Broadcom Stock

Good afternoon, everybody. Welcome to the Goldman Sachs Communacopia Technology Conference lunch session. My name is Jim Schneider. I'm the semiconductor analyst here at Goldman Sachs. It's my pleasure to welcome Broadcom and CEO, Hock Tan to the stage today. Welcome, Hock. It's really thrilled to have you here.

Thank you. Thank you for inviting me.

Hock, before we dive into talking about your AI business, I want to get some of your broader thoughts on AI and its adoption just across the world here. Given your deep view into the enterprise in all aspects, I think it's worth asking, how would you characterize the state of AI adoption across the average enterprise customer today?

That's a very interesting question, Jim. And I guess I will start off with Broadcom as a company, we -- like most large enterprises out there today, we have been looking closely reviewing how we can adopt generative AI through use of those frontier models and possibly even those open weight models that are floating around these days, but we've been looking at it. And we've been doing that in earnest now for 1 year. And part of the benefit we have is we have quite good full access to Mythos Preview that became available. And we look at how does it help our businesses.

And where we've been looking at very closely is obviously, we're a technology company in how we create products, how we design products, hardware as well as software. And we have engaged 15,000 of our 30,000 engineers in the company in looking at both in hardware and software. And a lot of insights, a lot of learnings. The biggest issue is finding use case that generates what you call a return on investment that is clearly meaningful and significant. That's not easy to do. And I suspect a lot of companies out there face the same issue.

The second thing is -- to frontier models, any model created as applications, it's just a tool. They're not meant to literally replace your people, not at least until you realize it's a tool and how would you apply it. And we find out that different users, depending who they are, could use the tool very differently, very effectively, very efficiently as well as all others who are having time really getting anywhere. It's like us use ChatGPT. And frankly, depending on how I ask, how I prompt ChatGPT, I get different sets of answers and some more useful than others. And over time, you get better and better at it, but that's a limit.

And so what we really find is to find great use cases, my thinking is simply one immediate use case, and that's probably from our perspective is creating new products, creating or help us create new technology. And what we're trying to ask for is not for it to create it out of thin air as much as create it through our engineers. And what we need -- what we find valuable out of it is it created with very high quality. So quality is very critical, and that's what those AI tools enable us to do and on a more accelerated fashion.

And for a semiconductor chip, that makes a difference because our engineers can still design the chip, but if they have a bug and have to respin that's 6 months and $30 million for leading-edge silicon. OpenAI, Anthropic tools, Gemini tools could help us eliminate the need for that respin. So there's a return on investment. So we find that as purposeful. But in terms of just simply productivity improvement, the return on investment is debatable. So it's really about quality, accelerated time to deployment or creation of that thing -- of leading-edge technology. You don't go back and redo what's in the past.

And the other application, of course, is cybersecurity through Mythos to scan to basically secure our infrastructure better. And finally, and it's something I haven't used yet -- we haven't used yet, but we can imagine you can use something like a Fable 5 in the minds of critical thinkers in specific functional areas to really replicate the software stack that we're using in specific areas. That definitely is big, and that can generate a big return. But that is something that will require a lot more investment, and we haven't reached that part of the thing yet.

So sum total of it is still early stage, but we do find it's important, most of all the users. And good users I define as critical thinkers, more in simple terms, the architects, senior people who are able to understand the context, understand the environment of where we are, the workflows and being able to prompt the models to do it. And I can imagine the day when they start using agents, which we are starting to last 3 months, that becomes even more important. So to sum it up, it helps smart -- very smart people become smarter.

That's wise. I won't say the converse. Now as the models get more efficient, more powerful, you talked about token cost coming down quite dramatically. What's your personal opinion about where the greatest value is going to accrue across the AI stack, model layer, application layer, someplace else?

Well, I'm a believer in the fact that intelligence and the more there is of it is unique. It's like doing semiconductors. When we design a switch, a network switch, that's the best of its generation at a timely basis, "the world beats a path to our door." There's no substitute for the best. Second best doesn't get anything. So I look at this the same way. For AI as we -- as long as there is continual improvement, I will almost say exponential improvement in the LLMs that keep coming out from the frontier model guys, the value is going to accrue over time to the guys who do the best frontier models.

And we look at it this way. Take the case of a very good frontier model. 1 gigawatt could generate $30 billion ARR, as you know. The cost to run 1 gigawatt, as you all know, is about $10 billion a year. So you have $20 billion that accrue to the model guys and the application that goes with it and $10 billion fought over by the power guys, the cloud provider guys, even the chiplet and memory guys. So the value comes, I think, in creating the best models and through it the best products to monetize the applications.

Understand. Okay. Now clearly, the pace and scale of this spending and infrastructure build-out is exceeding a lot of people's expectations. In your custom silicon business, you've consistently talked about working with a very limited number of large customers who have their own models can support very large spending programs over time. What do you think your customers are trying to achieve when they work with you? Is it supply diversification, cost optimization, some performance advantage by marrying software with custom silicon? And maybe talk a little bit about or quantify even the advantages they can get by working with you.

