute that is not being deployed.

And the next question comes from the line of Aaron Rakers with Wells Fargo.

Congrats on the results. I want to stick on the topic of CPU to GPU. And as we think about the chart that you had outlined at the Analyst Day, there was obviously broken out between traditional CPUs and then the AI bucket on top of that. Obviously, I think the new forecast has a lot to do with the AI CPU expansion. I'm just curious, when you're doing a CPU in an AI workload, is there structurally a different level of ASP tied to that kind of CPU optimized for AI relative to a general purpose server CPU? Any kind of color or help on that would be useful.

Sure, Aaron. So let me start with the broader question. The broader question regarding -- the way we think about the CPU TAM is, again, think about it as 3 categories. So there is a traditional CPUs, let's call it, general purpose CPU TAM that is increasing, but let's call it, increasing at low rate, maybe, let's call it, low double digits, then you have your AI head node, which is connecting to accelerators, which is also growing, but it's smaller. And then the largest piece of the growth is this agentic AI piece, which we think is really stemming from all of the agentic processes. I don't have a number that I can tell you in terms of relative ASPs because it really depends on the workload that is being run. And what we see going forward is as core counts increase, obviously, we will see ASP increase. And that's the direction that we're going in as we go forward. But the main point is the largest portion of this is the agentic AI, the CPUs that are serving these agent AI workloads in terms of the TAM increase.

And as a quick follow up, I'm curious, how do you characterize the competitive landscape as we see some of the ARM introductions in the market. Just curious of your views on the competitive landscape and server CPU.

Yes. Aaron, the best way to think about the server CPU landscape is, again, number one, everyone is talking about CPUs. So that tells you how critical they are for the AI infrastructure. And I think that's a good thing. We feel like we're very well positioned. No question, ARM is good architecture. It has a place in the data center market. We view it as more point products relative to a portfolio, where, from an AMD standpoint, we've built this broad portfolio of CPUs, going forward, what you're going to need for all of these different workloads. And we have, in the Venice time frame, added an AI-optimized CPU with the Verano in addition to our throughput optimized and sort of cost optimized point. So from that standpoint, I think we're very competitive. We're continuing to innovate on architecture. We're continuing to innovate on both advanced packaging as well as all of the architectural pieces. So we feel very well positioned going forward. And the key is the TAM is much, much larger than anybody thought. And so there's a lot of opportunity for different products to be successful in this area.

And the next question comes from the line of CJ Muse with Cantor Fitzgerald.

I guess first question, I was hoping to speak a bit more about client for all of calendar '26. You talked about growth -- expected growth, but would love to hear your thoughts around seasonality in the second half. And I'm assuming that you are repurposing certain logic tiles from clients over to the data center and would love to kind of better understand what the implications are for ASPs on the client side looking into the second half.

Sure. So CJ, I think the client business has performed really well for us. I think if we look at Q1, it actually was a little bit stronger than what we expected. We are seeing some mix shift in the client business. The mix that we're seeing is the M&C or the notebook business is actually growing, especially the premium portion. We're making very good progress in the commercial PC arena with our AI PCs. We did see desktops a little bit softer just given desktop is a more consumer-focused market. And so in that market, it's more impacted by some of the memory pricing and the component price increases.

When we look at the full year, our commentary is we are planning for some demand impact in the second half due to the memory pricing. But even in that environment, what we're focused on is ensuring that we continue to make good progress on the commercial business and continuing to focus on the premium segments of the market. So we believe that we will continue to grow on a year-over-year basis for the client business compared to last year. And as it relates to ASPs, again, it's a little bit of puts and takes between notebook and desktop. But overall, I think we're feeling good about our opportunity to outperform the market and clients going forward. Does that answer it?

That was perfect. And then I guess a question on Instinct gross margins. With compute essentially sold out and obviously, you're building a business, so one has to be, I guess, conservative on that front. But I would think outside of kind of passing through HBM that given the very tight wafer environment that this would be a place where you could look to drive your Instinct margins closer to your corporate average? How are you thinking about that either today or in the coming 1, 2, 3 years?

CJ, at this stage, we really focus on driving the top line revenue growth on our Instinct family for product. I think on the gross margin side, you're absolutely right, it's really -- the demand for compute is tremendous. We actually are very strategic in how we think about the -- how we work with the customers. And of course, the different customers also have a different gross margin. I think, over time, once we start to ramp our revenue, we'll have a lot of opportunities to improve gross margin, both on the ASP side, but also, more importantly, on the cost side when we scale our business.

And the next question comes from the line of Stacy Rasgon with Bernstein Research.

For the first one, I just wanted to make sure I have the near-term AI GPU trajectory correct. So I know you said it was down sequentially in Q1 because of China. You had like $390 million of China revenue in Q4. So the AI business in Q1 actually grow sequentially ex China because it doesn't feel like it given the server outlook? And then I look at what's maybe suggested for Q2, are you thinking GPUs and servers kind of grow similar rate sequentially because it would probably put GPUs in Q2 below the overall revenue in Q4, which seems low to me. I'm just trying to tie all that out. Could you help me with that, please?

Yes. So I think, Stacy, I appreciate the question. I think if you look at Q1, we did mention data center AI was down modest pace sequentially primarily due to lower China revenue in the quarter. I think on your second question regarding Q2, you're right, both data center AI and the server will grow double digit in Q2.

Yes. But you didn't answer my question. In Q1, did it grow sequentially ex the China step down, I guess, is what I'm asking.

The China, for our business, in Q1, it's not material. So I think I will repeat what I just said. Yes, the revenue -- the China revenue in Q1 is not material.

Okay. Okay. So you don't want to -- okay. Second question, OpEx [indiscernible] for spending -- but it sort of continues to blow past the targets. You kind of give an OpEx guide and then it blows through it and then you guide higher. So again, I'm not bothered by the [indiscernible]. I'm just wondering why is the OpEx been so hard to forecast? And how should we be thinking about that through the rest of the year in revenue growth?

Yes. Thanks, Stacy, for that question. I think the most important thing is given the tremendous market opportunities we have, we actually are investing aggressively. If you look at the past several quarters, we're really leaning in, in investing, but all the AI investments are driving the revenue momentum. So if you look at the Q1, revenue was 38% up; but then Q2, we guided 46% up. The investments are driving the revenue momentum. Some of the OpEx increase, of course, is tied to the revenue. When you look at our bid on the revenue side versus our guidance, we did bid on the revenue side, right? So that impacted a little bit. But also, at the same time, we have a lot of customer engagement with our data center AI business, we do continue to make sure we have the resource to support our different customers.

Thank you very much. Operator, I think we have time for one more caller on the call. Thank you.

Our final question comes from the line of Blayne Curtis with Jefferies.

Lisa, I just want to go back to the supply side. There was a lot of story about your competitor restarting 7-nanometer. I'm just kind of curious as you look at that landscape which is quite robust through the decade, do you think that the older products will stay around longer? And is there a way to think about the implications for gross margin in such a strong market. Is that actually a negative?

Actually, Blayne, I don't think we see the older products hanging around longer. In our case, I think it might be company-specific stuff. In our case, we actually see -- first of all, Turin is very strong. We actually crossed over 50% of our revenue being Turin this quarter. Generalist is very strong. We're still shipping some along, but I would say that's come down over time. So in general, people want to use the newer products because they're just more efficient in every aspect from performance, from cost structure, from a power standpoint. So that's what we're seeing.

By the way, I should also mention, in addition to what we're seeing in the cloud segment of server, we're seeing really nice strong pickup in enterprise. And there as well, we're seeing our newer products do very well. So from our standpoint, it is all about ensuring that we ship what the customer needs. And in this case, it typically is our newer products, and we expect that to continue.

As we transition into Venice later this year, we will expect Turin and general to continue shipping, but there's a lot of goodness in going to the new products.

And on the supply chain side, I know there's been a lot of discussion about how tight the supply chain is. The supply chain is tight. I would definitely say that. But I also think this is an area where we excel. We have very deep relationships across the supply chain on the wafer side, on the back end capacity side. And we are seeing meaningful improvements in that. And as our customers come to us with more demand, we are getting more supply. And the good thing about this is we're now talking about '27 CPU demand, we're talking about '28 CPU demand. And so that allows us to just plan much better as we go forward.

And then just a quick one for Jean. I'm just curious to follow up on Stacy's question on OpEx. I guess I was a little surprised that SG&A is kind of outpacing R&D. I was just kind of curious, is that start-up costs, because in a strong market, you wouldn't think you would have to discount or have a big sales effort. So I'm just kind of curious for the year, how you think about R&D growth versus SG&A?

I think for the year, you should expect us to grow R&D much faster than SG&A. I think in the past few quarters, we have been really building our go-to-market machine, and we have been investing more in sales and marketing side. But going forward, you should expect the year-over-year growth R&D will grow faster than SG&A growth.

Yes. And if I just add to that, Blayne, the places that we invest -- Jean is absolutely right. We're investing in R&D ahead of sales and marketing. But the places that we're investing in sales and marketing are paying off. So the investments are going into enterprise servers. They're going into commercial PCs. They're going into mid-market, small and medium business. These are places where AMD traditionally didn't invest, but now that we have a much broader portfolio, both on the server CPU and on the commercial PC side, it makes sense for us to invest because that's sort of the very best part of those markets.

All right. Thank you very much, everybody, for joining and your interest in AMD. John, you can go ahead and close the call now. Thanks.

Thank you. And ladies and gentlemen, that does conclude the question-and-answer session, and that also concludes today's teleconference. We thank you for your participation. Please disconnect your lines, and have a wonderful day.

All right. Good morning. Welcome back, everybody. I'm Joe Moore, Morgan Stanley Semiconductor Research. And one of the real highlights of this conference for me, CEO of AMD, Lisa Su, is here right on the back of signing a major deal with Meta. So very excited for this conversation. Lisa, thank you for coming.

Thank you, Joe.

Maybe before we go into details, let me just start off with the overview of what you're excited about for 2026, how you think this year is going to play out.

Great. Well, first of all, it's great to be here. Nice to see everyone bright and early at 7:00 a.m. I would say, look, I think what we saw coming out of 2025 was just a lot of momentum, a lot of demand for high-performance compute and really an environment where it favors strong product cycles and deep customer relationships. So 2026, if I think about the first few months, it is shaping up to be, again, a very exciting year. We're very excited about the data center business, the overall growth potentials. We're launching MI450. It's looking really good this year.

We have some great customer -- deep customer relationships to talk about there. And then frankly, we see just a tremendous demand for traditional compute as well. So if you look at the CPU cycle, we've always believed that the computing stack is heterogeneous and you're going to need CPUs and GPUs and FPGAs and all of these components. And that's really coming to fruition here in 2026. So a few months ago, we had our Financial Analyst Day. We put out an ambitious financial model to grow sort of 35% CAGR over sort of the next 3-, 4-, 5-year period. And I think as we look at the market dynamics, as we look at the product dynamics, I think we are very much on track to that and with an ambitious target of over $20 of earnings per share in that time frame. So lots to be excited about.

Great. Well, thank you for that. Maybe the big news last week, the deal that you signed with Meta. Maybe give us a little bit of an overview of that deal first.

Yes, absolutely. Very excited about the deepening our strategic relationship with Meta. If you think about what it really takes to build long-term lasting partnerships, it's really about road map alignment, technology alignment and aligning sort of our capabilities with what the customer is trying to achieve. Meta is a customer that we've had a long-term relationship with. They've been a deep user of CPUs throughout their data center portfolio. They've also been early adopters of MI300 and MI350 series. But what we wanted to do at this point is we actually see an inflection point in AI infrastructure.

What we're seeing is the world is much more complicated. Frankly, there is a much more workload specification. So workloads, whether you're talking about training or inference or large models or media models, you need different types of compute. And so we were looking for a way to really turbocharge, strategically deepen our relationship with Meta, and that's what we announced a few weeks ago, really a 6-gigawatt long-term strategic partnership where we're actually doing a semi-custom GPU for Meta along with all of the rest of the work that we do with them on CPUs and other parts of the system.

But it was really a vertically integrated discussion in the sense that we started from the workload first and then work through what is Meta trying to achieve with their workloads, what do they see the future of their data center infrastructure look like and then using our very flexible architecture to design something specifically tailored to their needs, which really allows us to increase our footprint in the Meta ecosystem. So very excited about that. I think as a long-term strategic partnership, it enables us to really build on each generation and frankly, get even more tailored to where the workloads are going in the future.

Thank you. And obviously, a great deal, a lot of enthusiasm for it. There have been some questions around the warrants. Can you talk about the warrants that you issued and how those warrants unlock value for you guys?

Yes, absolutely. So the way to think about it is, as I said, the AI infrastructure ecosystem is at an inflection point where deep partnerships really make a difference. And I have to say we have a lot of customers that we work with very deeply across CPUs and GPUs, and most of them don't get warrants. Warrants are a very special instrument that we use for what I would say are transformational partnerships. So when we look at Meta, what we see is a company that truly has sort of the view of the application stack. They are a foundational model builder.

They are betting big in infrastructure. And there's an opportunity not just as a consumer of chips. I mean, obviously, we'd like people to buy chips. We're talking about triple-digit billion dollar deals. Those are great things. But we actually see an opportunity to go much broader than that in the sense that you're actually charting the path for where AI infrastructure is going in the future. So the value that's accrued to AMD of a deal like this is, yes, we get to accelerate purchases, which is a great thing. But we also get to accelerate our technology ecosystem, our software ecosystem that accrues benefits beyond just the work that we're doing with Meta, but really to the overall AMD ecosystem. And the key with how we've designed these warrants is they're very, very performance-based.

So in some sense, both companies are incented to help each other win. We win when Meta's foundational models are super successful and they need lots and lots of chips. So we are motivated to give them the best infrastructure for their workloads. And they're motivated to ensure that our ecosystem is as strong as it can be, and there's a very good win-win synergistic partnership. But it is a special thing. I think the conversation is around we want to build a very rich ecosystem.

The AI infrastructure world is growing by leaps and bounds, and we have an opportunity to significantly accelerate and align with one of the strongest model builders in the ecosystem as well as our deep partnership with OpenAI, very similar from the standpoint of -- we think the model builders that are driving foundational models going forward have the opportunity to significantly align our road map with that really benefits the overall AMD ecosystem across all customers.

Okay. It's great. So those 2 being special relationships, do you anticipate that other customers for MI455 would have a similar warrant structure?

I don't. I mean you should expect that there are lots of other customers interested in MI450 and MI455. They are great products. I want to make sure that we start with the foundation of -- at the end of the day, the product has to be extremely competitive and frankly, leadership for anyone to spend gigawatts of power on our systems. But what we have is, again, lots of great partnerships across the board. I think we have a very, very competitive road map. We're excited about where MI450 is positioned when we look at the landscape.

We've always been very optimized for inference. You're now seeing the growth of inference exceed training, which is what we all expected, but that's a great thing because that means people are actually using the -- all of these models to now do real work. We're seeing the growth of agentic AI. All of these things favor our architecture. And so when I look at our combination of CPU, GPUs, networking, rack scale systems, we really have all of these pieces coming together. So lots of excitement around MI450, but I would say that the relationships with OpenAI and Meta are pretty special in how they are framed as multigenerational partnerships.

Great. Well, thank you for the overview of that deal. Maybe we could delve a little bit more into your AI products. Starting with the foundation that you had, you've done really well with MI300, MI350. Those have been leadership products that have gotten you to several billion per quarter now, over $2 billion per quarter. But now you're doing big investments into rack-scale Helios, the ZT acquisition. What's different between where you've been and where you're going within this?

Well, I think we've been -- we've made a lot of progress in the data center AI space. I think with each generation, we really increased the capability and the set of workloads that we address. I think MI300 and MI350 were great opportunities for us to really optimize, let's call it, infrastructure for inference. I think our inference capacity and capability has been really exceptional, and we've seen that adoption.

We've also been very focused on building a software ecosystem. So the idea of we want to make it super simple for people to adopt AMD technologies. And we've gone from, let's call it, in the early stages of MI300, it might have taken a number of months for customers to optimize to AMD. And now we're at the point where you can do that in a short number of days. So the tools are that good. I think the libraries are that good.

And frankly, we're using AI extensively in that ecosystem building. And when you go forward to MI450, that's why this year is so exciting for us. It really is a huge step function in capability. It's something that we planned. So we acquired ZT Systems because we believe that the rack-scale infrastructure, the whole goal, if I think about the large investments that it takes in AI infrastructure is to get our systems in the hands of users running workloads in as short a time as possible.

So it's really time to workload. And for that, the more we can do for the customers in terms of the full solution building, the easier it is for customers to deploy. So that's what we've done with the MI450 series. We've taken our view that an open ecosystem is a good thing. So the Helios rack is actually based on a standard that we developed jointly with Meta. And what that allows us to do is really leverage that entire rack-scale system.

The ZT team has been very, very active from day 1 as they joined AMD, really building that rack-scale system infrastructure. And when we look at MI450 today and the progress that we're making, it's just looking really, really good in the labs and running lots of workloads and working very closely with our lead customers.

Yes. You've made some really interesting investments into rack-scale that I know ASIC competitors, for example, aren't going to be able to make. You've laid that foundation for MI455. I guess can you just give us an update? You talked about it working well in the lab. You've talked about revenue in Q3, a bigger volume ramp in Q4. I know your competitor started building racks and there were challenges in the beginning, it took longer than they thought. What's your confidence in sort of the ability to have silicon out on time and then meet those time lines?

Yes. Well, we should start with -- these are very complex systems. So I will be clear with that, Joe. But I think we've done all of the planning and a lot of, let's call it, risk mitigation in terms of building the rack scale system. So even before we had final silicon in place, we were validating the rack-scale system. We've had a significant amount of cycles that are now being run. We've learned a lot from the ecosystem. Frankly, our partners have also been very helpful in sort of some of the early teething pains with rack-scale systems have given us a lot of feedback.

And we've designed Helios with some of the, let's call it, the previous issues in mind so that we do think it is going to come up smoothly. No question that we have a lot of work to do but we feel very good about sort of the steps that we've taken. And the most important thing is to be running workloads on these systems, and that's really what we're doing now. And so we feel good about our positioning. I think we have all the pieces in place. I think we have a strong set of relationships throughout the ecosystem to ensure that Helios ramp goes very smoothly.

And you've talked about this as a leadership product. Is that really -- is that leadership everywhere, leadership in training, leadership in inference. It doesn't seem like you're attacking some segment of the market that your competitor isn't, you're really going right at the center of the market.

Well, look, the MI450 series is a very general purpose capability. I think the way we've designed it, because of our chiplet architecture, it is quite special in how we put it together. What chiplets allow you to do is optimize for different workloads. So if you look at our standard products, we've always had an advantage in memory and memory bandwidth. I think we're going to continue to do that. Those are very, very important when you're talking about large-scale distributed inference and capabilities there.

We've also, in the same family, designed a HPC specific part, so our MI430 series. And the reason I mentioned that is there are going to be these workloads that require different data formats and different capabilities. Because of our chiplet architecture, we can fundamentally mix and match different components that allow us with, let's call it, very incremental work to get very significant workload benefit.

And then with Meta, we did -- we took that to another level to do customer-specific optimization. But we're bullish. I mean we're very bullish about the positioning of MI450. I think it's the right time, it's the right product. We have the customers who are anxious to get it in their data centers. We're now planning, as you can imagine, when you're planning multi-gigawatt deployments, we have to be planning together with the data center build-outs that are happening. And it's exciting to see all that come together.

And how do you think about the positioning versus custom silicon versus ASICs? You talked about some of the customization capabilities you can provide. But it's also -- it seems to me that this is not [indiscernible] We're all sort of focusing on the same types of workloads. So I don't know that we have the same role for ASIC customization than we've had in the past. And yet, the 2 customers that you have big deals with have deals with NVIDIA, have deals with ASIC, have deals with AMD. How do you see all that interrelating with you?

Yes. It's a very good question, Joe. And maybe we can take a minute to kind of break it up in a couple of pieces. So let's start with the workloads. I think what we're seeing in the market and what is clearly the next phase of AI infrastructure is there is like no one chip that does everything the best. It is a heterogeneous world out there. There's actually a continuum of capability going from, let's call it, the largest training clusters to inference to more specific inference workloads to even breaking up the workloads.

And I think this is a natural evolution. When you get into high-volume running AI workload, you want it to be as efficient as possible. And that efficiency comes from performance, but it's also performance per watt, it's also performance per watt per dollar. And so at the volumes that these hyperscalers are running at or these large foundational model companies are running at, you're going to want to do that optimization.

I think what we've always believed is that in that continuum, our portfolio plays really, really well. We're seeing significant CPU demand, frankly, as a result of the inference demand picking up. We're seeing significant demand for our standard product, but we're also seeing this continuum where we can do customizations for specific workloads. And frankly, I think there is always a place for ASICs as well for some more tailored applications. The key is how do we get we want to get for the best of both worlds, right? You want to be able to have flexibility and time to market. That's what we believe our chiplet ecosystem does and our overall ecosystems investments do.

But you also want to be able to tailor for specific workloads. And so that's kind of why we really believe that this world is going to come to a place where you do have different chips that are being optimized for different workloads and the capability that allows you to optimize the quickest where you get, let's call it, maybe not full tailoring or full ASIC, but you're able to get, let's call it, 80%, 90% of the benefit at a shorter time with similar economics is a great thing.

Great. Can you talk about the systems level things that you need to provide? I mean in networking, you're scaling up with UALink, but there's sort of UA link through Ethernet. There's a CPO migration to think about. Can you just kind of talk about your networking road map? And how important is that to the AMD rack-scale road map going forward?

Yes. No question, networking road map is very, very important. What we're trying to unlock is systems performance and systems performance includes all the elements of compute as well as the networking infrastructure to scale up and scale out. I think we have a great team internally part of the acquisition of Pensando that we did. We've done quite a bit of work on understanding the networking workloads. We have our own scale-up NIC as well. We work across the ecosystem in terms of some of the switching partners.

I think the key for this is, again, to -- from our standpoint, it's about open standards and it's about giving the customer choice. So I think UALink is a very specific AI optimized network that we believe can be beneficial. We also believe that there's a large set of customers who gravitate towards Ethernet because of its compatibility. And so we support UALink over Ethernet. We'll continue to support that Ethernet ecosystem. And the key for us is to be very mindful about rack-scale infrastructure performance and capability. Lots of optimization on both the hardware and software ecosystem. I think we're deeply partnered across the ecosystem to deliver that rack-scale performance.

Great. And then before you had the Meta deal, you had the OpenAI deal, which was the same basic size, 6 gigawatts. Everything the same with that deal. I know NVIDIA has kind of moved from a little bit more focus on provisioning their own data centers versus what you're doing, which is more cloud-centric. Just is that OpenAI deal tracking to what you thought it would?

Well, I have to say, first of all, I think our relationship with OpenAI is better than it has ever been. I think our strategic relationship was definitely we're much more tied in terms of road map. We're actively planning what are the installations of the first gigawatt of capability, and it's really playing out as we expected.

So I would say nothing has changed with the overall deal structure. I would say that I am quite, quite pleased. It's actually clearly paying off in terms of the technical alignment that we have, the prework that we're doing across the MI450. We are basically co-validating together. We're planning those installations together. So yes, we feel great about that relationship.

Okay. Great. You've talked about this is a trillion-dollar market end of the decade, and you've talked about $120 billion of AI revenue for AMD. I guess the market seems to be concerned about the sustainability of the strength we're seeing now and people look at the hyperscale cash flows as being sort of neutral to negative and the markets kind of understands that things are strong near term, but worried about the duration. What's your view on that? Why you continue to believe in trillion-dollars we talked backstage. There's a lot of indications of that. What gives you the confidence in the sizing of that market?

Yes. Look, we feel really good about the market, the build-out, sort of the most important thing when you're sitting on my side is to make sure that as the infrastructure is being deployed, it's solving real-world problems. And that's what we see. I mean we see that the investment in AI infrastructure, in some sense, we're equating that investment with productivity and intelligence -- and that's a great thing.

So yes, we are all investing ahead of the curve, but well within the reason of where we think the payoff is going to be. I can just tell you like every week, every month, we are seeing significant new enterprise use cases that are showing the payoff of what AI can give us. And as I talk to enterprise customers, like we're still in the very early innings of deployment. So all of the infrastructure that we are building out, and really, these are planned builds, right? So if you think about CapEx discussions today are planned builds later in '26, '27 and beyond, they're really to address that enterprise demand and really delivering the payoff of AI.

We are seeing it. We are seeing the early signs of it. We're seeing the early signs of it in our own business. We're seeing the early signs of it in our customers' businesses. And I think the thing that's a little bit different, Joe, which maybe people need to understand is it's just not all about GPUs. Like this is not just about deploying accelerators. This is actually about deploying the entire compute infrastructure you need to service all of those agents that we're all going to be spawning with our new AI capabilities, right?

So if you think if a company has 10,000 people and they add another 10,000 agents on top of that, they're going to need a lot more compute to satisfy what all of those agents are doing. And we're seeing that. We're seeing -- actually, as much as I'm very, very excited about the GPU portion of the business. I mean the CPU portion of the business has actually far exceeded my expectations in terms of demand. And I was pretty bullish to begin with.

We talked about -- I talked about like a high teens CAGR in the compute market at our Analyst Day. And I can tell you that every indication that I'm seeing today is that, that compute market is even much larger than that. The ratio of traditional compute to accelerated compute is such that you need really a very balanced system overall.

Yes. That seems like we've moved from theory to seeing that play out in real time now with that.

Did you not believe me when I said that, Joe?

I believed you, but we certainly are seeing evidence of it now. Can you talk about that? And it seems like the microprocessor market is dealing with shortages at the moment. What's your visibility into being able to meet that demand that's out there?

Well, first of all, we have a very strong road map. I think we have executed very well. As we ramped Turin, it was a very, very fast ramp. And with each generation of our EPYC processors, we've actually increased the workload coverage. So we started with, let's call it, the main cloud workloads at the hyperscalers. But we've now really expanded to the breadth and depth that you would expect with strong products, and that's across both hyperscalers as well as enterprise.

I think what we're seeing is really that build-out continue and continue in a very positive fashion. So back to your comment about are there -- is it supply tightness? Yes, there is supply tightness, but that's really because the market sizing is bigger than what we had forecasted 3 or 6 months ago. And so it always takes time for the supply chain to catch up with what the market wants. I can say that we are very, very well positioned from a supply standpoint to meet a large percentage of that demand. We are still working very closely with our supply chain partners to expand that capability as we go through '26 and '27.

And what we're looking for is, again, durable demand that is not just, hey, are we just catching up because we haven't upgraded CPUs? No. I mean that is the wrong way to look at it. Yes, we are upgrading. You did say that, too, didn't you? No. But look, I mean, I think we never know until we actually see what happens in the workload. I would say even the hyperscalers are surprised. So if you talk to our top customers, they're like, wow, Lisa, like the demand for CPU compute sitting along AI was perhaps something that was under forecasted.

And so we are in the process of catching up. But I think it's a great time. It's a great time because, one, we were already, from an AMD standpoint, expanding our workload coverage. And then two, you're seeing the customer demand really strengthen as well. So we will continue to increase our supply coverage as we go through this year and into next year, but this definitely feels like a very durable cycle, and it's a very pleasant -- I won't say a surprise, but it's a pleasant development as we think about our overall goal is to provide the right compute for the right workload. I think our data center business clearly shows that we have all the right pieces for this AI cycle.

And I know Forrest had talked recently about sort of Turin versus Granite Rapids is as close as this is going to get and Venice is a clear indication of AMD pulling ahead. Can you talk about that? And I feel like I guess your competitor would say, we have fabs, we can meet this demand, whereas AMD may be constrained on wafers. Any concern about that?

Well, I would say from a competitive standpoint, we feel really good about Venice. I think we were -- we continue to be very aggressive with each generation of our CPU build-out. We continue to broaden the workloads that we're covering. Venice was one of the very, very first products in TSMC 2-nanometer using our chiplet architecture. It is on track to ramp very nicely in the second half of the year. I think we feel very good about our ability to expand to the demand out there.

I would say what we're seeing about Venice, which tells you a little bit about the competitive position is each generation, what we're trying to do is align customer ramps with our ramp, right? We want customers to have the best technology that they can. And frankly, -- that doesn't always happen because customers have their own cycles that they're going through. With Venice, like every one of our large customers want Venice the moment it comes out. And that kind of gives you a sense of how good it is because if you have power to spend, you want to spend your power on the best technology out there. And that's what Venice will be when it comes out.

Great. And then on the CPU side, can you talk about competition within ARM? Your bigger hyperscale customers do have some ARM deployments that are out there. Just how do you see them fitting into this ecosystem?

Well, look, I think ARM has always been a part of the data center ecosystem. I would say it tends to be on the lower performance side of it. We view it as not about ARM versus x86. We view it as you want the right processor for the right workload. And so the performance per watt capabilities, performance per dollar capabilities, the overall TCO are what's critical. And we think with the broad coverage that we have as we go into the Venice generation, I see our TAM expanding, and I see our share expanding because of the capabilities of Venice.

Okay. Great. So I have one more question, and I'll turn it to the audience. The role of memory in all of this, are you seeing impact of memory shortages on the GPU side, on the CPU side, any part of your business?

Yes, it's a dynamic world right now. I think when you look at the memory market, first of all, we plan with the memory vendors many years in advance. So we've been planning for the MI450 ramp. We're planning our HBM4 ramp across the memory ecosystem. So we feel good about where we're positioned from an HBM standpoint. There are other knock-on effects on the memory market right now. Certainly, if you talk to any of the memory vendors in terms of where DDR4 and DDR5 are positioned as well as some of the consumer grades.

The impact that we're seeing is certainly the memory prices are affecting system prices. So you see system prices going up. I will say that the enterprise demand on the data center side seems, again, very durable. I think people are wanting the compute, needing the compute. So although they're paying a bit more than they might have 6 or 9 months ago, I think that is the main impact. I am watching the impact on the PC market. So we would expect that there might be more sort of cost pressures.

And as those cost pressures, they may change a little bit the PC market dynamics. So whereas our overall sell-through in the PC market is actually quite good. We are expecting that in the second half of the year, we may see a more muted part of the market just as memory prices are volatile. But we'll have to see how it plays out. I mean at the end of the day, -- the one thing about the industry is we tend to like demand, and we tend to like fulfilling that demand. So I think there is a lot to still play out. But on the data center side, it is full speed ahead, and we'll have to see what happens in the consumer markets.

Very helpful. Let me see if we have any questions from the audience.

Following OpenAI and Meta, can you just talk about the propensity for other kind of gigawatt scale deals with other hyperscalers, AI lab types?

Yes. Absolutely. So look, we are very ambitious with what we can do in the data center AI market. I think from a road map standpoint, as much as we're excited about the MI450 series, we're actually super excited about what's beyond as well. These days, I think there are multiple gigawatt scale customers. I think every lab is looking for choice at this point. The -- if you -- this idea of diversity of compute is important.

So with Meta and OpenAI, I think we've built foundational multi-generational deals that will absolutely help pull the entire ecosystem and enhance the entire AMD ecosystem. I think there are a number of other customers that are, let's call it, in that scale that we see as very strongly interested in MI450 and beyond. Back to this comment of what are we trying to do. The -- if you think about the ambition that we have in the data center AI segment, it is a very, very large TAM, and we are currently at the very early stages of building out our business. What we're trying to do is accelerate it. So we've talked about growing over 80% CAGR over the next 3 to 5 years in our data center AI segment. I think with the visibility that we have with some of these large deals as well as the other -- the broader customer set, I think we have a very good confidence to not only meet those but exceed those targets as we go forward.

Can you comment on the status of the Chinese market opportunity and also maybe also competition from China?

Sure. So we've always stated that the Chinese market is an important market to us. We are broadly -- we have a broad set of customers in the Chinese market on CPUs in other areas. On the GPU side, it is still a little bit complicated. We were able to ship some MI308s last quarter in the fourth quarter that we reported, and we talked about approximately $100 million this quarter. We're in the process of applying for licenses for the next generation of the MI325 chips.

I think the Commerce Department and the U.S. government are still going through the approval processes for that. So it's very, very hard to predict. And for that reason, we're not forecasting additional revenue going forward. We would certainly like to satisfy our customers in China. I think there's a lot that we learned by participating in the Chinese market because they also have a set of models that are somewhat different than the U.S. models, and we want to be able to service that, but we'll have to see how that whole environment plays out from a licensing standpoint over the next couple of months.

And do those limitations kickstart Chinese competition?

Well, I think Chinese competition was always going to be fierce. So we should expect that in a world as competitive as AI, we have to give the Chinese chip providers credit for what they're able to achieve. That being said, I think the road maps that we have from a U.S. technology standpoint are very, very strong. We want to be able to participate in the global market, and we need to continue to work with both governments to enable that to happen.

Last question.

There seems to be a debate out there as to whether you're able to ship rack-scale solutions in volume in the second half of this year? Or do you have enough -- I mean, you have enough CoWoS capacity to do that? Or is this much more of a 2027 story?

We definitely have enough CoWoS capacity. I know that there's lots of people trying to check various things. The best thing I can tell you is we have the capacity. We have the technology. We have the deep customer relationships. We have the data center providers have allocated space for it. So we have to execute that ramp.

And we've always said the ramp is second half weighted. Think about it a little bit in Q3, but really ramping sharply as we get into Q4. This is no question, a very, very important ramp for us. It's something that we've been planning for, for many quarters, and we feel good about the ramp.

Great. Well, we'll wrap it up there, Lisa. Congratulations on everything you achieved, and thank you so much for being here today.

Thank you.

Greetings, and welcome to the AMD Conference Call. [Operator Instructions] As a reminder, this conference is being recorded.

I would now like to turn the conference over to your host, Matt Ramsay, Vice President, Financial Strategy and Investor Relations. Thank you. You may begin.

Good morning. Thank you for joining on such short notice, and welcome to our call to discuss the significant new AI partnership between AMD and Meta. By now, you should have had the opportunity to review a copy of our press release and our Form 8-K filing discussing this announcement. If you have not had the opportunity to review these materials, they can be found on the Investor Relations page of amd.com.