Actually, it's really very, very direct and straightforward. When we design, we develop, I would say, what I call a custom accelerator. We call it XPU, as you all have heard, for a customer. It's really very effective when we actually co-design it, work closely with that particular customer, who happens -- who has to be, in our view, to be a customer, a frontier model developer. Because what happens is we are closely embedded our engineering teams with the team of our customer who help create those frontier models. And it starts with the frontier models.

They understand what they're trying to create. They understand what breakthroughs they are making to make the -- for their algorithms to create workloads that run their models much more effectively, more productively, more efficiently. And they turn the specs over to us. And we work with them to create a chip all the way through to the physical part of the chip that enables the chip, the transistors to be able to handle the kind of workloads that model does. So that co-design, co-working together, collaboration, tight collaboration really makes the difference. And if you do that tight collaboration, you have seen a walking example of what happened recently and is news out there with OpenAI. We work closely with the OpenAI team and within a year, we created an XPU called the Jalapeno, which performs as well, if not better than the best state-of-the-art general purpose accelerator out there.

Fair enough. Now I want to get to the heart of the forecast you provided at your earnings call last week. You basically laid out the path to doubling your AI revenue for each of the next 2 years, $115 billion in '27, $230 billion in 2028. And you said those are supply constrained estimates. So I want to take a minute to sort of unpack some of that.

In terms of the supply constraint or the constraint part of it, you mentioned the ability to deploy physical data centers as one of those, land, power, shell. What is the time frame that you think is kind of the pinch point for data center availability on the infrastructure side? What's the biggest limiter inside that as you see it, grid power, gas turbines, construction? And then how would you sort of throw this sort of wildcard about political backlash into the equation?

That's what we are seeing now. The positive side to this phenomenon happening is you cannot build up easily an AI data center without a lot of lead time. So it gives us pretty good visibility out there, what needs to be done. But equally, that lead time leads to a fair degree of sometimes uncertainty on preciseness of the timing. And one of the biggest -- I mean, before I even say that, now for creating the chips and even that we do, whether it's accelerators, AI accelerators or network chips, connectivity chips, we understand what it takes to the supply chain for chips.

We understand the need for wafers -- leading-edge wafers and memory as well as even substrates. And that's pretty well locked in. We can pretty well lock that in for sure in '27 and the process of locking in '28 gives us -- within some limits. That is more within our control. What is interestingly is we find out that's not enough when you want to put together AI data centers. You need power, you need a site, you need building -- you need a facility to house your chips for it to run. And that's the other part of the constraint. I'm not saying chips and memory are not constraints. They are constrained.

But at least there are constraints we can work with and more deterministic. What is less deterministic is a power site that is power-ready, no that power available, but not ready. So you want to make it power ready, say, in '28, got to start construction on the sites now. And that's not us, our customers has to start construction on the site now to make it available and ready by '28. And that's 2 years out, almost 2 years, 1.5 years out. And it's -- a lot of it is not just equipment, whether they are transformer, gas turbines, that obviously are the pieces that go in it. It's also construction.

Construction, including local permitting, though some of it exists, but also the rate of construction, you're talking about those that you're trying to renovate or build your own houses, you know how construction is, right? That's what at large scale, this is happening. And so there's a lot of other factors that come in that kind of make it -- I describe it as perhaps constrained because if I turn around and give you guys a forecast, which we do, we want to be sure that's a forecast that is very clear line of sight. And that's what translates to the numbers you mentioned. It's a judge number based on not just supply of memory wafers, but availability and readiness of power sites.

Fair enough. Okay. Now in terms of customers, Google is one of your long-standing strategic partners for custom silicon. Thomas Kurian was just up here a couple of hours ago, outlining their strategy. Back in April, you disclosed a long-term agreement with Google to supply TPUs and networking products through 2031. Can you maybe talk about how this long-term agreement deepens the partnership between you and Google over that time frame?

Google and us have a very strong relationship. We've been doing this for over 10 years ever since the first version of the TPU. And we've been doing every version since then. And we will continue to do so under this LTA because if anything else, it expands and strengthen the relationship, very technical relationship, strategic relationship we have had with Google now for 10 years, and we see that continuing.

And one of the things that we offer, we believe we offer to Google is the strength, the breadth of our technology in -- especially in semiconductor design because with every new generation of GPU or XPU in our case, we are pushing. Our engineers are pushing the limits of semiconductor technology. And part of one example is, as you know, in the past, 10 years ago, you get better performance out of a CPU in chips by moving to a next-generation process node.