Participants on today's conference call are Dr. Lisa Su, our Chair and CEO; and Jean Hu, our Executive Vice President, CFO and Treasurer. This is a live call and will be replayed via webcast on our website.

Today's discussion may contain forward-looking statements based on our current beliefs, assumptions and expectations, speak only as of today and as such, involve risks and uncertainties that could cause actual results to differ materially from our expectations. Please refer to the cautionary statement in our press release for more information on the factors that could cause these actual results to differ materially.

With that, I'll hand the call to Lisa.

Great. Thank you, Matt, and good morning, and thank you all for joining the call today. We are announcing a significant expansion of our strategic partnership with Meta, including a new multi-year multigeneration agreement that positions AMD at the core of their next-generation AI infrastructure.

2025 was a defining year for AMD with record results across the business, and we're carrying that momentum into 2026. AI demand is accelerating rapidly as customers scale modern AI infrastructure across both accelerated and general-purpose compute. Through the leadership technology and consistent execution, we have built one of the premier data center AI franchises in the industry, anchored by our differentiated Instinct GPU road map and our leadership EPYC CPU portfolio.

AMD is uniquely positioned to deliver high-performance, energy-efficient compute across the full spectrum of AI workloads. Meta has been a close partner over multiple generations, deploying millions of EPYC CPUs and hundreds of megawatts of MI300 and MI350 Series GPUs across their global infrastructure. Meta was also an early definition customer for our MI450 series, and we developed our Helios rack-scale architecture on the OCP Open Rack Wide standard in collaboration with Meta.

Today, we are significantly expanding our relationship. Under this agreement, Meta is expected to deploy 6 gigawatts of AMD Instinct GPUs across multiple product generations. To meet their evolving AI requirements, we are co-engineering a custom GPU accelerator based on our MI450 architecture optimized specifically for Meta's workloads. Initial shipments supporting the first gigawatt deployment are scheduled to begin in the second half of 2026 and will leverage our Helios rack-scale architecture with the custom MI450-based Instinct GPU and our 6th Gen EPYC CPU code named Venice.

This partnership firmly establishes AMD at the center of one of the industry's most significant AI infrastructure deployments and highlights the strength of our end-to-end platform strategy. As AI workloads scale, customers are increasingly looking for solutions tailored to their specific architectures and performance requirements. AMD's leadership in chiplet design and advanced packaging is a key differentiator and enables us to rapidly leverage core building blocks of our AI platform and tailor them for the optimal compute, memory and networking needs of specific customer workloads.

The custom MI450-based GPU we are developing with Meta is a direct result of this capability, delivering workload-specific optimizations while leveraging the MI450 platform, the Helios rack-scale system infrastructure and our open ROCm software ecosystem, giving Meta the advantages of a custom solution with the benefits of the broader MI450 ecosystem and GPU programmability.

As these platforms deploy at gigawatt scale, the ecosystem optimizations across ROCm, AI frameworks and system software will extend well beyond this engagement, strengthening our broader Instinct franchise and expanding opportunities across our entire customer base.

In addition to expanding our GPU engagement with Meta, we are further deepening our EPYC CPU partnership. We are seeing accelerating CPU compute demand driven by the rapid scaling of AI infrastructure across model development, inferencing, data processing and the rise of agentic AI. As deployments grow in scale and complexity, CPUs remain a strategic foundation of the compute stack, driving efficiency, orchestration and system-level performance. EPYC is well positioned to capture outsized value in this next phase of AI expansion.

Meta is already a multi-generation EPYC customer with EPYC processors powering the majority of core services across their global data center footprint. Building on our deep road map alignment, Meta will be a lead customer for our 6th Gen EPYC Venice processor at launch later this year.

We have also partnered closely on a new addition to our Zen 6 family code named Verano, incorporating workload-specific optimizations to deliver leadership performance per watt and compelling TCO. The expansion of our partnership across GPUs and CPUs is another strong proof point that the world's most ambitious AI builders are choosing AMD Instinct and EPYC platforms as the foundation of their AI infrastructure. I want to thank Mark and the entire Meta team for their collaboration and partnership. We are extremely proud to work together to advance the future of AI at scale.

Importantly, this engagement reflects the strong and growing demand we are seeing for our MI450 Series and Helios architecture. Overall, there is significant excitement in the market for the MI450 Series and Helios. And from an execution standpoint, we are making excellent progress. MI450 and Helios are currently in hardware and software validation, running the latest inference and training workloads.

We are working closely with our lead customers, supply chain and ecosystem partners to ensure a smooth ramp. We expect to begin customer sampling shortly and remain on track to begin production shipments of both the standard MI450 series and the custom MI450-based CPU for Meta in the second half of 2026.

Now I'll turn the call over to Jean to provide additional details on the agreement.

Thank you, Lisa. Today's announcement of a 6-gigawatt agreement with Meta is another significant step in scaling our data center AI business, consistent with the ambitious plan we set out at our Financial Analyst Day.

Let me provide some context on the financials. The Meta deployment is expected to generate data center AI revenue of significant double-digit billions of dollars per gigawatt. Revenue will begin in the second half of 2026 and ramp alongside our MI450 deployment with other customers.

As part of the agreement, to further align strategic interests of both companies, AMD has issued Meta a warrant for up to 160 million shares of AMD common stock. The warrant is performance-based. The first tranche vests with the initial 1-gigawatt of shipments of AMD Instinct GPU, with additional tranches vesting as Meta's purchases of Instinct GPUs to scale to 6 gigawatts. Vesting is further tied to AMD, achieving certain stock price thresholds with the final tranches vesting at a price of $600 per share.

In addition, exercise of the warrant is tied to Meta achieving key technical and commercial milestones. The unique structure aligns Meta and AMD, driving significant long-term revenue growth and is accretive to our non-GAAP earnings per share, while enabling Meta to share directly in the upside of our mutual success. This partnership marks another significant step forward in delivering our ambitious long-term financial model, including greater than 80% CAGR of our data center AI business and generating more than $20 in annual earnings per share within the next 3 to 5 years.

Let me now turn the call back to Lisa.

Thank you, Jean. Let me close by saying that today represents another major milestone for AMD's AI strategy. Through this multiyear multi-generation agreement to deploy 6 gigawatts of AMD Instinct GPUs, we are significantly expanding our partnership with Meta, broadening our AI footprint and deepening our co-development and road map alignment with one of the world's leading AI companies that is building at massive scale.

We believe that the scale of this deployment and the ecosystem benefits that it drives will further strengthen our AI platforms and expand opportunities across both existing and new customers. The current AI infrastructure build-out is one of the most significant technology investment cycles in decades. The expanded partnership with Meta, together with our previously announced partnerships with OpenAI, Oracle and others, demonstrates the strength of our multi-generation Instinct and Helios road maps and firmly establishes AMD at the center of this next phase of AI growth.

From silicon to systems to software, we are executing with scale, speed and discipline, leveraging our differentiated end-to-end portfolio and deep strategic partnerships to capture the AI opportunity, accelerate sustained data center growth and deliver long-term shareholder value.

Now I'll turn the call back to Matt for Q&A.

Thank you, Lisa. Thank you, Jean. For today's Q&A session, please focus your questions on today's announcement. [Operator Instructions]

Thank you very much. Operator, you can please poll for the first question.

[Operator Instructions] Our first question comes from the line of Joshua Buchalter with TD Cowen.

Congrats on the deal. For my first one, I just wanted to hit one that's gotten in my inbox a couple of times. It seems like it wouldn't be the case since Meta doesn't have a third-party cloud business like the other hyperscalers, but could you confirm whether or not there's any overlap between this deal and the OpenAI deal you announced a few months ago? And are there any details you're able to share on how the 6 gigawatts compares to the OpenAI deal from a timing and value standpoint?

Yes, absolutely, Josh. So thanks for the question. If I give you the overall context, look, I think Meta's plans are extremely ambitious in terms of what they're trying to do in terms of AI infrastructure build-out. We are super excited about this partnership. Meta has been a long-standing partner of AMD over the last few generations, but this agreement really takes our relationship to the next level and significantly expands on what we were doing before.

In terms of your question about whether there's any overlap, no, there's no overlap from the standpoint that Meta's -- this agreement is really for Meta's workloads. Our -- the custom GPU that we're building is optimized for Meta's workloads. It's the first time we are doing a custom GPU of this style using our MI450-based architecture.

So I think, look, both OpenAI and Meta are incredible companies. They're doing tremendous innovation, and we are super happy to be deeply partnered with them. And what we are really trying to do, which we've always been is if you think about the AI space, there are lots of different workloads and workloads of different types. And what we want to do is partner deeply so that we're providing the technology that each company needs to satisfy their ambitions.

Matt, are we doing follow-ups?

No, we can just try to keep this to one question per caller given we got a [indiscernible] call. But thank you, Josh. Operator, the next caller please.

Our next question comes from the line of Vivek Arya with Bank of America.

Lisa, I had a more fundamental question, which is where is the economic value add here? You're giving away $30 billion roughly of value in your stock. And in return, you're getting about $30 billion, $35 billion of net income. So it seems like an even swap. So where is the value add here? So if I ask it in a different way, if the product is so good, why does AMD need to give up 10% of your equity? And is now every MI450 customer going to ask for this kind of a deal?

Yes, sure, Vivek. So let me -- let me start with some context. If you look at the context of these partnerships, every deal of this scale is very unique. We -- as Jean said, this deal is very accretive to AMD's earnings. So it is a very good deal for AMD shareholders. But what we're setting out to create is a strategic partnership where we're going much, much deeper together. And I view this as a very transformational partnership from an AMD standpoint. Meta is operating at gigawatt scale. We are actually working with them deeply in terms of their technology road map, and we're working together on hardware, software systems.

And with that, it accrues significant value to our overall road map going forward. And I think you have to look beyond just these particular economics, which, as I said, are very accretive to our earnings, but you also have to look at what it's doing across our entire road map and our future road map going forward.

So if you look at the structure of our warrants in this case, is -- again, it's a very aligned incentive structure. Meta is making a big bet on deploying at large scale for AMD, which is great. AMD benefits from this large-scale deployment, which brings revenue scale, ecosystem maturity, software maturity. And assuming that we satisfy all of the purchases as well as the share price thresholds, AMD shareholders will benefit significantly, and Meta gets to benefit as part of that.

So to your question, no, I don't think every MI450 customer is of this size and scale. Our standpoint is we are looking for defining partnerships for our AI franchise, and Meta is one of those defining partnerships.

Our next question comes from the line of Timothy Arcuri with UBS.

Lisa, I wanted to ask a question along the same line. So it sounds like the exact same deal that OpenAI got right down to the same kind of stock prices as well. So it is a lot of revenue. But speaking to the precedent here, would Meta have done this without this deal. So when they saw the deal from OpenAI, did they say, okay, I'll do a deal if I get the same deal. So I guess speaking to the precedent if you're going to sign with Amazon over somebody else, does this sort of open up the discussion that they would want the same kind of deal.

Tim, I think the way -- if I give you some sort of background on how we came up with this -- this wasn't about, hey, are we trying to do something similar with other customers. It's not like that.

What it is about is Meta has been a tremendous partner for AMD over the last several generations, and we appreciate that, by the way. I mean they have been a big adopter of EPYC. They were an early adopter of MI300 and MI350. But without this strategic agreement, I think we would have done well. I think MI450 would have done well.

But what we are looking to do is do something transformational. And when you talk about gigawatt scale deployments and 6 gigawatts over 5 years, that is transformational in terms of where we see our business. And in addition to that, they're at the forefront of what's happening with models and model builders, they are optimizing workloads for their future, and we are optimized alongside with them.

So I think if you look in that context, they are different deals, but they are very important strategic deals in terms of the shape of AI going forward. And if you think about this market, there's probably only a handful of companies that are deploying at this scale. And so to have Meta today and OpenAI as well as strategic partners that anchor AMD's AI strategy, I think, is a really, really good place for us to be.

Our next question comes from the line of Blayne Curtis with Jefferies.

Congrats. Lisa, I just wanted to kind of dive in a little bit more on the economics. Should we think about this deal, you've previously talked about a range of like dollars per gigawatt and maybe OpenAI was at the lower end of that. Is that the same kind of economics? And then I'm just kind of curious in terms of support of these 2 deals, how do you think about the OpEx part of the equation?

Yes. So the way I would say it is we've talked about -- I think Jean might have mentioned that when we think about the revenue from, let's call it, the GPUs, we're talking about something like double-digit billions per gigawatt. And so that is the range that we're talking about. When you -- I'm sorry, the second part of your question?

OpEx [indiscernible]...

OpEx consideration to support of these 2 deal?

Yes, Blayne, on the OpEx side, because it is based on MI450, you can think about it's just another variant of MI450. So the incremental OpEx actually is quite minimal. That drives significant operating model leverage.

Our next question comes from the line of Aaron Rakers with Wells Fargo.

Yes. In the prepared comments, obviously, this is all about the Instinct GPU road map, but you've also emphasized the importance that the CPU plays in these architecture deployments going forward, particularly probably around AI inferencing.

I'm curious, Lisa, if you could talk a little bit about how you see the evolving competitive landscape in CPUs, in particular, one of your key competitors trying to push Arm-based architecture more prolifically on a stand-alone basis. I'm curious how you see that playing out in these architecture deployments.

Sure, Aaron. So thanks for the question. I think what we've said and what we said on the last earnings call and the last couple of earnings calls is the CPU market is absolutely on fire. I mean there is significant demand. It has continued to grow, and it really is a result of the AI infrastructure deployments as inferencing scales, as agentic AI scales. And our EPYC portfolio is in an extremely good position. I mean everything that we see, certainly with our current Zen 5 class [ Trent ] products, we are very widely deployed in AI infrastructure.

And as we go into our Zen 6 family with Venice and the addition of Verano, we actually see our workload coverage increasing across a number of our largest customers. So I think it's great that there's so much interest around CPUs. I think that is something that we've always believed that you need all different types of compute. And we feel we're very, very well positioned from a competitive standpoint.

Our next question comes from the line of Tom O'Malley with Barclays.

Congrats on the deal. Lisa, I wanted to dive in on the custom commentary from the release and then also in your answer to one of the questions earlier. What does it mean specifically about a custom design? Is it just a different flavor of a tape-out? Is the system architecture going to look a little bit different? And then in the future, are you going to be doing more custom-style tape-outs and/or systems with other customers that come on board? Just a little curious what that means specifically.

Yes, Tom, I think this is actually a pretty interesting new thing that we're doing here. Look, we've always believed that when you look at AI infrastructure, there's no one chip that does it all, especially when you look across training and inference and big models and small models and different workloads that you're trying to optimize.

So what's unique about this deal is we started with the workload first. We didn't start with the chip. We started with the workload. What is most important to Meta for their future workloads, their highest volume workloads. And then we worked back from that with our chiplet architecture. What's unique about our chiplet architecture is we have all the building block pieces, but you can put them together and configure them in different ways to give you sort of different performance and system characteristics.

And so what we've done together with Meta is we're using our chiplet architecture to come up with a new variant based on the MI450 architecture. As Jean said, it's highly leveraging the base capability, which gives us both development scale as well as, frankly, a lot of leverage as we're bringing up these technologies. But it's not just the chip optimization. I think it's chip level, board level, system level, and that comes together in a solution that I think gives Meta is the best of both worlds, which is something that is highly tuned to their workload, but takes advantage of the entire infrastructure and supply chain and everything else that we are developing for the base MI450 architecture.

So to your question of do we expect to do more? I would expect that for high-volume workloads, there will be benefits to doing something that is, let's call it, more customized in a GPU format. Note that we're not doing full ASIC, which, again, usually takes a lot more time, but we're doing something based off of our foundational architecture. So I think it's a good opportunity to expand our portfolio as customers are increasing their volumes.

Tom, just added to what Lisa said, there's no additional tape-out needed for this customer chip.

Our next question comes from the line of Antoine Chkaiban with New Street Research.

Actually, I'd like to follow up on the prior question. I'm wondering, like given this is a custom deployment, how much of Meta's software investment in ROCm is truly transferable versus custom? And maybe if I ask the question differently, does this deal really create a self-reinforcing ecosystem that makes subsequent AMD deployments progressively easier, as I imagine, custom hardware also means custom software.

Antoine, actually, it's extremely leverageable. So I don't know if I would say 100%, but 95-plus percent is the way you should think about it because all of the underlying software is using the MI450 architecture. So if you think about all of the work that we have to do in terms of the base libraries, the kernel optimizations, all of that stuff is highly leverageable to the rest of the AMD ecosystem.

And again, this is a GPU. So from a GPU standpoint, it's already highly programmable. So Meta has already been a very strong partner with us on the software ecosystem. If you look at all of their work with PyTorch and the open ecosystem. And I expect that we will continue to work closely together as we scale these deployments to gigawatts plus.

Our next question comes from the line of Mark Lipacis with Evercore ISI.

Lisa, can you help us understand what is -- what kind of visibility do you have under this deal? Like I'm trying to understand what hard orders do you have right now for the near term? And is there like a take-or-pay element on Meta's part here? Or like can they elect to opt out? Can you just help us understand like how hard are the orders? I appreciate you anticipate getting the full 6 gigawatts. But if you could help us understand the near-term visibility and the longer-term visibility.