And each time you move to a new generation process node every 18 months or so, you double the transistor count, you double performance. Well, that's Moore's Law. Well, Moore's Law has come to an end. Going -- moving down process node, you do not -- you cannot achieve double transistor performance anymore. You barely eke out 5% improvement, but you get power. So that's one positive thing. But building the kind of accelerators we are building today, and these are monster multipliers, very, very complex chips.

What we are doing with our customers is we're trying to what I call substitute Moore's Law with something else. And the simple way to look at it is Moore's Law is one chip -- semiconductor chip is limited by the radical law of physics on optics, 800 square millimeters. That's the largest size die you can make. So we can only trim so much transistors into one. And the transistors are obviously mostly multipliers used in AI accelerator.

What we're doing now is we're putting 2 die together in 1 chip. So effectively, you now have 1,600 square millimeters, double as a way of getting the double the performance. Well, the current generation we are working on is 4 dies. Can you just imagine. And then one after that probably will go to maybe no limit 8 die, 8 die in 1 chip. So your chip size is equivalent of, say, 800 times 8 over 6,000 square millimeters, running as one single chip.

The amount of technology, the complexity of technology needed to be created to enable that to run and it's technology that's not just created at our wafer fab or with the memory guys, it's us designing it. It's very, very challenging, very interesting. And that's the relationship we find very exciting with every one of our customers because we are able to keep enabling them to create more and more high-performance chip, which enables them to drive as a key part of their model -- of their stack, much higher, better -- exponentially better and better LLMs, which drives towards what I said before, more and more value out of .

And my view of it is with the ability to create those kind of chips and the imagination and intelligence of the R&D engineers, the data scientists writing those algorithms and finding breakthroughs towards Superintelligence. We do not see a limit at all yet towards the exponential improvement of those frontier models.

Yes. And then as you know, there's been some investor concern about the potential for some of your customers, including Google, to in-source more elements of their custom silicon strategy. So maybe help us understand what are some of the deep technical and/or commercial moats that you have that give you confidence that, that custom silicon business will grow with that customer?

It's -- well, I have expressed some of that, and they continue to grow. And the example of it is literally, we have more IP out there in semiconductors than probably most of our peers out there across the Board, whether it's in IOs through very high-speed interconnect SerDes, whether it's in creating much more density of multipliers within every square millimeter and as well as putting multiple die in one chip, you have to have dies communicate very well, very fast with each other and being able to do it very low power on and on, and then advanced packaging, how do you package a die in 1 chip.

All these are, I consider deep moats that our peers have to come forward with to develop it to challenge us with the advantage of being earlier than most other guys. And most of all, it's no different than what we have been going through as a semiconductor company for the last 30 years, where we have picked different areas, whether it's switching, whether it's networking from a different point of view, Ethernet and computing. We have been able to essentially compete our engineer and compete against our competitors. It's no different here, whether it's another peer semiconductor company or more interestingly enough, even our customer themselves.

Okay. You mentioned OpenAI before, so maybe mention them again. You previously signed a supply agreement for 10 gigawatts of custom silicon capacity last October. And then at the end of the June, as you referenced, you announced, you delivered your very first ASIC sample called Jalapeno within 9 months, I believe, which I think is pretty fast. So maybe discuss the partnership with OpenAI and what enabled you to deliver on that accelerated time line.

Well, the biggest thing is like what we do with our customers who are frontier model developers, we -- they have a strong team in developing the models. They also have a team that translates what the model wants to specifications to architecture of what the hardware -- the compute hardware needs. And we provide that creation of that chip with architecture from the customer side as well as from our side to create the micro architectures and the physical design.

When you have that very tight, almost seamless flow, it's not that much a challenge to be able to create chips of that caliber that quickly. So the key is the tight engagement with the customer. So needless to say, relationship is very strong and continues to be very strong because we are working on the next 2 generations to the Jalapeno.

Okay. Then in June, you also announced a special purpose vehicle in collaboration with Apollo and Blackstone to provide initial $35 billion in financing for over 20 gigawatts of compute capacity to various AI labs, starting with a gigawatt for Anthropic this year and take 5 in 2027. Maybe help us understand the parameters behind that agreement? And how broad -- what's enabling Anthropic to sort of extend their compute capacity and maybe how broad could this agreement get?

It comes back to our vision of this -- of what we see in generative AI. One we look at is we have only 6 customers today at this point, very concentrated group, but I believe these 6 customers all each of them are creating their own frontier models, LLMs. And they each continue to invest to try to create a state-of-the-art leading edge. And that's what makes this all very interesting because that's the creation of this phenomenon of Superintelligence because from the models comes the products, applications, as you call them, from which you monetize and from which enterprise consumers benefit.

And what -- with this focus in mind, we also look at and look at it very practically. We want to support all these 6 guys because they do the best. And we can't tell who's going to win. And by the way, they may coexist in where they land in their models as it is a continual state of evolution, I would almost say revolution in a way they exponentially improve. Of the 6, 2 of them, 2 of the strongest guys do not have the cash flows to be able to generate the compute needed to train and create inference productize the models.