Sure, Mark. So I think as we said, we are signing a long-term strategic agreement that goes across 5 years for 6 gigawatts. The first gigawatt is committed and will start shipments in the second half of 2026. As you know, the supply chain overall is tight, and so we're planning it very tightly together for their data center builds in terms of which data centers these things are going in.

But I think we have a very good near-term visibility. The important thing I want to mention, again, Mark, is we had already a very strong relationship with Meta. But what this agreement does is it really does take our near-term work together to the next level. And that has been very, very positive. And I really feel that it is a big addition to what we see in terms of MI450 adoption to include the custom GPU optimized for Meta's workload.

So in terms of long-term, we are actively working on beyond MI450 as well. As you know, we've already been well into the development of MI500 and beyond. And I expect with each generation, we can get even more optimized as we learn more about their workloads and they learn more about our architecture.

Very helpful.

Thanks.

Our next question comes from the line of Harsh Kumar with Piper Sandler.

I was curious about how the deal was won, especially the custom GPU part. This is not the norm for you guys. So was this sort of a bake-off against some other people? Or did Meta approach you specifically with an idea of custom chip? I'd be curious how this played out.

Harsh, I think the best way to say it is we're always in deep discussions with Meta. They are a very close partner. What has been very good about this engagement is Mark and I have spent a good amount of time together on how we really align our road map with their next-generation infrastructure. I think Mark is extremely ambitious with what he wants to do with Meta in terms of what they're doing with their models. I think they have a set of requirements -- he has a great team that works with him, and we have been working closely with them on how can we both optimize as well as expand our relationships.

So from that standpoint, we started with workload first, like what are you trying to accomplish? And we came up with ideas for how we could even enhance the base standard product of MI450 to broaden our workload adoption. So it was a very collaborative effort and one that I think talks about the types of bets that companies are making. Like we are making certainly a big bet on Meta, and I think Meta is making a big bet on AMD.

Operator, I think we have time for just one more question before we end today's call. Thank you.

Our last question today will come from the line of Srini Pajjuri with RBC Capital Markets.

Lisa, I just want to clarify, you gave us some targets at the Analyst Day last year, $20-plus EPS. Just curious if those numbers already included the Meta deal. And then you also talked about potential other customers, and obviously, Meta is one of them. I'm just curious if that $20 number includes any other future deals as well.

Yes. So going back to our financial model, I think we put some very ambitious goals out there in terms of our revenue as well as our EPS and $20 EPS. The -- when we came up with that number in November, I wouldn't say we had this Meta deal specifically baked in. This was still very, very much in the works. This does expand our relationship with Meta, which is a great thing. And when we look at these financial models, we have a broad set of customers and a broad set of customers that we are actively engaging.

So I think what this should give you is clearer visibility into how we intend to both achieve and exceed our financial model, having a great strategic partnership like Meta as well as some of the other partnerships that we have talked about. And you should imagine that the overall interest in MI450 is very high. And in addition to the partnerships that we're talking about right now, there are a number of other strategic partnerships that are underway. And I feel very good about our trajectory towards that long-term financial model.

Thank you very much, everyone. I think this is the end of the Q&A period. We really appreciate everybody jumping on, on short notice. It's an exciting day for AMD, and thank you for your interest and your time.

Operator, please go ahead and close the call. Thank you.

Thank you. This concludes today's conference call. You may disconnect your lines at this time. Thank you for your participation.

Greetings, and welcome to the AMD Fourth Quarter and Full Year 2025 Conference Call. [Operator Instructions]. And please note that this conference is being recorded.

I will now turn the conference over to Matt Ramsay, VP of Financial Strategy and IR. Thank you. You may begin.

Thank you, and welcome to AMD's Fourth Quarter and 2025 Full Year Financial Results Conference Call. By now, you should have had the opportunity to review a copy of our earnings press release and accompanying slides. If you have not had the opportunity to review these materials, they can be found on the Investor Relations page of amd.com.

Today, we will refer primarily to non-GAAP financial measures on the call. The full non-GAAP to GAAP reconciliations are available in today's press release and in the slides posted on our website. Participants in today's conference call are Dr. Lisa Su, our Chair and CEO; and Jean Hu, our Executive Vice President, CFO and Treasurer. This is a live call and will be replayed via webcast on our website.

Before we begin, I would like to note that Mark Papermaster, Executive Vice President and CTO, will present at Morgan Stanley's TMT Conference on Tuesday, March 3. Today's discussions contain forward-looking statements based on our current beliefs, assumptions and expectations, speak only as of today and as such, involve risks and uncertainties that could cause actual results to differ materially from our current expectations. Please refer to the cautionary statement in our press release for more information on factors that could cause actual results to differ materially.

With that, I will hand the call to Lisa.

Thank you, Matt, and good afternoon to all those listening today. 2025 was a defining year for AMD with record revenue, net income and free cash flow driven by broad-based demand for our high-performance computing and AI products. We ended the year with significant momentum with every part of our business performing very well. We saw demand accelerate across the data center, PC, gaming and embedded markets, launched the broadest set of leadership products in our history, gained significant server and PC processor share and rapidly scaled our Data Center AI business as Instinct and ROCm adoption increased with cloud, enterprise and AI customers.

Looking at our fourth quarter, fourth quarter revenue grew 34% year-over-year to $10.3 billion, led by record EPYC, Ryzen and Instinct processor sales. Net income increased 42% to a record $2.5 billion and free cash flow nearly doubled year-over-year to a record $2.1 billion. For the full year, revenue grew 34% to $34.6 billion, and we added more than $7.6 billion of Data Center segment and Client revenue.

Turning to our fourth quarter segment results. Data Center segment revenue increased 39% year-over-year to a record $5.4 billion, led by accelerating Instinct MI350 Series GPU deployments and server share gains. In server, adoption of fifth gen EPYC turn CPUs accelerated in the quarter, accounting for more than half of the total server revenue. Fourth gen EPYC sales were also robust as our prior generation CPUs continue to deliver superior performance and TCO compared to competitive offerings across a wide range of workloads. As a result, we had record server CPU sales to both cloud and enterprise customers in the quarter and exited the year with record share.

In cloud, hyperscaler demand was very strong as North American customers expanded deployments. EPYC-powered public cloud offerings grew significantly in the quarter with AWS, Google and others launching more than 230 new AMD instances. Hyperscalers launched more than 500 AMD-based instances in 2025, increasing the number of EPYC cloud instances more than 50% year-over-year to nearly 1,600.

In the enterprise, we are seeing a meaningful shift in EPYC adoption, driven by our leadership performance, expanded platform availability, broad software enablement and increased go-to-market programs. The leading server providers now offer more than 3,000 solutions powered by fourth and fifth gen EPYC CPUs that are optimized for all major enterprise workloads. As a result, the number of large businesses deploying EPYC on-prem more than doubled in 2025, and we exited the year with record server sell-through.

Looking ahead, server CPU demand remains very strong. Hyperscalers are expanding their infrastructure to meet growing demand for cloud services and AI while enterprises are modernizing their data centers to ensure they have the right compute required to enable new AI workflows. Against this backdrop, EPYC has become the processor of choice for the modern data center, delivering leadership performance, efficiency and TCO. Our next-generation Venice CPU extends our leadership across each of these metrics. Customer pull for Venice is very high with engagements underway to support large-scale cloud deployments and broad OEM platform availability when Venice launches later this year.

Turning to our Data Center AI business. We delivered record Instinct GPU revenue in the fourth quarter, led by the ramp of MI 350 Series shipments. We also had some revenue from MI308 sales to customers in China. Instinct adoption broadened in the quarter. Today, 8 of the top 10 AI companies use Instinct to power production workloads across a growing range of use cases. With the MI350 series, we are entering the next phase of instinct adoption, expanding our footprint with existing partners and adding new customers.

In the fourth quarter, hyperscalers expanded MI350 Series availability, leading AI companies scale their deployments to support additional workloads and multiple neocloud providers launched MI350 Series offerings that deliver on-demand access to instinct infrastructure in the cloud.

Turning to our AI software stack. We expanded the ROCm ecosystem in the fourth quarter, enabling customers to deploy Instinct faster and with higher performance across a broader range of workloads. Millions of large language and multimodal models run out of the box on AMD with the leading models launching with day zero support for Instinct GPUs. This capability highlights our rapidly expanding open source community enablement, including new upstream integration of AMD GPUs in DLLM, one of the most widely used inference engines.

To drive Instinct adoption with industry-specific use cases, we're also adding support for domain-specific models in key verticals. As one example, in health care, we added ROCm Support for the leading medical imaging framework to enable developers to train and deploy highly performing deep learning models on Instinct GPUs. For large businesses, we introduced our enterprise AI suite, a full stack software platform with enterprise-grade tools, inference microservices and solutions blueprints designed to simplify and accelerate production deployments at scale.

We also announced a strategic partnership with Tata Consultancy Services to co-develop industry-specific AI solutions and help customers deploy AI across their operations. Looking ahead, customer engagements for our next-gen MI400 series and Helios platform continue expanding. In addition to our multi-generation partnership with OpenAI to deploy 6 gigawatts of Instinct GPUs, we are in active discussions with other customers on at-scale multiyear deployments starting with Helios and MI450 later this year.

With the MI400 series, we are also expanding our portfolio to address the full range of cloud, HPC and enterprise AI workloads. This includes MI455X and Helios for AI superclusters, MI430X for HPC and sovereign AI and MI440X servers for enterprise customers requiring leadership training and inference performance in a compact eight-GPU solution that integrates easily into existing infrastructure.

Multiple OEMs publicly announced plans to launch Helios systems in 2026 with deep engineering engagements underway to support smooth production ramps. In December, HPE announced that they will offer Helios racks with purpose-built HPE Juniper Ethernet switches and optimized software for high-bandwidth scale-up networking. And in January, Lenovo announced plans to offer Helios racks. MI430X adoption also grew in the quarter with new exascale class supercomputers announced by GENCI in France and HLRS in Germany.

Looking further ahead, development of our next-generation MI500 series is well underway. MI500 is powered by our CDNA 6 architecture built on advanced 2-nanometer process technology and features high-speed HBM4E memory. We are on track to launch MI500 in 2027 and expect MI500 to deliver another major leap in AI performance to power the next wave of large-scale multimodal models.

In summary, our AI business is accelerating with the launch of MI400 series and Helios representing a major inflection point for the business as we deliver leadership performance and TCO at the chip, compute tray and rack level. Based on the strength of our EPYC and Instinct road maps, we are well positioned to grow Data Center segment revenue by more than 60% annually over the next 3 to 5 years and scale our AI business to tens of billions in annual revenue in 2027.

Turning to Client and Gaming. Segment revenue increased 37% year-over-year to $3.9 billion. In client, our PC processor business performed exceptionally well. Revenue increased 34% year-over-year to a record $3.1 billion, driven by increased demand for multiple generations of Ryzen desktop and mobile CPUs. Desktop CPU sales set a record for the fourth consecutive quarter. Ryzen CPUs topped the bestseller list at major global retailers and e-tailers throughout the holiday period with strong demand across all price points in every region, driving record desktop channel sell-out.

In mobile, strong demand for AMD-powered notebooks drove record Ryzen PC sell-through in the quarter. That momentum extended into commercial PCs, where Ryzen adoption accelerated as we established a new long-term growth engine for our client business. Sell-through of Ryzen CPUs for commercial notebooks and desktops grew by more than 40% year-over-year in the fourth quarter, and we closed large wins with major telecom, financial services, aerospace, automotive, energy and technology customers.

At CES, we expanded our Ryzen portfolio with CPUs that further extend our performance leadership. Our new Ryzen AI 400 mobile processors deliver significantly faster content creation and multitasking performance than the competition. Notebooks powered by Ryzen AI 400 are already available with the broadest lineup of AMD-based consumer and commercial AI PCs set to launch throughout the year. We also introduced our Ryzen AI Halo platform, the world's smallest AI development system, featuring our highest-end Ryzen AI MAX processor with 128 gigabytes of unified memory that can run models with up to 200 billion parameters locally.

In gaming, revenue increased 50% year-over-year to $843 million. Semi-custom sales increased year-over-year and declined sequentially as expected. For 2026, we expect semi-custom SoC annual revenue to decline by a significant double-digit percentage as we enter the seventh year of what has been a very strong console cycle.

From a product standpoint, Valve is on track to begin shipping, its AMD-powered steam machine early this year and development of Microsoft's next-gen Xbox featuring an AMD semi-custom SoC is progressing well to support a launch in 2027. Gaming GPU revenue also increased year-over-year with higher channel sell-out driven by demand throughout the holiday sales period for our latest generation Radeon RX 9000 series GPUs. We also launched FSR 4 Redstone in the quarter, our most advanced AI-powered upscaling technology, delivering higher image quality and smoother frame rates for gamers.

Turning to our Embedded segment. Revenue increased 3% year-over-year to $950 million, led by strength with test and measurement and aerospace customers and growing adoption of our Embedded x86 CPUs. Channel sell-through accelerated in the quarter as end customer demand improved across several end markets, led by test, measurement and emulation. Design win momentum remains one of the clearest indicators of long-term growth for our embedded business, and we delivered another record year. We closed $17 billion in design wins in 2025, up nearly 20% year-over-year as we've now won more than $50 billion of embedded designs since acquiring Xilinx.

We also strengthened our Embedded portfolio in the quarter. We began production of our Versal AI Edge Gen 2 SoCs for low-latency inference workloads and started shipping our highest-end Spartan UltraScale+ devices for cost-optimized application. We also launched new embedded CPUs, including our EPYC 2005 series for network security and industrial edge applications, Ryzen P100 series for in-vehicle infotainment and industrial systems and Ryzen X100 series for physical AI and autonomous platforms.

In summary, 2025 was an excellent year for AMD, marking the start of a new growth trajectory for the company. We are entering a multiyear demand super cycle for high performance and AI computing that is creating significant growth opportunities across each of our businesses. AMD is well positioned to capture that growth with highly differentiated products, a proven execution engine, deep customer partnerships and significant operational scale.

And as AI reshapes the compute landscape, we have the breadth of solutions and partnerships required for end-to-end leadership from Helios in the cloud for at-scale training and inference to an expanded instinct portfolio for sovereign, supercomputing and enterprise AI deployment. At the same time, demand for EPYC CPUs is surging as Agentic and emerging AI workloads require high-performance CPUs to power head nodes and run parallel tasks alongside GPUs. And at the edge and in PCs where AI adoption is just beginning, our industry-leading Ryzen and Embedded processors are powering real-time on-device AI. As a result, we expect significant top line and bottom line growth in 2026, led by increased adoption of EPYC and Instinct, continued client share gains and a return to growth in our Embedded segment.

Looking further ahead, we see a clear path to achieve the ambitious targets we laid out at our Financial Analyst Day last November, including growing revenue at greater than 35% CAGR over the next 3 to 5 years, significantly expanding operating margins and generating annual EPS of more than $20 in the strategic time frame, driven by growth in all of our segments and the rapid scaling of our Data Center AI business.

Now I'll turn the call over to Jean to provide additional color on our fourth quarter results and full year results. Jean?

Thank you, Lisa, and good afternoon, everyone. I'll start with a review of our financial results and then provide our current outlook for the first quarter of fiscal 2026.

AMD executed very well in 2025, delivering record revenue of $34.6 billion, up 34% year-over-year, driven by 32% growth in our Data Center segment and 51% growth in our Client and Gaming segment. Gross margin was 52%, and we delivered record earnings per share of $4.17, up 26% year-over-year while continuing to invest aggressively in AI and the data center to support our long-term growth.

For the fourth quarter of 2025, revenue was a record $10.3 billion, growing 34% year-over-year, driven by strong growth in the Data Center and Client and Gaming segment, including approximately $390 million in revenue from MI308 sales to China, which was not included in our fourth quarter guidance. Revenue was up 11% sequentially, primarily driven by continued strong growth in Data Center from both server and data center AI business as well as a return to year-over-year growth in the Embedded segment.

Gross margin was 57%, up 290 basis points year-over-year. We benefited from the release of $360 million in previously writing down MI308 inventory reserves. Excluding the inventory reserve release and MI308 revenue from China, gross margin would have been approximately 55%, up 80 basis points year-over-year, driven by favorable product mix.

Operating expenses were $3 billion, an increase of 42% year-over-year as we continue to invest in R&D go-to-market activities to support our AI road map and long-term growth opportunities as well as higher employee performance-based incentives. Operating income was a record $2.9 billion, representing a 28% operating margin. Tax, interest and other resulted in a net expense of approximately $335 million.

For the fourth quarter, diluted earnings per share was a record $1.53, an increase of 40% year-over-year, reflecting strong execution and operating leverage in our business model.

Now turning to our reportable segment. Starting with the Data Center segment. Revenue was a record $5.4 billion, up 39% year-over-year and 24% sequentially, driven by strong demand for EPYC processors and the continued ramp of MI350 products. Data Center segment operating income was $1.8 billion or 33% of revenue compared to $1.2 billion or 30% a year ago, reflecting higher revenue and inventory reserve release, partially offset by continued investment to support our AI hardware and software road maps.

Client Gaming segment revenue was $3.9 billion, up 37% year-over-year, driven primarily by strong demand for our leadership AMD Ryzen processors. On a sequential basis, revenue was down 3% due to lower semi customer revenue. The client business revenue was a record $3.1 billion, up 34% year-over-year and 13% sequentially, led by strong demand from both the channel and the PC OEMs and continued market share gains.

The gaming business revenue was $843 million, up 50% year-over-year, primarily driven by higher semi customer revenue and strong demand for AMD Radeon GPUs. Sequentially, gaming revenue was down 35% due to lower semi customer sales. Client and Gaming segment operating income was $725 million or 18% of revenue compared to $496 million or 17% a year ago.

Embedded segment revenue was $950 million, up 3% year-over-year and 11% sequentially as demand strengthened across several end markets. Embedded segment operating income was $357 million or 38% of revenue compared to $362 million or 39% a year ago.

Before I review the balance sheet and cash flow, as a reminder, we closed the sale of ZT Systems manufacturing business to Sanmina in late October. The fourth quarter financial results of the ZT manufacturing business are reported separately in our financial statement as discontinued operations and are excluded from our non-GAAP financials.

Turning to the balance sheet and cash flow. During the quarter, we generated a record $2.3 billion in cash from continuing operations and a record of $2.1 billion in free cash flow. Inventory increased sequentially by approximately $607 million to $7.9 billion to support strong data center demand. At the end of the quarter, cash, cash equivalents and short-term investments were $10.6 billion.

For the year, we repurchased 12.4 million shares and returned $1.3 billion to shareholders. We ended the year with $9.4 billion authorization remaining under our share repurchase program.

Now turning to our first quarter 2026 outlook. We expect revenue to be approximately $9.8 billion, plus or minus $300 million, including approximately $100 million of MI308 sales to China. At the midpoint of our guidance, revenue is expected to be up 32% year-over-year, driven by strong growth in our Data Center and Client and Gaming segments and modest growth in our Embedded segment.

Sequentially, we expect revenue to be down approximately 5%, driven by seasonal decline in our Client Gaming and Embedded segment, partially offset by growth in our Data Center segment. In addition, we expect fourth quarter non-GAAP gross margin to be approximately 55% non-GAAP operating expense to be approximately $3.05 billion. Non-GAAP other net income to be approximately $35 million, non-GAAP effective tax rate to be 13% and diluted share count is expected to be approximately 1.65 billion shares.

In closing, 2025 was an outstanding year for AMD, reflecting disciplined execution across the business to deliver strong revenue growth, increased profitability and cash generation while investing aggressively in AI and innovation to support our long-term growth strategy. Looking ahead, we are very well positioned for continued strong top line revenue growth and earnings expansion in 2026 with a focus on driving data center AI growth, operating leverage and delivering long-term value to shareholders.

With that, I'll turn it back to Matt for the Q&A session.

Yes. Thank you very much, Jean. Operator, please go ahead and open the Q&A session. Thank you.

[Operator Instructions] And the first question comes from the line of Aaron Rakers with Wells Fargo.

Lisa, at your Analyst Day back in November, you seem to kind of endorse the high $20 billion AI revenue expectation that was out there on the Street for 2027. I know today, you're reaffirming the path to strong double-digit growth. So I guess my question is, can you talk a little bit about what you've seen as far as customer engagements, how those might have expanded? I think you've alluded to in the past, multiple multi-gigawatt opportunities. Just any -- just double-click on what you've seen for the MI455 and Helios platform from a demands shaping perspective as we look into the back half of the year?

Yes. Sure, Aaron. Thanks for the question. So first of all, I think the MI450 Series development is going extremely well. So we're very happy with the progress that we have. We're right on track for a second half launch and beginning of production. And as it relates to sort of the shape of the ramp and the customer engagements, I would say the customer engagements continue to proceed very well. We have obviously a very strong relationship with OpenAI, and we're planning that ramp starting in the second half of the year going into 2027. That is on track.

We're also working closely with a number of other customers who are very interested in ramping MI450 quickly, just given the strength of the product, and we see that across both inference and training. And that is the opportunity that we see in front of us. So we feel very good about sort of the data center growth overall for us in 2026 and then certainly going into 2027, we've talked about tens of billions of dollars of data center AI revenue, and we feel very good about that.

The next question comes from the line of Tim Arcuri with UBS.

Jean, I'm wondering if you can maybe give us a little bit of detail under the hood for the March guidance. I know you basically told us that -- you told us about what Embedded is going to be up a bit year-over-year. Client sounds like it's down seasonally, which I take to be maybe down 10%. So can you give us a sense maybe of the other pieces?

And then also, can you give us a sense of how Data Center GPU is going to ramp through the year? I know it's a back half loaded year, but I think people are thinking at least somewhere in the $14 billion range this year. That's what investors are thinking. I don't -- I'm not asking you to endorse that. But if you can give us a little flavor for sort of how the ramp will look for the year, that would be great.

Tim, thanks for your question. We're guiding one quarter at a time, but I can give you some color about our Q1 guide. First is right, sequentially, we guided a decline around 5%, but Data Center is actually going to be up. And when you think about it, right, our CPU business seasonal actually in a regular seasonal pattern, it's going to be down high single digit. And in our current guide, we actually guide CPU revenue up sequentially very nicely.

Also with the Data Center GPU side, we also feel really good about GPU revenue, including China, will be also up. So very nice guide for the Data Center overall. On the Client side, we do see seasonality sequentially decline. Embedded and Gaming, they also have a seasonal decline.

And maybe, Tim, if I just give you a little bit on the full year commentary. I think the important thing, as we look at the full year, we're very bullish on the year. We're not -- if you look at the key themes, we're seeing very strong growth in the Data Center, and that's across 2 growth vectors. We see server CPU growth actually very strong. I mean we've talked about the fact that CPUs are very important as AI continues to ramp. And we've seen the CPU order book continue to strengthen as we go through the last few quarters and especially over the last 60 days.

So we see that as a strong growth driver for us. As Jean said, we see server CPU growing from Q4 into Q1 in what normally is seasonally down, and that continues throughout the year. And then on the Data Center AI side, it's a very important year for us. It's really an inflection point. MI355 has done well, and we were pleased with the performance in Q4, and we continue to ramp that in the first half of the year. But as we get into the second half of the year, the MI450 is really an inflection point for us. So that revenue will start in the third quarter, but it will ramp significant volume in the fourth quarter as we get into 2027. So that gives you a little bit of sort of what the data center ramp looks like throughout the year.

And the next question comes from the line of Vivek Arya with Bank of America.

First, just a clarification on what you're assuming for your China MI308 sales beyond Q1. And then Lisa, specific to 2026, can your Data Center revenues grow at your target 60% plus growth rate? I realize that, that's a multiyear target, but do you think that there are enough drivers, whether it's on the server CPU side or GPU side for you to grow at that target base even in 2026?

Yes. Sure, Vivek. So let me talk a little bit about China first because that's, I think, important for us to make sure that's clear. Look, we were pleased to have some MI308 sales in the fourth quarter. They were actually a license that was approved through work with the administration. And those orders were actually from very early in 2025. And so we saw some revenue in Q4, and we're forecasting for about $100 million of revenue in Q1. We are not forecasting any additional revenue from China just because it's a very dynamic situation.

So given that it's a dynamic situation, we're still waiting for -- we've submitted licenses for the MI325, and we're continuing to work with customers and understanding sort of their customer demand. We thought it prudent not to forecast any additional revenue other than the $100 million that we called out in the Q1 guide.

Now as it relates to overall Data Center, as I mentioned in the question to Tim, like we're very bullish about data center. I think the combination of drivers that we have across our CPU franchise, I mean, the EPYC product line, both Turin and Genoa continue to ramp well. And in the second half of the year, we will be launching Venice, which we believe actually extends our leadership and the MI450 ramp, which is also very significant in the second half of 2026. We're not obviously guiding specifically by segment, but the long-term target of, let's call it, greater than 60% is certainly possible in 2026.

[Operator Instructions] The next question comes from the line of C.J. Muse with Cantor.

I'm curious on the server CPU side of the house and given the dramatic tightness, curious your ability to source incremental capacity from TSMC and elsewhere. And I guess, how long will it take for that to see wafers out? And how should we think about the implications for kind of the growth trajectory throughout all of calendar '26? And I guess as part of that, if you could speak to how we should be thinking about inflection in pricing as well, that would be very helpful.

Sure, C.J. So a couple of points about the server CPU market. First of all, we think the overall server CPU TAM is going to grow, let's call it, strong double digits in 2026, just given the -- as we said, the relationship between CPU demand and overall AI ramp. So I think that's a positive.

Relative to our ability to support that, we've been seeing the trend for the last couple of quarters. So we have increased our supply capacity capability for server CPUs. And that's one of the reasons we're able to increase our Q1 guide as it relates to the server business, and we see the ability to continue to grow that throughout the year. There's no question that demand continues to be strong. And so we're working with our supply chain partners to increase supply as well. But from what we see today, I think the overall server situation is strong, and we are increasing supply to address that.

C.J., do you have a follow-up question?

I do maybe for Jean, if you could kind of touch on gross margins through the year and as you balance kind of strengthening server CPU with perhaps greater GPU accelerating in the second half. Is there kind of a framework that we should be working off of?

Yes. Thank you for the question. We are very pleased with our gross margin Q4 performance and the Q1 guide at 55%, which actually is 130 basis points up year-over-year, while we continue to ramp our MI355 year-over-year very significantly. I think we are benefiting from a very favorable product mix across all our business. If you think about in Data Center, we're ramping our new product, new generation product, Turin and MI355, which helps the gross margin in client. We continue to move up the stack and also gaining momentum in our commercial business.

Our client business gross margin has been improving nicely. In addition, certainly, we see the recovery of our Embedded business, which is also margin accretive. So all those tailwinds we are seeing, we continue to see in the next few quarters. And when MI450 ramp, of course, in Q4, our gross margin will be driven largely by mix. And I think we'll give you more color when we get there. But overall, we feel really good about our gross margin progression this year.

The next question comes from the line of Joe Moore with Morgan Stanley.

On the MI455 ramp, will 100% of the business be racks? Will there be kind of an 8-way server business around that architecture? And then is the revenue recognition when you ship to the rack vendor? Or is there something to understand about that?

Yes, Joe. So we do have multiple variants of the MI450 series, including an 8-way GPU form factor. But for 2026, I would say the vast majority of it is going to be Rack Scale solutions. And yes, we will take revenue when we ship to the rack builder.

Okay. Great. And then can you talk to any risks that you may have in terms of once you get silicon out, turning that into racks, any potential issues as you ramp that? I know your competitor had some last year, and you said you learned from that. Is there anything you've done with kind of prebuilding racks to sort of ensure you won't have those issues? Just any risk that we need to understand around that?

Yes. I mean, I think, Joe, the main thing is the development is going really well. It is -- we're right on track with the MI450 series as well as the Helios rack development. We've done a lot of testing already both at the Rack Scale level as well as at the silicon level. So far, so good. We are getting, let's call it, a lot of input from our customers on things to test so that we can do a lot of testing in parallel. And our expectation is that we will be on track for our second half launch.

Our next question comes from the line of Stacy Rasgon with Research.

First, Lisa, I just wanted to ask about OpEx. Like every quarter, you guys are guiding it up and then it's coming in even higher and then you're guiding it up again. And I understand, given the growth trajectory that you need to invest. But how should we think about the ramp of that OpEx and that spending number, especially as the GPU revenue starts to inflect? Do we get leverage on that? Or should we be expecting the OpEx to be growing even more materially as the AI revenue starts to ramp?

Yes. Sure, Stacy. Thanks for the question. Look, I think in terms of OpEx, we're at a point where we have very high conviction in the road map that we have. And so in 2025, as the revenue increased. We did lean in on OpEx, and I think it was for all the right reasons. As we get into 2026 and as we see some of the significant growth that we're expecting, we should absolutely see leverage. And the way to think about it is we've always said in our long-term model that OpEx should grow slower than revenue, and we would expect that in 2026 as well, especially as we get into the second half of the year and we see inflection in the revenue. But at this point, I think the -- if you look at our free cash flow generation and the overall revenue growth, I think the investment in OpEx is absolutely the right thing to do.

For my follow-up, I actually have 2 sort of one line answers I'm looking for. Just first, the $100 million in China revenue in Q1, does that also drop through a 0 cost basis like we had in Q4? And is that a margin headwind? And number two, I know you don't give us the AI number, but could you just give us the annual like 2025 Instinct number now that we're through the year? Like how big was it?

So Stacy, let me answer your first question on the $100 million revenue in Q1. Actually, the inventory reserve reversed in Q4, which was $360 million, not only associated with the Q4 revenue, China revenue, but also covers the $100 million revenue we expect to ship in Q1 to China with our MI308. So the Q1 gross margin guide is a very clean guide.

And Stacy, for your second question, as you know, we don't guide at the business level, but to help you with your models, I think you can -- if you look at the Q4 data center AI number, even if you were to back out the China number, which was, let's call it, not a recurring number, you would still see growth -- you'll see growth from Q3 to Q4. So that should help you a little bit with your modeling.

And the next question comes from the line of Joshua Buchalter with TD Cowen.

I want to ask about clients. So the segment beat pretty handily in the fourth quarter. and I recognize you guys have been gaining share with Ryzen. But I think given what we've been seeing in the memory market, there's a lot of concern about inflationary costs and the potential for pull-ins. Were there any changes in your order patterns during the quarter? And maybe bigger picture, how are you thinking about client growth and the health of that market into 2026?

Yes. Thanks for the question, Josh. The client market has performed extremely well for us throughout 2025, very strong growth for us, both in terms of ASP mixing up the stack as well as just unit growth. Going into 2026, we are certainly watching the development of the business. I think the PC market is an important market. Based on everything that we see today, we're probably seeing the PC TAM down a bit just given some of the inflationary pressures of the commodities pricing, including memory.

The way we are modeling the year is, let's call it, second half a bit subseasonal to first half, just given everything that we see. Even in that environment with the PC market down, we believe we can grow the -- our PC business. And -- our focus areas are enterprise. That's a place where we're making a very nice progress in 2025, and we expect that into 2026 and just continuing to grow sort of at the premium higher end of the market.

And then I want to ask about the Instinct family. So we've seen your big GPU competitor make a deal with an SRAM-based spatial architecture provider and then OpenAI has reportedly been linked to one as well. Could you speak to the competitive implications of that? You've done well in inferencing, I think, partly because of your leadership in HBM content. So I was wondering if you could maybe address the pull seemingly motivated by lower latency inference and how Instinct is positioned to service this if you're indeed seeing it as well.

Yes. I think, Josh, it's really, I think, the evolution that you might expect as the AI market matures. What we're seeing is as inference ramps, the -- really the tokens per dollar or the efficiency of the inference stack becomes more and more important. As you know, with our chiplet architecture, we have a lot of ability to optimize across inference training and even across sort of the different stages of inference as well.

So I think I view this as very much as you go into the future, you'll see more workload optimized products. And you can do that with GPUs as well as with other more ASIC-like architectures. I think we have the full compute stack to do all of those things. And from that standpoint, we're going to continue to lean into inference as we view that as a significant opportunity for us in addition to ramping our training capabilities.

And the next question comes from the line of Ben Reitzes with Melius Research.

Lisa, I wanted to ask you about OpenAI. I'm sure a lot of the volatility out there is not lost on you. Is everything on track for the second half for starting the 6 gigawatts and the 3.5-year time line as far as you know? And is there any other color that you'd just like to give on that relationship? And then I have a follow-up.

Yes. I mean I think, Ben, what I would say is we're very much working in partnership with OpenAI as well as our CSP partners to deliver on MI450 series and deliver on the ramp. The ramp is on schedule to start in the second half of the year. MI450 is doing great. Helios is doing well. We are in, let's call it, deep co-development across all of those parties. And as we look forward, I think we are optimistic about the MI450 ramp for OpenAI. But I also want to remind everyone that we have a broad set of customers that are very excited about MI450 series. And so in addition to the work that we're doing with OpenAI, there are a number of customers that we're working to ramp in that time frame as well.

All right. I appreciate that. And I wanted to shift to the server CPU and just talk about x86 versus ARM. There's some view out there that x86 has a particular edge in agents, big picture, do you agree with that? And what are you seeing from customers? And in particular, obviously, your big competitor is going to be selling an ARM CPU separately now in the second half. So if there's just anything on that competitive dynamic versus ARM and what NVIDIA is doing and your views on that, that would be great to hear.

Yes, Ben, what I would say about the CPU market is there is a great need for high-performance CPUs right now. And that goes towards Agentic workloads where when you have these AI processes or AI agents that are spinning off a lot of work, in an enterprise, they're actually going to a lot of traditional CPU tasks and the vast majority of them are on x86 today. I think the beauty of EPYC is that we've optimized. We've done workload optimization. So we have the best cloud processor out there. We have the best enterprise processor. We also have some lower-cost variants for storage and other elements. And I think all of that comes into play as we think about the entirety of the AI infrastructure that needs to be put in place.