We create the chips for them. We step forward to enable them to have the compute needed to support their models. And we do that by not just taking cash from our books over there. We do it by creating a platform, we call it XPV, where we will enable them to be financed. We do not finance directly. So -- and I'll be direct, this is not circular financing. When a customer sees the demand that drives up their own demand and they don't have to use us necessarily, they can get it somebody else. That's not circular finance. We're using financing to create demand. Demand is there.

What we're doing here is we harness a lot of financial -- a bunch of financial partners and banks who are willing to come in and essentially take the risk on financing Anthropic and OpenAI, at least partially. And on the balance of it, which is what we did with Apollo and what we'll do on the XPV going forward, we will provide backstop in the form of residual guarantees on what is the equipment that is secured against that. So there's a combination of financial partners will come in. And why they come in, a big part of it is they believe in the demand, but two is rates, right? They get better rates creating through this. And our partners, OpenAI and Anthropic, have the capability to handle that as they grow, as they generate more and more revenues to be able to take on this additional enabling this financing capacity.

Okay. And then last question on this topic. We've seen a lot of rapid development in China, specifically around open source, open-weight models, a lot of activity from players in the region, whether it be Moonshot, DeepSeek or others. I mean how do you see the role of open source or open weight AI models in the market over time? And from where you stand, do you think any of those customers could actually become customers of Broadcom in future?

That's an interesting question. You're going to set me off, by the way, about this. As I said in my -- in all this question you asked, AGI believes in reaching -- generative AI in reaching Superintelligence. We're not there yet. We are not at AGI yet, but we can get there. And the frontier models, the guys who do frontier models are pushing us towards doing that. And when we -- and as I said, too, the better the models, the better applicate use case, the better results, outcomes you get out of it. But equally, to get to those models and get to those outcomes, those are just tools. You need to train up and you need to have users that know how to do it.

So obviously, I'm a believer in closed frontier models. That's not to say open-weight models, as you put it, can coexist. They could. But the value chain, I think over time, over the next 5 years, I believe, will reside with the leading-edge frontier models. As opposed to the open web models, even though they may coexist because I hear all the time as I go sell products to enterprises, how they don't like the fact token cost is getting expensive when they start using it for very complex application that maybe open source, will give them a cheaper source.

But my thinking and my experience as a technology company is, no, the best is what matters at the end and value accrues to the best. I'll give you an example of the math, the economics we see today, just today. Now I agree things might change a year, 2 years down the road, but things have changed a lot since then between open-weight models and closed models, roughly. And you can get these sources of data from OpenRouter-Vercel.

The total amount of compute tokens consumed and generated, the cost of generating those tokens to be consumed globally. And this data is obviously for inference. Trading is not easily monitored. So it's not that just inference. It's around $200 billion a year now, give or think, rough numbers. That's the cost of generating those tokens, producing those tokens in infrastructure, $200 billion. Ask yourselves, what's the revenue you could attribute being created and earned by those model guys. I give you that, you can count and look at where the models are either around $150 billion.

Interesting, isn't it? This generative AI today, I know it's early stage, things are still growing, but we're generating $200 billion of token costs to generate $150 billion. But look below the surface, you can split it up. On that tokens generated, half of it is generated through a frontier model, closed systems. The other half, roughly, in fact, more than half generated through players offering open source, open-weight models. So that's about it. But you look at the revenue, 75% at least is coming from frontier models. That's $120 billion, and they spend roughly $100 billion. It's not so bad, it's still growing.

Let's look at open-weight models. So spending $100 billion, the other $100 billion, they generate revenue, $30 billion. Do you think that's a sustainable model? We don't know, but it proves one thing. The value goes to where intelligence continue to improve.

Fair enough. Just 1 minute left, but I wanted to ask you about capital allocation for a moment. From your financial forecast, you're going to generate a tremendous amount of free cash flow over the next couple of years given the numbers you've laid out. Can you maybe update us on the Board's thinking around capital allocation in light of all that?

Well, it's -- give me 3 months. We usually do a capital allocation annually in the December Board meeting. But your question is very right on in the sense because I can imagine, given our forecast and we already start seeing that in 2026 because we'll end 2026 with a record amount of cash sitting on our books end of this quarter, end of October, just simply because our revenue is growing so fast, driven a lot by AI.

Go to next year, we're talking as our guidance forecast, $115 million (sic) [ billion ] for AI revenue alone. You add on it our non-AI revenue and software, you're talking about pretty high dollar number, which 45 -- mid-40s free cash flow. So we're going to generate a lot of cash, to your point exactly, and more so in '28. And the question we have to look at as management and the Board is, one, we could give it out as increased dividends, but how much of that will be increased dividends, still a long way off. And the other likelihood is we'll do a stock buyback. These are probably the other 2 choices we have to probably bring down debt makes sense. But our debt -- total debt of about $56 billion was taken up when interest rates were much lower. So it's almost like -- why would I want to take -- prepay low-cost interest rate debt, which really leaves 2 choice of capital allocation, increase dividends and probably do stock buyback.