I think the CPUs are going to continue to be as important as a piece of the AI infrastructure ramp. And that's one of the things that we mentioned at our Analyst Day back in November, is really this multiyear CPU cycle, and we continue to see that. I think we've optimized EPYC to satisfy all of those workloads, and we're going to continue to work with our customers to expand our EPYC footprint.

And the next question comes from the line of Tom O'Malley with Barclays.

Lisa, I just wanted to ask, you mentioned on memory earlier as a sticking point in terms of inflationary cost. Different customers do this in different ways, different suppliers do this in different ways. But can you maybe talk about your procurement of memory, when that takes place, particularly on the HBM side? Is that something that gets done a year in advance, 6 months in advance? Different accelerator guys have talked about different time lines. I would be curious to kind of hear when you do the procurement.

Yes. I mean, given the lead times for things like HBM and wafers and these parts of the supply chain, I mean, we're working closely with our suppliers over a multiyear time frame in terms of what we see in demand, how we ramp, how we ensure that our development is very closely tied together. So I feel very good about our supply chain capabilities. We have been planning for this ramp. So independent of the current market conditions, we've been planning for a significant ramp in our -- both CPU as well as our GPU business over the past couple of years. And so from that standpoint, I think we're well positioned to grow substantially in 2026. And now we are also doing multiyear agreements that extend beyond that given the tightness of the supply chain.

And just as a follow-up, you've seen a variety of different things in the industry in terms of system accelerators, so cache offload, more discrete ASIC style compute, CPX. If you look at what your competitors are doing and you look at your first generation of system architecture coming out, maybe spend some time on -- do you see yourself following in the footsteps of some of these different type of architectural changes? Do you think that you'll go in a different direction? Anything just on the evolution of your system-based architecture and then the adjoining products and/or silicon within?

I think, Tom, what we have is the ability with a very flexible architecture, with our chiplet architecture, and then we also have a flexible platform architecture that allows us to really have different system solutions for the different requirements. I think we're very cognizant that there will be different solutions. So there's no -- I've often said there's no one size fits all, and I'll say that again, there's no one size fits all.

But that being the case, it's clear that the Rack Scale architecture is very, very good for the highest end applications when you're talking about inference -- distributed inference and training. But we also see an opportunity with enterprise AI to use some of these other form factors. And so we're investing across that spectrum.

And the next question comes from the line of Ross Seymore with Deutsche Bank.

A couple of I guess my first question is back on the gross margin side of things. As you go from the MI300 to the 400 to the 500 eventually, do you see any changes in the gross margin throughout that period? In the past, you've talked about optimizing dollars more so than percentages. But just on the percentage side, does it go up, down? Or is there volatility as you go from one to the next for any reason? Just wondered on the trajectory there.

Ross, thank you for the question. At a very high level, each generation, we actually provide much more capabilities, more memory, help our customers more. So in general, the gross margin should progress each generation when you offer more capabilities to your customers. But typically, when you first ramp at the beginning of ramp of generation, it tends to be lower when you get to the scale, get to the yield improvement, the testing improvement and also overall performance improvement, you will see gross margin improving within each generation. So it's kind of a dynamic gross margin. But in the longer term, you should expect each generation should have a higher gross margin.

And then one on a small segment of your business, but it seems quite volatile and you talked a little bit about further off than you usually do is the gaming side of things. What is the magnitude down you're talking about this year? Because in 2025, you thought it was going to be flat and it ended up growing 50%, which was a nice positive surprise. But now that you're talking about this year being down, but then the next-gen Xbox ramping in 2027, I just hope to get some color on what you see as kind of the annual trajectory there.

Yes. So Lisa can add more. So 2026, actually, it's the seventh year of current product cycle. Typically, when you're at this stage of the cycle, revenue tend to come down. We do expect the revenue on the semi customer revenue side to come down significantly double digit for 2026, as Lisa mentioned in her prepared remarks, the next generation?

Yes. I think we'll -- I mean we'll certainly talk about that going forward. But as we ramp the new generation, you would expect a reversal of that.

Operator, I think we have time for one more caller on the call, please.

And our final question comes from the line of Jim Schneider with Goldman Sachs.

Relative to the ramp of your recallable systems, would you expect any kind of bottleneck in terms of supply constraints in terms of the ramp as you ramp the second half of the year to potentially impact or limit the revenue growth? In other words, maybe talk about whether you expect supply to really kind of mute the growth in Q4 sequentially relative to -- sorry, Q3 relative to Q4?

Yes. Jim, we are planning this at the -- every component level. So I think relative to our data center AI ramp, I do not believe that we will be supply limited in terms of the ramp that we put in place. I think we have an aggressive ramp. I think it's a very doable ramp. And as we think about the size and scale of AMD, clearly, our priority is ensuring that the data center ramps go very well, and that's both on the data center AI, the GPU side as well as on the CPU side.

And then maybe as a follow-up to the earlier question on OpEx. Could you maybe address what are some of the largest investment areas you made in 2025? And then what are the largest incremental OpEx investment areas for '26?

Yes, Jim, on the 2025 investment, the priority and the investment -- largest investment in Data Center AI. Our hardware road map, we accelerated that road map. We expand our software capabilities. We also acquired ZT Systems, which added significant system-level solutions and capabilities. Those are the primary investment in 2025.

We also invest heavily in go-to-market to really expand our go-to-market capabilities to support the revenue growth and also expand our commercial business and enterprise business for our CPO franchise. In 2026, you should expect us to continue to invest aggressively. But as Lisa mentioned earlier, we do expect revenue to expand faster than operating expense increase to drive the earnings per share expansion.

All right. Thank you, everybody, for participating on the call. Operator, I think we can go ahead and close the call now. Thank you. Good evening.

Thank you. And ladies and gentlemen, that does conclude the question-and-answer session, and that also concludes today's teleconference. You may disconnect your lines at this time, and have a great rest of the day.

So ladies and gentlemen, it is my privilege to welcome to the stage a globally respected technologist, an industry-defining CEO and a leader whose work continues to shape the very trajectory of modern computing. Ladies and gentlemen, please join me in welcoming to the stage, Chair and CEO of AMD, Dr. Lisa Su.

All right. What an audience. How are you guys doing tonight? That sounds wonderful.

First of all, thank you, Gary, and welcome to everyone here in Las Vegas and joining us online. It's great to be here with all of you to kick off CES 2026. And I have to say, every year, I love coming to CES to see all the latest and greatest tech and catch up with so many friends and partners. But this year, I'm especially honored to be here with all of you to open CES.

Now we have a completely packed show for you tonight, and it will come as no surprise that tonight is all about AI. Although the rate and pace of AI innovation has been incredible over the last few years, my theme for tonight is "you ain't see nothing yet". We are just starting to realize the power of AI. And tonight, I'm going to show you a number of examples of where we're headed, and I'll be joined by some of the leading experts in the world from industry giants to breakthrough start-ups, and together, we are working to bring AI everywhere and for everyone. So let's get started.

At AMD, our mission is to push the boundaries of high performance and AI computing to help solve the world's most important challenges. Today, I'm incredibly proud to say that AMD technology touches the lives of billions of people every day. From the largest cloud data centers to the world's fastest supercomputers to 5G networks, transportation and gaming, every one of these areas is being transformed by AI. AI is the most important technology of the last 50 years, and I can say it's absolutely our #1 priority at AMD. It's already touching every major industry, whether you're going to talk about healthcare or science or manufacturing or commerce, and we're just scratching the surface. AI is going to be everywhere over the next few years. And most importantly, AI is for everyone. It makes us smarter. It makes us more capable. It enables each one of us to be a more productive version of ourselves. And at AMD, we're building the compute foundation to make that future real for every company and for every person.

Now since the launch of ChatGPT a few years ago, I'm sure we all remember the first time we tried it, we've gone from 1 million people using AI to now more than 1 billion active users. This is just an incredible ramp. It looked -- it took the Internet decades to reach that same milestone. Now what we are projecting is even more amazing. We see the adoption of AI growing to over 5 billion active users as AI truly becomes indispensable to every part of our lives, just like the cell phone and the Internet of today. Now the foundation of AI is compute. With all of that user growth, we have seen a huge surge in demand in the global compute infrastructure, growing from about 1 zetaflop in 2022 to more than 100 zetaflop in 2025. Now that sounds big. That's actually 100 times in just a few years. But what you're going to hear tonight from everyone is we won't have -- we don't have nearly enough compute for everything that we can possibly do. We have incredible innovation happening. Models are becoming much more capable. They're thinking and reasoning, they're making better decisions, and that goes even further when we extend that to agents overall.

So to enable AI everywhere, we need to increase the world's compute capacity another 100x over the next few years to more than 10 yottaflop over the next 5 years. Now let me take a survey. How many of you know what a Yottaflop is? Raise your hand, please. A Yottaflop is a followed by 24 zeros. So 10 Yottaflop is 10,000x more compute than we had in 2022. There's just never, ever been anything like this in the history of computing. And that's really because there's never been a technology like AI.

Now to enable this, you need AI in every compute platform. So what we're going to talk about tonight is the whole gamut. We're going to talk about the cloud where it runs continuously delivering intelligence globally. We're going to talk about PCs where it helps us work smarter and personalize every experience that we have. And we're going to talk about the edge where it powers machines that can make real-time decisions in the real world. AMD is the only company that has the full range of compute engines to make this vision a reality. You really need to have the right compute for each workload. And that means GPUs, that means CPUs, that means NPUs, that means custom accelerators. We have them all. And each of them can be tuned for the application to give you the best performance as well as the most cost-effective solution.

So tonight, we're going to go on a journey. So you're going to go with me through several chapters as we showcase the latest AI innovations across cloud, PCs, health care and much more. So let's go ahead and start with the first chapter, which is the Cloud. The cloud is really where the largest models are trained and where intelligence is delivered to billions of users in real time. For developers, the cloud gives them instant access to massive compute, the latest tools and the ability to deploy and scale as use cases take off. The cloud is also where most of us experience AI today. So whether you're using ChatGPT or Gemini or Grok or you're coding with Copilots, all of these powerful models are running in the cloud.

Now today, AMD is powering AI at every level of the cloud. Every major cloud provider runs on AMD EPYC CPUs. And 8 of the top 10 AI companies use Instinct accelerators to power their most advanced models. And the demand for more compute is just continuing to go up. Let me just show you a few graphs. Over the past decade, the compute needed to train the leading AI models has increased more than 4x every year, and that trend is just continuing. That's how we're getting today's models that are dramatically smarter and more useful. At the same time, as more people are using AI, we've seen an explosion over the last 2 years of inference, growing the number of tokens 100x, really hitting an inflection point. You can just see how much that inference is really taking off. And to keep up with this compute demand, you really need the entire ecosystem to come together. So what we like to say is the real challenge is how do we put AI infrastructure at [indiscernible] scale, and that requires more than just raw performance. It starts with leadership compute, CPUs, GPUs, networking coming together. It takes an open modular rack design that can evolve over product generations. It requires high-speed networking to connect thousands of accelerators into a single unified system. And it has to be really easy to deploy. So we want full turnkey solutions.

That's exactly why we built Helios, our next-generation rack scale platform for the yota scale AI era. Helios requires innovation at every single level, hardware, software and systems. It starts with our engineering teams who designed our next-generation Instinct MI455 Accelerators to deliver the largest generational performance increase we've ever achieved. MI455 GPUs are built using leading-edge 2-nanometer and 3-nanometer process technologies and advanced 3D chiplet packaging with ultrafast high-bandwidth HBM4 memory. This is integrated into a compute tray with our EPYC CPUs and Pensando networking chips to create a tightly integrated platform. Each tray is then connected with high-speed ultra accelerator link protocol tunneled over ethernet, which enables the 72 GPUs in the rack to function as a single compute unit.

And then from there, we can connect thousands of Helios racks to build powerful AI clusters using industry standard Ultra Ethernet NICs and Pensando-programmable GPUs that can accelerate AI performance even more by offloading some of the tasks from the GPUs. Now we are at CES. It is a little bit about show and tell. So I am proud to show you Helios right here in Vegas, the world's best AI rack.

[Video Presentation]

Now for those of you who have not seen a rack before, let me tell you Helios is a monster of a rack. This is no regular rack, okay? This is a double-wide design based on the OCP Open Rack Wide standard developed in collaboration with Meta, and it weighs nearly 7,000 pounds. So Gary, it took us a bit to get it up here, just so you know. But we wanted to show you what is really powering all this AI. It is actually more than 2 compact cars. Now the way we've designed Helios was really working closely with our lead customers, and we chose this design so that we could optimize serviceability, manufacturability and reliability for next-generation AI data centers. Now let me show you a few other things.

At the center of Helios is the compute tray. So let's take a closer look at what one of those trays look like. Now I can tell you, I probably cannot lift this compute tray, so it had to come out. But let me just describe it a little bit. Each Helios compute tray includes 4 MI455 GPUs, and they're paired with the next-gen EPYC Venice CPU and Pensando networking chips. And all of this is liquid cooled so that we can maximize performance. At the heart of Helios is our next-generation Instinct GPUs. And you guys have seen me hold up a lot of chips in my career. But today, I can tell you, I am genuinely excited to hold up this chip. So let me show you mi455X for the very first time. 455 is the most advanced chip we've ever built. It's pretty darn big. It has 320 billion transistors, 70% more than mi355. It includes 12 2-nanometer and 3-nanometer compute and I/O chiplets and 432 gigabytes of ultrafast HBM4, all connected with our next-gen 3D chip stacking technology. So we put 4 of these into the compute trays up here.

And then driving those GPUs is our next-generation EPYC CPU code named Venice. Venice extends our leadership across every dimension that matters in the data center, more performance, better efficiency and lower total cost of ownership. Now let me show you Venice for the first time. I have to say this is another beautiful chip. I do love our chips, so I can say that for sure. Venice is built with 2-nanometer process technology and features up to 256 of our newest high-performance Zen 6 cores. And the key here is we actually designed Venice to be the best AI CPU. We doubled the memory and GPU bandwidth from our prior generation. So Venice can feed mi455 with data at full speed even at rack scale. So this is really about co-engineering. And we tie it all together with our 800-gig Ethernet Pensando Vulcano and Salina networking chips, delivering ultra-high bandwidth as well as ultra-low latency, so tens of thousands of Helios racks can scale across the data center. Now just to give you a little bit of the scale of what this means, that means that each Helios rack has more than 18,000 CDNA 5 GPU compute units and more than 4,600 Zen 6 CPU cores, delivering up to 2.9 exaflops of performance.

Each rack also includes 31 terabytes of HBM4 memory, an industry-leading 260 terabytes per second of scale-up bandwidth and 43 terabytes per second of aggregate scale-out bandwidth to move data in and out incredibly fast. Suffice it to say, those numbers are big. When we launched Helios later this year, and I'm happy to say Helios is exactly on track to launch later this year, we expect it will set the new benchmark for AI performance. And just to put this performance in context, just over 6 months ago, we launched mi355, and we delivered up to 3x more inference throughput versus the prior generation. And now with mi455, we're bending that curve further, delivering up to 10x more performance across a wide range of models and workloads. That is game changing.

mi5455 allows developers to build larger models, more capable agents and more powerful applications. And no one is pushing faster and further in each one of these areas than OpenAI. To talk about where AI is headed and the work that we're doing together, I'm extremely happy to welcome the President and Co-Founder of OpenAI, Greg Brockman to the stage.

Greg, it is so great to have you here. Thank you for being here. OpenAI truly started all of this with the release of ChatGPT a few years ago, and the progress you've made is just incredible. We're absolutely thrilled about our deep partnership. Can you just give us a picture of where are things today? What are you seeing? And how are we working together?

Well, first of all, it's great to be here. Thank you for having me. ChatGPT is very much the overnight success that was 7 years in the making, right? -- that we started OpenAI back in 2015 with a vision that deep learning could lead to artificial general intelligence to very powerful systems that could benefit everyone. And we wanted to help to actually realize that technology and bring it to the world and democratize it. And we spent a long time just making progress where year-over-year, the benchmarks look better and better. But the first time we had something that was so useful that many people around the world wanted to use it was ChatGPT. And we were just blown away by the creativity and the ways in which people found how to really leverage the models we have produced in their daily lives. And so just out of curiosity, how many people in the room are ChatGPT users?

That -- pretty much like the whole room.

I'm glad to hear it. But very importantly, how many of you have had an experience that was very key to your life or the life of a loved one, whether it's in health care, in helping manage a newborn and any other walk of your life?

And to me, that's the metric that we want to optimize and that seeing that number go up and to the right has been something that has been really different over 2025, right, that we really moved from being just a text box that you ask a question, you get an answer, something very simple contained to people really using it for very personal, very important things in their lives. And it's not just in personal lives, health care and aspects like that. It's also in the enterprise, right, and really starting to bring models like [indiscernible] to be able to transform software engineering. And I think that this year, we're really going to see enterprise agents really take off. We're seeing scientific discovery start to be really accelerated, whether it's developing novel math proofs. The first time we saw that was just a couple of months ago, and the progress is continuing. And it's really across every single endeavor of human knowledge work, where there's human intelligence that can be leveraged, right? You can amplify it. We now have an assistant, right? We now have the tool, we have an adviser that is able to amplify what people want to do.