Very good. I think that, we're out of time, but thank you very much, Hock, for being with us. We really appreciate it.

Thank you.

Welcome to Broadcom Inc.'s Third Quarter Fiscal Year 2026 Financial Results Conference Call. At this time, for opening remarks and introductions, I would like to turn the call over to Ji Yoo, Head of Investor Relations of Broadcom Inc. Please go ahead.

Thank you, Cherie, and good afternoon, everyone. Joining me on today's call are Hock Tan, President and CEO; Amie Thuener, Chief Financial Officer; and Charlie Kawwas, President, Semiconductor Solutions Group.

Broadcom distributed a press release and financial tables after the market close, describing our financial performance for the third quarter fiscal year 2026. If you did not receive a copy, you may obtain the information from the Investors section of Broadcom's website at broadcom.com. This conference call is being webcast live, and an audio replay of the call can be accessed for 1 year through the Investors section of Broadcom's website.

During the prepared comments, Hock and Amie will be providing details of our third quarter fiscal year 2026 results, guidance for our fourth quarter of fiscal year 2026 as well as commentary regarding the business environment. We'll take questions after the end of our prepared comments.

Please refer to our press release today and our recent filings with the SEC for information on risk factors that could cause our actual results to differ materially from the forward-looking statements made on this call. In addition to U.S. GAAP reporting, Broadcom reports certain financial measures on a non-GAAP basis. A reconciliation between GAAP and non-GAAP measures to the extent possible is included in the tables attached to today's press release. Comments made during today's call will primarily refer to our non-GAAP financial results.

I will now turn the call over to Hock.

Well, thank you, Ji, and thank you, everyone here for joining us today. We delivered an exceptional quarter with revenue, operating income and free cash flow all exceeding prior records. And driving this was our Q3 AI semiconductor revenue, which grew 221% year-on-year and up 54% sequentially. This brought our consolidated revenue to $29.6 billion which was up 86% year-on-year. Operating income grew even faster at 92% year-on-year with operating margin at a record 68% of revenue, reflecting strong operating leverage.

Q3 demand was simply hot and we're just getting started. Our 6 XPU customers are accelerating the adoption of custom accelerators. And AI semiconductor revenue more than tripled year-over-year to $16.7 billion. During the quarter, we delivered Ironwood TPU v7, version 7 in high volume to both Anthropic and Google. At the same time, we began production shipments of the next-generation TPU version AI for Google. This new TPU generation has been designed with more memory and bandwidth compared to Ironwood. And just like Ironwood is optimized for inference workloads. And in performance, it is comparable, if not surpasses the [indiscernible] GPU.

Just to reiterate the major technology challenges of developing these complex accelerators, Broadcom is now shipping the TPU version 8i, ahead of the [ MediaTek ] version [ v8t, ] which, in fact, was initiated earlier.

In Q3, we also shipped Jalapeno, OpenAI's first-generation custom accelerator, which outperforms [ Grace Blackwell ] GPU for inference workloads. Overall, our XPU shipments for the third quarter were up over 3.5x year-on-year and represented 73% of AI revenue during the quarter. Our AI networking revenue was up over 2.5x year-on-year.

And this strong momentum continues into Q4. During the quarter, we expect to accelerate shipments of Ironwood to Anthropic. We also expect to ramp up high-volume shipments of the TPU version 8i to Google. Shipments of Jalapeno for Anthropic will continue, and for Meta, we expect production shipments of their custom MTIA accelerator optimized for inference and recommendation at scale.

In Q4, we expect both XPUs and AI networking revenue to triple year-on-year. And together, we expect these deployments to drive our Q4 AI revenue to $21.7 billion, which is up 236% year-on-year. Based on all this -- on this Q4 guidance, I should say, we expect our fiscal 2026 AI revenue to be $58 billion for the year, up 186% year upon year and above our prior guidance of $56 million.

We are continuing to see exponential growth in demand from our XPU customers. We believe the vast majority of compute demand for AI workloads today originates from these concentrated group who develops state-of-the-art frontier models, and we expect the need for compute infrastructure to inflect even more in 2027 and 2028. And they are all growing with XPUs to achieve superior performance, cost and power.

Let me now walk you through each of their journey -- each of our customers' journey towards using XPUs at scale to run their frontier models worldwide [indiscernible]. Our engagement with Google has never been stronger. This has been recently strengthened by a long-term agreement to develop and supply future generations of TPUs and AI networking. Under this agreement, we're planning to deliver multi tens of billions of dollars of TPUs annually over the next several years. We expect this growing demand in '28 and '29 to be fulfilled through successive generations of the increasingly complex TPUs we are developing today with Google.

Our partnership with Google will continue to sustain because we have the strongest IP portfolio in semiconductor design, including industry-leading SerDes, chip-to-chip interconnect, leading-edge HBM and SRAM integration and simply differentiated advanced packaging. Most of all, we have been consistently -- we have consistently, I should say, delivered the fastest time to market for TPUs from product definition to production. We found the need for [indiscernible]. We believe these are very deep moats for any competitor to cross.

Moving on to Anthropic, starting with the 1 gigawatt of Ironwood we are deploying in 2026. We expect Anthropic to deploy another 5 gigawatts of TPU, V -- version 8i in 2027. And in 2028, we have clear line of sight to deliver another -- an incremental 10 gigawatts even as we expect Google to grow for us, Anthropic is on track to become our largest XPU customer in 2027 and sustain that in 2028.

For OpenAI, Jalapeno is on track for the planned deployment of 1.3 gigawatts in 2027. Together with OpenAI, we are deep in development of the next-generation XPU beyond Jalapeno, which is approaching tape-out. In 2028, we have line of sight for OpenAI to deploy over 5 gigawatts of Jalapeno and its successful generation of XPU, which would make OpenAI our second largest XPU customer. In addition, we are in development with OpenAI on the third generation XPU.

As OpenAI announced last week, Jalapeno outperformed the Grace Blackwell Ultra in performance, but [indiscernible] latency, throughput and power and it's actually comparable to [indiscernible] GPUs in running OpenAI workloads. The lesson here is when you codevelop a chip that is optimized for your particular LLM workloads, you will outperform any GPU. Like TPUs, Jalapeno demonstrates that it can also run other frontier models, and you can do all this at half the cost of a GPU.

Our partnership with Meta to deliver multiple generations of MTIA XPUs remains on track. Between now and the end of 2027, we will be delivering 3 generations of MTIA accelerators to Meta. Across these 3 generations, we have line of sight to deploy 3 gigawatts through '28.

Our content in AI, as you know, goes beyond XPUs. We're the leader in AI networking, and we continue to extend our lead. In ethernet switching, for scale up and scale out. We were first to market with our 100 terabit Tomahawk 6, and we just take out Tomahawk 7, the industry's first 200 terabit per second ethernet switch. For scaling in, we continue to be the leader in every generation of PCI Express switching. We're now the leader in leading-edge optical DSPs and are rapidly expanding our capacity and market position as a leader in EMLs, VCSELs and CW lasers for optical interconnects.

In sum, we continue to invest and invest heavily to provide the broadest and most leading-edge AI portfolio. In fact, our AI networking revenue is expected to grow just as fast as XPUs over the next few years.

Reflecting our excellent progress with this key group of LLM customers, here is our outlook for our AI semiconductor revenue: in 2027, we have secured the supply to again double AI revenue to approximately $115 billion. Our demand actually exceeds this outlook, and we will work to improve supply. In 2028, we expect the trajectory of growth to continue. We have line of sight for fiscal 2028 AI semiconductor revenue growth to again double to $230 billion. Here again, we have secured the supply to meet this outlook.

This AI revenue guidance through '28 is being provided to give you the trajectory of our growth. That demand for compute continues to be extremely strong. As a result, I got to say we are very much on target to exceed $30 in earnings per share in fiscal 2028.

Now turning to non-AI semiconductors. Q3 revenue of $4.2 billion was up 5% year-on-year and flat sequentially. Broadband and server storage together were up partially offset by a decline in wireless. In Q4, we forecast non-AI semiconductor revenue to be approximately $4.3 billion, again, up 5% sequentially.

Talking about infrastructure software, Q3 revenue of $8.8 billion was up 29% year-on-year. And we sustain ARR growth of 15% year-on-year. For Q4, we forecast infrastructure software revenue to stabilize at approximately $8.7 billion.

We announced VMware private AI cloud, giving enterprises a secure cost-effective platform to build and run AI alongside their existing applications. It brings together AI infrastructure, security and compliance and the tools to build and operate trusted AI agents. All while protecting enterprise data. VCF, VMware cloud that is, is also making it easier for customers to repatriate workloads from public cloud to private cloud where they can gain greater control and significantly improve infrastructure economics. Enterprise consumption of AI is in fact, opening a new opportunity for our infrastructure software business.

So to sum it all, for Q4 '26, we expect consolidated revenue to grow to $34.8 billion, up 93% year-on-year. We expect Q4 AI revenue to be $21.7 billion, up 236% year-on-year, and we expect operating margin to be approximately 66% of revenue.

And with that, let me turn it over to Amie.