I completely agree with you, Greg. I think we have seen just an enormous acceleration of what we're using this tech for. Now I would say, I think every single time I see you, you tell me you need more compute.

It's true. It's true.

It's almost like a broken record, you could just just Greg wants more compute. Can you talk about just some of the things that you're seeing in the infrastructure, some of the bottlenecks? And where do you think we should be focusing as an industry?

Well, the why, why do we need more compute is the most important question, right, which is really when the models are not that capable, right, and where we were in 2015, 2016, 2017 and so forth is that you basically just want to train a model and evaluate it, right? Maybe there would be a very narrow task to be useful for. But as we've made this exponential progress on the models, then there's actually exponential utility to them. People want to bring it into their lives in a very scalable way. And I think that what we're seeing is as we move from -- you ask a question, you get an answer to agentic workflows where you ask the model to write some software for you, and it goes off for minutes or hours or so even days. And you're not just operating one agent, you're operating a fleet of agents. right?

You can have 10 different work streams all going at once on your behalf for a single developer, right? And it should be the case that you wake up in the morning, and this is the kind of thing we are going to build, by the way, and ChatGPT has taken items off your to-do list at home and at work. And that all of that, that's going to require that big graph that you had of how much compute the world is going to need. That's going to require far more compute than we have right now. Like I would love to have a GPU running in the background for every single person in the world because I think it can deliver value for them, but that's billions of GPUs. No one has a plan to build that kind of scale. And so what we're really seeing is benefits in people's lives, right?

We're seeing, for example, some of my favorite applications and some of the ones I think are the most important are in healthcare, right, that we actually see people's lives being saved through ChatGPT. Just over the holidays, one of my coworkers says that her husband had leg pain. They went to the hospital, they went to the ER and they got it X-ray and the doctors are like, "Oh, it's a cold muscle, just wait it out, you'll be fine. They went home, it got a little bit worse, type symptoms into ChatGPT. ChatGPT said, go back to the ER, this could be a blood clot. And in fact, it was. It's deep [indiscernible] thrombrosis in the leg and in addition to 2 blood clots on the lungs. And if they just waited it out, that would have been likely fatal. And it's not a unique story.

CG, our CEO of Applications, I work very closely with every day, ChatGPT literally saved her life, too, right? She was in the hospital for a kidney stone and had an infection. They're about to inject an antibiotic. And she said, wait just a moment, she asked ChatGPT whether that one is safe for her. ChatGPT, which has all of her medical history, said, no, no, because you had this other infection 2 years ago, that could re-trigger it, and that could actually be life-threatening as well. And so she showed it to the doctor. The doctor is like, wait, what you had this condition. I didn't know. I only had 5 minutes to review your medical history.

I completely agree, Greg. I mean that's one of the things that we can always -- all of us can use a helper, and that's really what we have here. Look, I mean, I think you've painted a vivid picture of why we need more compute, of what we can do with AI. I think we feel exactly the same way.

Now we've also done an incredible amount of work with your engineering teams. mi455 and Helios is actually -- a lot of it is through some of the feedback from our engineering teams working closely together. Can you talk a little bit about that infrastructure? And what are your customers wanting? And how are you going to use mi455?

Well, one of the key things with how AI is evolving is thinking about the balance of different resources on the GPU. And so we have a slide to show how we've seen the evolution of this -- the balance of resources across different mi generations. So you see the slide that I very painstakingly put together. Actually, I did not painstakingly put it together. I asked ChatGPT to put it -- to go create the slide. And so it literally did all the research, and you can see some sources at the bottom and it actually went and read a bunch of different AMD materials, created these charts, put together, the title put together all of these headers and produce not just an answer for me to then go do a bunch of work and produce an artifact, an artifact that I can show. And this is just one simple example that you can do today with ChatGPT, right, that we are moving to a world where you are going to be able to have an agent that does all this work for you.

And for that, we're going to need to have hardware that is really tuned to our applications. What we have in mind is that we're moving to a world where human attention, human intent becomes the most precious resource. And so that there should be very low latency interaction anytime that human involved, or there should be an ocean of agentic compute that's constantly running. It's very high throughput. And these two different regimes of low-latency and high throughput yield a bunch of different pressures on hardware manufacturers such as yourself. So it's a pleasure to be working together.

We like building GPUs for you. That works well. Look, lastly, Greg, let's talk a little bit about the future. Paint the picture, one of the things that we've talked about is there are some people out there who are wondering, is the demand really there? Can AI compute like do we really need all of this AI compute? And I know you and I have talked about it. I think people don't have a view of the future that you see. I mean you have like a special seat. So paint the world for what this looks like in a few years?

Well, looking backwards, we have been tripling our compute every single year for the past couple of years, and we've also tripled our revenue. And the thing that we find within OpenAI is every time we want to release a new feature, we want to produce a new model, we want to bring this technology to the world. We have a big fight internally over compute, because there are so many things we want to launch and produce for all of you that we simply cannot because we are compute constrained.

And I think we're moving to a world where GDP growth will itself be driven by the amount of compute that is available in a particular country, in a particular region. And I think that we're starting to see the first [indiscernible] of this. And I think over the next couple of years, we'll see it start to hit in a real way. And I think that AI is something where I think data centers can actually be very beneficial to local communities. I think that's a really important thing for us to really prove to people. But also the AI technology we produce, that is also something where in terms of scientific advances, you think about what has been the most fundamental driver of increase in quality of life, right? It really is about science.

And every time we've gone into specific domains, you just see how much limitation there is from how things are done because that's just -- there's a particular discipline that's built a bunch of expertise. There are a small number of experts and then it's hard for them to propagate that to future generations. For example, in biology, we hooked up GP5 to a [indiscernible] setup and had humans described what the [indiscernible] looked like. The model said here, a couple of ideas to try, the humans will go try it, and it actually produced a 79x, almost 100x fold improvement in the efficiency of a particular protocol. And that's just one particular reaction that people have spent some time actually optimizing, but not like a ton and ton of time because it's just there's so much surface area available in biology that no human can possibly get to all of it. No human can be an expert across every single subfield.

And I think what we're going to see is AIs that really bridge across disciplines that humanity has been unable to bridge, right? You see this within health care, where as humans learn more, we specialize more, but we're going to have AI that's going to amplify. And so I think it will be for hard problems that AI will be brought to bear. This will be true for enterprise. Every single application, I think we'll have an agent that is accelerating what people want to do. And I think the hardest problem for humanity will be deciding how do we use the limited resources we have to get the most benefits for everyone.

That is an incredible vision. Greg, we are so excited to be working with you. I think there's no question in the world that we have the power to really change people's lives. Thank you for the partnership and really look forward to it.

Thank you.

So as you heard from Greg, compute is key, and mi455 is a game changer. But with the mi400 series, we designed a full portfolio of solutions for cloud, enterprise, supercomputing and sovereign AI. At the top is Helios that's built for the leading-edge performance, hyperscale training and distributed inference at rack scale. For enterprise AI deployments, we have instinct mi440X GPUs that deliver leadership training and inference performance in a compact 8-GPU server designed for easy use in today's existing data center infrastructure. And for sovereign AI and supercomputing, where extreme accuracy matters the most, we have the mi430X platform that delivers leadership hybrid computing capabilities for both high-precision scientific and AI data types. This is something unique that we at AMD do because of our chiplet technology, we can actually have the right compute for the right application.

Now hardware is only part of the story. We believe an open ecosystem is essential to the future of AI. Time and time again, we've seen that innovation actually gets faster when the industry comes together and aligns around an open infrastructure and shared technology standards. And AMD is the only company delivering openness across the full stack. That's hardware, software and the broader solutions ecosystem. Our software strategy starts with ROCm. ROCm is the industry's highest performance open software stack for AI. We have [indiscernible] support for the most widely used frameworks, tools and model hubs, and it's also natively supported by the top open source projects like PyTorch, LLM, SG Lang, Hugging Face and others that are downloaded more than 100 million times a month and run out of the box on Instinct, making it easier than ever for developers to build, deploy and scale on AMD.

One of the exciting AI companies using AMD and ROCm to power their models is Luma AI. Please join me in welcoming the Luma AI CEO and Co-Founder, Amit Jain, to the stage. Hello, Amit, how are you? It's great to have you here with us.

You're doing some incredible work in video generation and multimodal models. Can you tell us a little bit about Luma and what you're doing?

Absolutely. Lisa, thank you so much for having me here. Of course. Luma's mission is to build multimodal and general intelligence. So AI can understand our world and help us simulate and improve it. Most AI video and image models today, they're in early, early stages, and they're used to generate pixels. They're used to produce pretty pictures. What is needed in the world are more intelligent models that combine audio, video, language, image all together.

So at Luma, we are training these systems that simulate physics, causality, are able to go out, do research called tools and then finally, render out the results in audio, video, image, text, whatever is appropriate for the information that you're trying to work with. In short, we are modeling and generating worlds. So as an example, let me show you some of our -- some results from our latest model, Ray3. By the way, Ray3 is the world's first reasoning video model. So what it means is it actually is able to think first in pixels and [indiscernible] and decide whether what it's about to generate is good. And it's also the world's first model that can generate in 4K and HDR. So please take a look.

[Video Presentation]

Ican say Amit, that looks pretty incredible. So tell us, how are customers using Ray3 today?

So we are working with very large enterprises as well as individual creators across the spectrum, and we work with them in advertising, media, entertainment and industries where like you want to tell your story. 2025 was the year when they started to deploy our models and experiment with them. And towards the end of it, we are seeing large-scale deployments where people are using our models for as much as actually making a 90-minute feature-length movie.

What customers are also asking us a lot for as they're using it more and more is control and precision. How can they get their particular vision out onto the screen. And what we have realized and through our research is that control comes from intelligence, not just better prompts. You can't keep like typing in again and again and actually do those things. So we have built a whole new model on top of Ray3 called Ray3 Modify that allows you to edit the world. So let me show you actually what that looks like. This won't have audio, and I'm going to tell you a little bit about what you're seeing.

So what's playing on the screen is a demo of Ray3's world editing capabilities. It can take any real or AI footage from cameras or footage that you generated and change it as little or as much as you want to realize the creative goals. It's a powerful system that we have developed for our most ambitious customers who are most demanding, and they spread the [indiscernible] across entertainment, advertising, and this has allowed us to enable a new era of hybrid human AI productions. The human becomes the prompt through motion, timing and direction like, you act it out and then the model can produce it. What that means in practice is that filmmakers and creators can create entire cinematic universes now without elaborate steps and then edit and modify anything to get to the result they want. This has never been possible before.

But in 2026, we are focused on actually going much further. 2026 will be the year of agents where AI will be able to help you to accomplish more of the task or hopefully, the full end-to-end of the task rather than doing some patchwork. So our teams have been working diligently building the world's most powerful multimodal agent models. Using Luma, will suddenly feel like you have a large team of capable creators who are working with you in your creative pursuit. I want to show you a brief demo of what that would feel like.

So what you're seeing here is a new multimodal agent that can take a whole script from -- of ideas with like characters and everything and start imagining that in front of you. Now this is not script to movie. This is human AI interaction. And our next generation of models provide the ability to analyze multiple frames, long-form video and make selection and maintaining the fidelity of the character scenes and story and only editing when it's needed. Here, you're seeing human and AI collaborate in designing characters, environments, shots and the whole world.

And with our agents, we believe that creators will be able to make entire stories, what used to take a large production before. Again, this has never been possible before, and we have been using it heavily internally, and we couldn't be more excited. Individual creators or small teams that suddenly have the power of doing what entire Hollywood studios do.

That's pretty amazing. It's really nice to see how these Luma agents come together and make this happen. Now I know you have a lot of choices in compute. And when we first started talking, actually, you called me and said, you needed compute, and I said, I thought I could help. Can you tell us a bit about why you chose AMD? And what has your experience been?

Yes. We bet on AMD very early on. That call was in 2024, early 2024. And since then, our partnership has grown into a large-scale collaboration between our teams, so much so, that today, 60% of Luma's rapidly growing inference workloads actually run on AMD cards. Today, we also -- so initially, when we started out, we used to do a bunch of engineering. But today, we are at a point where most -- any operators, most any workloads that we can imagine run out of the box on AMD. And this is huge props to your software teams and the diligent work that is going into the ROCm ecosystem.

We are building multi-modal models. And actually, these workloads are very complex compared to text models. One example of that is these consumes hundreds of times, thousands of times more tokens. A video that you saw, a 10-second video, it's about 100,000 thousand tokens easily. Compare that to a response from an LLM, it's about 200 to 300 tokens. So when we are working with this much information, TCO and inference economy is absolutely critical to our business. Otherwise, there's no way to serve all the demand that is coming our way.

Through our collaboration with the AMD team, we have been able to achieve some of the best TCO -- total cost of ownership that we have ever seen in our stack. And we believe as we build these more complex models that are able to do auto regressive diffusion and that are able to do text and image and audio and video all at the same time, this collaboration will allow us to significantly differentiate on cost and efficiency, which, as you know, in AI is a big deal.

So through this collaboration, we have developed such a degree of confidence that in 2026, we are expanding our partnership to a tune of about 10x what we have done before. And these 455 cards, and I cannot be more excited for mi455X because this -- the rack scale solution and the memory and the infrastructure that you're building is essential for us to be able to build these [indiscernible] simulation models.

Well, we love hearing that, Amit. And look, our goal is to deliver more powerful hardware. Your goal is to make it do amazing things. So just give us a little brief view of what do you see customers doing over the next few years that just isn't even possible today?

Right. So as Greg was mentioning, early on in LLM land, right, like in 2022, 2023, they were great for writing copies, small e-mails, things like that. We could have never imagined that we would actually put these models into real-time systems, into health care and these kind of things through accuracy and architecture and scaling, LLMs have now gotten to that point. Video models are currently in that early stage. So they're great for generating video and 3D pictures. But soon, by scaling these models up, by improving the accuracy and data, we would end up with a place where these models will help us simulate real physical processes in the work, like CAD, architecture, fluid flows, help us design entire rocket engines, plan cities. And this is not [indiscernible]. This is what we do today manually with big giant teams in simulation environments.

These models will allow us to do that and automate that to a great degree. And as they become more and more accurate, multimodal models is what we need for the backbone of general purpose robotics. Your home robot will run hundreds of simulations in its head in image and video and then try, how do I do this? How do I solve this, so that it's able to do a lot more than current generation of LLM and VLM robots are able to do. This is how the human brain works. Humans are natively multimodal. Our AI systems will be as well.

That sounds wonderful, Amit. Look, thank you so much for being here today. Thank you for the partnership, and we really look forward to all that you're going to do next.

Thank you so much.

Thank you. So you've heard from Greg and Amit, what they said is they need more compute to build and run their next-gen models, and it is the same across every single customer that we have, which is why the demand for compute is growing faster than ever. Now meeting that demand means continuing to push the envelope on performance far beyond where we are today.

Mi400 series was a major inflection point in terms of delivering leadership training across all workloads, inference, scientific computing, but we are not stopping there. Development of our next-gen mi500 series is already well underway. With mi500, we take another major leap in performance. It's built on our next-gen CDNA 6 architecture manufactured on 2-nanometer process technology and uses higher speed HBM4E memory. And with the launch of mi500 in 2027, we're on track to deliver 1,000x increase in AI performance over the last 4 years, making more powerful AI accessible to all.

So with that, let's -- so with that, now let's shift from the cloud to the devices that make AI more personal, PCs.

[Video Presentation]

So for decades, the PC has been about a powerful device, helping us be more productive, whether at work or at school. But with AI, the PC has become not just a tool, but it's a powerful essential part of our lives as an active partner. It learns how you work and it adapts to your habits, and it can help you do things faster than you've ever expected, even when you're offline. AI PCs are starting to deliver real value across a wide range of everyday tasks from content creation and productivity to intelligent personal assistance. Let's just take a look at a few of the AI PC applications today.

Starting with content creation. These videos were created from simple text prompts on a Ryzen AI Max PC, so not in the cloud, but in a local environment. Anyone can generate professional quality photos and videos in minutes with no design expertise. Microsoft has been a key enabler of AI PCs, helping bring next-generation capabilities directly into our productivity tools. For example, managing your meetings, summarizing meetings, summarizing e-mails, quickly finding files that you need, using real-time translation on video conferences. And with Microsoft Copilot, advanced AI capabilities are being built directly into the Windows experience to complete tasks faster. You just describe what you need and the PC takes it from there.

Now at AMD, we saw the AI PC wave early, and we invested. That's why we've led every inflection point. We were the first to integrate a dedicated on-chip AI engine in 2023 and the first to deliver Copilot Plus x86 PCs in 2024. And with Ryzen AI MAX, we created the first single-chip x86 platform that could run a 200 billion parameter model locally. And now we're extending that leadership again with our next-gen Ryzen AI notebook and desktop processors. So today, I'm proud to announce the new Ryzen AI 400 series, the industry's broadest and most advanced family of AI PC processors.