Thank you, Hock. Let me now provide additional detail on our Q3 financial performance. Consolidated revenue was a record $29.6 billion for the quarter, up 86% and year-on-year. Gross margin was 75% of revenue in the quarter, down 210 basis points sequentially as AI semiconductor revenue was a greater proportion of our total revenue mix. This was better than our guidance of 74%.

Q3 operating income was a record $20.1 billion, up 92% from a year ago. Even with the decline in gross margin due to revenue mix, operating margin increased 240 basis points year-over-year to 67.9% because of phenomenal operating leverage we are achieving. Aligned with this Q3 non-GAAP EPS of $3.32 was up 96% year-on-year.

Now I'll review the P&L for our 2 segments, starting with Semiconductors. Revenue for our Semiconductor Solutions segment was a record $20.8 billion, up 127% year-on-year and represented 70% of our total revenue. AI semiconductor revenue of $16.7 billion represented 56% of total revenue, up from 49% in Q2. Gross margin for our Semiconductor Solutions segment was approximately 76%. Operating expenses of $1.2 billion reflected investments in R&D with OpEx representing 6% of segment revenue. Operating margin of 61% was up 440 basis points year-on-year as revenue growth of 127% outpaced operating expenses, which grew 22% year-on-year.

Moving on to Infrastructure Software. Revenue of $8.8 billion was up 29% year-on-year and represented 30% of our total revenue. Gross margins for Infrastructure Software was 94% in the quarter, and operating expenses were over $900 million. Q3 software operating margin was up 650 basis points year-on-year to approximately 84%.

Moving on to the balance sheet. We ended the third quarter with $24 billion of cash compared to $19.6 billion in the prior quarter, up $4.3 billion sequentially. We ended the third quarter with inventory of $4.5 billion to support our strong semiconductor demand. Moving on to cash flow. Free cash flow in the quarter was a record $13.7 billion and represented 46% of revenue. We spent $532 million on capital expenditures in the quarter.

Turning to capital allocation. In Q3, we paid stockholders $3.1 billion of cash dividends based on a quarterly common stock dividend of $0.65 per share. In Q3, we also paid down $5.6 billion of long-term debt. Subsequent to quarter end, we paid down an additional $1.5 billion of senior notes upon maturity. The weighted average coupon rate and years to maturity of our gross principal fixed rate debt of $59.6 billion is 4% and 7.4 years, respectively.

In June, we established the AI XPV platform in partnership with Apollo and Blackstone to enable more than 20 gigawatts of compute infrastructure for OpenAI and Anthropic by the end of 2028. We closed the first $35 billion tranche in June for Anthropic 1 gigawatt deployment, which is already underway.

While future tranches will include unique features tailored to specific lab and investor needs, our core strategy remains consistent. First, we are empowering 2 of our most strategic customers, the leading AI labs, to bridge the gap between their current cash flow and the significant upfront investments required for their businesses. Our XPUs enable cost-effective, sustainable growth at multiples of their infrastructure costs. Second, this XPV platform enables highly investable customers and facilitate execution where demand is already locked in.

And third, we review any strategic financing through a commercial as well as balance sheet lens, consistent with our existing capital allocation framework. Through the XTB platform, we partner with sophisticated third-party financial partners to independently underwrite and capitalize the assets rather than providing the direct financing ourselves. Where necessary, we may provide modest residual value guarantees, which are contingent liabilities, we view as low risk, supported by the strong profitability trajectory of these labs and the sustaining value of the underlying assets.

Moving on to guidance. Our guidance for Q4 is for consolidated revenue of $34.8 billion, up 93% year-on-year. We forecast semiconductor revenue of approximately $26.1 billion, up 136% year-on-year. Within this, we expect Q4 AI semiconductor revenue of $21.7 billion, up over 236% year-on-year. We expect Q4 Infrastructure Software revenue of approximately $8.7 billion, up 24% -- 25% year-on-year.

Moving on to margins. As the proportion of AI revenue accelerates in Q4, we expect our Q4 consolidated gross margin to be approximately 73%, down from 78% a year ago. As we have discussed previously, this reflects the increasing mix of XPUs with their increasing memory content, which is diluting our consolidated gross margin. Regardless, we expect Q4 operating margin to be approximately 66%, flat from a year ago because our strong revenue growth drives substantial operating leverage.

We expect the non-GAAP tax rate for Q4 and fiscal year 2026 rate to be approximately 16% due to the impact of the global minimum tax and the geographic mix of income compared to that of fiscal year 2025. We expect the Q4 non-GAAP diluted share count to be approximately 4.94 billion shares, excluding the impact of potential share repurchases. And in Q4, we expect capital expenditures of $1.4 billion as we invest in capacity for semiconductors.

As Hock mentioned in his remarks, we expect AI revenue to double again to approximately $115 billion in fiscal 2027 and double again in fiscal 2028 to $230 billion. This AI revenue guidance is being provided to you to give you the trajectory of our growth, and we do not intend to update it on a quarterly basis.