Ryzen AI 400 combines up to 12 high-performance Zen 5 CPU cores, 16 RDNA 3.5 CPU cores, our latest XDNA 2 NTU delivering up to 60 [indiscernible] of AI compute and support for faster memory speeds. These flagship Ryzen AI mobile processors deliver significantly faster content creation and multitasking performance compared to the competition. Now there's a lot of excitement for the Ryzen AI 400 series. And if you're walking around CES this week, you're going to see many notebooks launching this week. The first Ryzen AI 400 series PCs begin shipping later this month with more than 120 Ultra-thin, gaming and commercial PCs launching throughout the year from every major OEM across every AI PC form factor.

Now powering the next generation of AI PC experiences takes more than just hardware. It takes smarter software with models that are lighter, faster and can run directly on device. These are different than what you're seeing in the cloud. So to talk more about this next wave of model innovation, please welcome Ramin Hasani, Co-Founder and CEO of Liquid AI.

Ramin, it's great to have you here. I'm very excited about the work that you guys are doing at Liquid. You are really taking a different approach to models. Can you talk a little bit to the audience about what Liquid is doing and why it's different from others?

Absolutely, Lisa. It is great to be here. We are a foundation model company spun out of MIT 2.5 years ago. We're building efficient generative AI models that can run fast on any processor inside and outside of data centers. We designed from scratch multimodal models with a hardware-in-the-loop approach that allows us to optimize neural architectures for a given hardware. We are not building transformer models. We are building liquid foundation models. Powerful, fast and processor-optimized generative models. The goal is to substantially reduce the computational cost of intelligence from first principles without sacrificing quality.

That means Liquid models deliver frontier model quality right on a device. Device could be a phone, it could be a laptop, could be a robot, could be a coffee machine and could be an airplane. Basically, anywhere compute exists with three value propositions: privacy, speed and continuity. It can work seamlessly across online and offline workloads.

Ramin, our teams have been working really closely on bringing more capable models to AI PCs. Can you share a bit about that work?

Absolutely. Today, I've got 2 new product announcements. One, we are excited to announce Liquid Foundation Models 2.5, the most advanced tiny class of models on the market. At only 1.2 billion parameters, the model performs best on instruction-following capabilities between its class and models that are larger in its class. LFM 2.5 instances are the building blocks of reliable AI agents on any device.

To put this in perspective for you, this model delivers instruction following capabilities better than the DeepSeek models and Gemini Pro kind of models, Gemini 2.5 Pro right on the device. We are releasing 5 model instances, a chat model, an instruct model, a Japanese enhanced language model, a vision language model and a lightweight audio model, audio language model basically. These are highly optimized for AMD, Ryzen AI, CPUs, GPUs and NPUs. And today, they are available for download on Hugging Face and on our own platform, Leap. Here you can enjoy them.

That's pretty cool.

So we can stack the [indiscernible] 2.5 instances together to build agentic workflows. But then it would be really amazing if we can bring in all these modalities into one place. So that brings me to my second announcement. LFM 3. LFM 3, it is designed natively multimodal to process text, vision and audio as input and deliver audio and text as an output in 10 different languages with sub-100 millisecond latency for audiovisual data. You will get LFM 3 later in the year.

All right. That's fantastic. So now I mean, help our audience understand like why should they be so excited about LFM 3? Like what can we do with these models on an AI PC?

Absolutely. So most assistants, AI assistants, Copilots today are reactive agents. You open an app, then you ask a question, it responds. But when the AI is running fast on the device and is always on, it can be working on the tasks proactively for you. The task can be done in the background. So let me show you a quick demo, a reference design to inspire what is possible to build on PCs with LFM instances. Let's jump in.

Imagine you're a sales leader working on your AMD Ryzen laptop with LFM 3 backbone proactive agents activated. You're in full focus mode, working on a spreadsheet, notifications start piling up, you get a calendar notification for a sales meeting, but want to continue your work in deep focus. A Liquid proactive agent notices the meeting and offers to join on your behalf. You allow the agent to join. And while you focus on your data analysis task in the background, the meeting is in progress with your agent representing you.

Are you sure we can trust this agent?

I think. I'm a little worried there. This system can actually transcribe more than transcribing your systems and really understanding what is going on. And the system can also be hooked up to your e-mail platform. It can analyze your e-mails as your e-mails are actually coming out, it can perform a deep research functionality.

So with the deep research functionality, you can analyze every e-mail and draft the response for you, again, everything under your own control. You're not going to go -- this is not going to go wrong. Everything is offline locally on the device. So this system can deliver a summary and can do the jobs better than what you have expected, what you have seen from reactive agents. I think this year is going to be the year of proactive agents. And I'm very excited to announce that we are bringing -- we are working or collaborating with Zoom to bring these features to the Zoom platform actually.

That's fantastic. Ramin, we're really excited about what you're doing. I think you've just given people just a glimpse of what we can do when we bring true AI capability to our PCs. So thank you. We're excited, and we look forward to all we're going to do together.

Thank you so much. Thank you for having me.

Thank you.

So now you see a little bit about what's possible with local AI. But the latest PCs aren't just running AI apps. They're actually building them. That's why we've created Ryzen AI MAX, the ultimate PC processor for creators, gamers and AI developers. It's the most powerful AI PC platform in the world with 16 high-performance Gen 5 CPU cores and 40 RDNA 3.5 GPU compute units and an XDNA 2 NPU delivering up to 50 [indiscernible] of AI performance, all connected by a unified memory architecture that supports up to 128-gigabytes of shared memory between the CPU and GPU.

In premium laptops, Ryzen AI Max is significantly faster in both AI and content creation applications compared to the latest MacBook Pro. In small form factor workstations, Ryzen AI Max delivers comparable performance to -- at much lower price than NVIDIA's DGX Spark, generating up to 1.7x more tokens per second per dollar when running the latest GPT-OSS models.

And because Ryzen AI Max supports both Windows and Linux natively, developers maintain full access to their preferred software environment, tools and workflows. Now there are more than 30 Ryzen AI Max systems in market today with new laptops, all-in-ones and compact workstations launching at CES and rolling out throughout the year. But our mission is to advance AI everywhere for everyone.

The truth is there are AI developers, many of you in this room, who want access to platforms that enable you to develop on the fly. So we took this one step further. Today, I'm excited to announce the AMD Ryzen AI Halo, a new reference platform for local AI deployment.

Now I would say this is pretty beautiful. Do you guys agree? So let me tell you what it is. This is the smallest AI development system in the world, capable of running models with up to 200 billion parameters locally, not connected to anything. It's powered by our highest-end Ryzen AI Max processor with 128 gigabytes of high-speed unified memory that is shared by CPU, GPU and NPU. This architecture accelerates system performance and makes it possible to efficiently run large AI models on a compact desktop PC that fits in your hand.

Halo supports multiple operating systems natively, ships with our latest ROCm software stack, comes preloaded with the leading open source developer tools and runs hundreds of models out of the box. And this really gives developers everything you need to build, test and deploy local agents and AI applications directly on the PC. Now for all of you who are wondering, Halo is launching in the second quarter of this year, and we can't wait for folks to get their hands on them.

So now let's turn to the world of gaming and content creation. A few gamers out there. I think there are a lot of gamers out there. Look, every day, gamers and creators rely on AMD across Ryzen and Radeon PCs, [indiscernible] workstations and consoles from Sony and Microsoft to deliver tens of billions of frames. And while the visual quality of those frames has advanced dramatically over the years, the way we build those worlds really hasn't. It still takes teams months or even years to bring a 3D experience to life. Now AI is really starting to change that. To show what's next in 3D world creation, I'm honored to introduce one of the most influential figures in AI. Known as the Godmother of AI, her work has transformed how machines see and understand the world. Please welcome the Co-Founder and CEO of World Labs, Dr. Fei-Fei Li. Fei-Fei, we are so excited to have you here. You've been one of the leaders shaping AI for decades. Can you just give us a little bit of your perspective? Where are we today? And why did you start World Labs?

Yes. First of all, thank you, Lisa, for inviting me to be here. Congratulations to all the announcements. I can't wait to you some of them. So it's true that there has truly been great breakthroughs in AI progress in the past few years. And as you said, I've been around the block for a while for more than 2 decades, and I really cannot be more excited than now by where things are going. So in the past few years, language-based intelligence in AI technology really has taken the world by storm. We're seeing the proliferation of all kinds of capabilities and applications.

But the truth is there's a lot more than just language intelligence. Even for us humans, there's more than passively looking at life in the world. We are incredible spatial intelligent animals. And we have profound capabilities that use our own spatial intelligence that connects perception with action. Think about all of you being here, how you break through airports this morning, I'm one of them, or woke up in your hotel room and get to the nice coffee shop or find your way in this maze in Vegas to be here. All this requires spatial intelligence.

So what excites me is that there's now a new wave of Gen AI technology for both embody AI and generative AI that we can finally give machines something closer to the human level spatial intelligence. It's the ability to not only perceive but create 3D or even 4D worlds, reason about objects and people and imagine entirely new environments that still abate the laws of physics and dynamics in worlds, virtual or real. So that's why I started World Labs. I really want to bring spatial intelligence to life and deliver value to people.

I remember the first time I talked to you about your concept for World Labs and your passion about what this could bring. Tell us a little bit about what your models do? So the audience gets a feel for what does this really mean?

Yes. Well, I heard that there are gamers out there. So this is very exciting. So traditionally, building 3D things requires laser scanners or calibrated cameras or hand-built models using pretty sophisticated and complicated software. But at World Labs, we're creating a new generation of models that can use the recent Gen AI technology to learn the structure, not only to flat pixel structure, I'm talking about 3D, 4D structure of the world directly from data, a lot of data. So give the model a few images and even one image, the model itself can fill in the missing details, predict what's behind objects and generate rich, consistent, permanent navigable 3D worlds.

So what you're seeing here on screen is a hobbit world that's created by our World Labs model called Marvel. We just give it a handful of images, and they created these 3D scenes that are persistent and you can navigate, you can even see a top view. And our system transformed a few visual inputs into a fully [indiscernible], expensive 3D world. And it shows how these models not just reconstruct the environment, they really imagine cohesive worlds, wonderous world. And once these worlds exist, they flow together and allowing effortless transition from one environment to the next and scaling into something much larger. And this is much closer to how humans piece together a place from a few glances.

Well, it looks pretty amazing that you can do that with such little input. Now can you just show us a little bit about how the technology works?

Yes, definitely. Let's just ground it in real world, a bit more from the hobbit world. And let's do something that you're very familiar with. So over the break, our team went to AMD's Silicon Valley office. I hope they got your permission.

They did not, but that's okay.

Okay. Well, now here we are. We just use some regular phone cameras. There's no special equipment, just phones to capture a few images. And then we put them into World Lab's generative -- 3D generative model called Marvel. And then our model that can use AMD's mi325X chip and the ROCm stack software stack can create a 3D version of that environment, including window, stores, furniture size and sense of depth and scale. And keep in mind, you're not looking at photos, you're not looking at videos. You're looking at truly 3D consistent world.

And then our team started to have a little more fun and decided to model free for you for different design styles, right? I don't know which one you guys like the most. I personally really like the Egyptian one, but -- maybe that's because I'm going there in a few months.

And while this transformation is keeping the geometric consistency and the 3D inputs. So you can imagine this can be such powerful tools for many use cases, whether you're doing robotic simulation or game development or design. This can -- what would traditionally take months to do in a typical workload, we really could do it in minutes now. And we can even navigate into an entirely different world, like actually the Venetian Hotel. We just did that yesterday by taking the picture and then put it in the model and then just had a little fun and they turned this whole place into a 3D imaginative space.

Now I'm sure you guys can take pictures and send it to Marvel and experience this yourself. But what you don't see here behind the scene is how much computation is happening and why inference speed really matters. The faster we can run these models, the more responsive the world becomes, instant camera moves, instant edits and a scene that stays coherent as you actually navigate and explore. And that's what's really important.

Fei, I think you're going to have a few people going out to your website to try Marvel. Keep the server up. But look, that looked really amazing. Can you just share a little bit about your experience working with AMD and our work on Instinct and ROCm?

Yes, of course. And even though we are old friends, our partnership is relatively new. And I got to be honest, we're -- I'm very impressed by how quickly this came together. Our -- part of our model is a real-time frame generative frame model. It was running on mi325X in under just a week. And then with AMD Instinct the ROCm, our teams were able to iterate really rapidly over a course of a few weeks to improve performance by more than fourfold. And that was really impressive.

That matters because spatial intelligence is fundamentally different from what came before. Teaching AI to understand and navigate 3D structure, have motion, understand physics requires enormous memory, massive parallelism and very fast inference. And I was seeing your announcement, I can't wait to see platforms like mi450 continue to scale, and they will give us the ability to train larger world models and just as importantly, to run them fast enough that these environments can feel alive, react instantly as the user or agents move, explore, interact and create.

No, that's wonderful. Thank you for those comments. Your team has been fantastic working together. So Fei-Fei, with all the compute performance that we're going to give you and all of the innovation in your models, give the audience a view of what to expect over the next few years?

Yes. I know. And as you know me, I don't like to hype. I think the word...

I call thunder hype. So...

No, I think we should just share what it is. It is going to be a changing world. A lot of workflows, a lot of things that were difficult to do will actually go through a revolution because of the incredible technology.

So for example, creators can now experience and create real-world things in real time, shaping what's in their mind's eye, experimenting with the space, the light, the movement as if they are sketching inside a living world. And the intelligent agents, whether it's robots or vehicles or even tools, can learn inside very rich physics-aware, digital world before they even need to be deployed into the real one, making them much safer, make the development much faster, more capable and more helpful to people.

And designers, for example, or architects can walk through ideas before anything is built, exploring form, flow, materials as navigate spaces rather than just looking at staring at static plans. So what excites me most is that this represents a shift in how AI shows up in our lives.

We're moving from systems that understand words and images passively to systems that not only understand but can help us to interact with the world.

So Lisa, what we are sharing today, which is turning a handful of images or photos into coherent explorable world in real time is not a glimpse of the distant future anymore. It is really the beginning of the next chapter. So you and I talk about this even offline. We know as powerful as AI technology is, it's also our responsibility to deploy and develop it in ways that reflect true human values, that augments -- that augment human creativity, productivity and our care for each other while keeping people firmly at the center of this story, however powerful technologies are. And I'm very excited to partner with AMD and with you on this journey.

Fei-Fei, I think I speak it for everyone. You are really an inspiration to the AI world. Congratulations on all the great progress, and thank you for joining us tonight.

Thank you, Lisa. Okay.

Next shift. Now let's turn to the world of health care.

[Video Presentation]

So look, as you saw in that video, AMD technology is already at work across health care. This is one of the most meaningful applications. You've already heard about some of the stories tonight of how high-performance computing and AI is one of the areas that I am most personally passionate about, is how you can bring health care there. There's nothing more important in our lives than our health and the health of our loved ones and using technology to improve healthcare outcomes means we measure progress in terms of lives saved.

I'm very happy to be joined tonight by three experts who are leading the way in applying AI to real-world health care challenges. Please join me in welcoming Sean McClain, CEO of Absci, Jacob Thaysen, CEO of Illumina, and Ola Engkvist, Head of Molecular AI at AstraZeneca.

All right, guys. Thank you so much for being here. You can see there's a lot of excitement about healthcare. Thank you for the tremendous partnership. Sean, at Absci, you're using generative models and synthetic biology to design new drugs from scratch. Can you walk through a little bit of how that works?

Yes. Thank you so much for having us here today. Biology is hard. It's complex. It's messy. Drug discovery and development is this archaic way of going about discovering drugs. Ultimately, it's this trial and error process where you ultimately are searching for a needle in the haystack. But with generative AI and what we're doing at Absci, you're actually able to start creating that needle and being able to actually engineer-in the biology that you want, being able to go after the diseases that have large unmet medical need and being able to have the manufacturability, the developability that you want in the drug.

We're actually able to start having precision engineering now because of AI with biology. Just like Apple is engineering an iPhone or you all are engineering the 455s, we're able to start engineering biology. And what is that actually doing? It's allowing us to start tackling some of the hardest, most challenging diseases that still exist that have high unmet medical need where standard of care is poor.

And at Absci, this is exactly what we want to tackle. We want to tackle these hard challenging diseases. Two of them that we're focused on at Absci is androgenic alopecia, so think common baldness. We actually have the opportunity in the not-too-distant future to have AI cure baldness. Wouldn't that be incredible?

And not only that, be able to focus in areas that have been neglected, women's health. For far too long, women's health has been pushed aside. And we have a drug that we are developing for endometriosis that affects 1 in 10 women with the opportunity to potentially deliver a disease-modifying therapy for these women. This is what AI and drug discovery is all about. And this wouldn't be possible without the compute partnership that we've had with AMD.

Lisa, you and Mark Papermaster invested in Absci roughly a year ago. And within that year, we've been able to scale the inference and be able to, in a single day, screen over 1 million drugs in one single day. That's incredible. And additionally, we're getting on to the 355s and the memory there is going to allow us to contextualize the biology in a way that we haven't been able to before and ultimately create better models for drug discovery. The future is really bright in AI and drug discovery.

That's fantastic, Sean. Well, look, thank you for the partnership. We're really excited about all the work we're doing together. Now Jacob, Illumina is really a leader in leading and understanding the human genome to improve health. How is AI helping in your work? And talk a little bit about what the impact is for the future of precision medicine?