That concludes my prepared remarks. Operator, please open up the call for questions.

[Operator Instructions] And that will come from the line of Joseph Moore with Morgan Stanley.

Congratulations on the results. You talked about the business doubling next year. and you said demand could be higher than that. Can you talk about the supply around that? And kind of what are the supply bottlenecks and what are the variables that could drive that number higher if you're able to resolve them?

Well, Joe, that's a loaded question. Whatever I tell you, you guys go on and bring a bigger number. I think that's funny. And we're very careful. And to be honest, we try to be conservative. So we're giving you an outlook. And yes, demand -- we can ship significantly more. Question to some in our mind sometimes is, are they going to be even as we ship the chips going to be deployed on a timely basis. And that's always very much in our mind when we give you that outlook.

But certainly, our customers want us to ship more. But we have secured supply, and we think it's the right number to put it to $115 billion. And in times, if circumstances change and our ability to scale more supply chain, of course, we will uplift. But at this point, that's our best outlook.

And by the way, the same applies -- the same thinking applies to 2028 when we give you that outlook of $230 billion. This is real demand, we believe, based on what data center sites locations are ready 2028 with respect to our customers, the size of what we have and against the supply chain we have in leading-edge wafers, substrates and HBM memory. And so this is, again, a carefully structured outlook that we believe we can achieve.

One moment for our next question. That will come from the line of Blayne Curtis with Jefferies.

I actually want to follow up on Joe's on supply, the key point. Can you just talk about from either a substrate or kind of interposer [indiscernible] replacement? Has any XPU decided to use your Singapore capacity? And how does that fit into that supply picture you put together?

Well, we are going to start deploying our Singapore fab for substrates. By the way, starting fiscal '27. And that would, I guess, address a key part of our supply bottlenecks. Anything else?

Maybe you...

Finish your question.

No. I was going to give another one because I was to answer -- just 1 clarification quickly. The Google number you gave and then you said 10 gigawatts for Anthropic, are those -- it's not 1 number, right? The 10 gigawatts is just for Anthropic?

10 gigawatts is just for Anthropic.

One moment for our next question. And that will come from the line of Harlan Sur with JPMorgan. .

Hock, your customers are driving their XPUs, their GPUs higher in performance. The networking bandwidth is also scaling accordingly, right? Your customers are now moving from 100 gigabits per second per lane to 200 gigabit per second per lane from Tomahawk 5 to Tomahawk 6. We've heard that Tomahawk 6 ramp has been the fastest ramp of your -- any of your switching families. And we've also heard that you guys are almost sold out for Tomahawk 6. But next year, I'm not sure if you can give us an update on that.

And it's also good to see the team picked out its next-gen Tomahawk 7 with 200 gig SerDes. Out of your 6 frontier model-based XPU customers, how many of them are using Tomahawk scale-out network? And can you also just give us an update on the adoption of your Tomahawk Ultra scale-up platform as well?

Charlie will take this question. He knows everything here. .

Thank you, Hock. And thanks, Harlan, as well. You're right, our Tomahawk 6 has been a phenomenal success. It started really first in scale out, as you mentioned, -- and it's available both in 100-gig and 200-gig SerDes. So we actually have 2 versions of Tomahawk 6. Both are very successful, and both are deployed in pretty much all of the AI hyperscalers that are building XPUs with us. And even those that are not using our XPUs, are using the Tomahawk 6, both 100 gig and 200 gig.

With respect to Tomahawk Ultra, we have quite innovated ahead of the market here by enabling scale-up using low-latency Ethernet. The adoption also on this device has surprised us, and we're starting to see it deployed starting actually this quarter and in FY '27 coming up in scale-up applications.

Just to amplify what Charlie is saying with Tomahawk Ultra is that we are enabling now scaling up within cluster of GPU, XPU in a rack on basically Ethernet for the first time because it won't -- Tomahawk Ultra will perform just as well as anything else that exists prior to that before being able to do it on Ethernet, particularly with respect to loss and latency.

One moment for our next question. And that will come from the line of Stacy Rasgon with Bernstein.

I just wanted to verify, if I add up the gigawatts for '27 and '28, [indiscernible] I'm still getting roughly 10 for '27, about 6 of which are Anthropic and OpenAI, and roughly 20 for '28 roughly 15 of which are Anthropic and OpenAI. I just want to make sure that was the case. And if it is, if I sort of calculate the revenue per gigawatt, I come out with your revenue guys, kind of somewhere between $11 billion and $12 billion per gigawatt. I mean your competitors suggested something like [ 40. ] And I'm just wondering, is this the right level that we should think about for your content and should that, I guess, trend as we go forward, as you said, into more generations of XPUs again, with higher performance and higher memory, like how do those trends. So I guess, first, just the gigawatt count and second, the content.

Right. Stacy, you're very clever. You put too tricky questions into one. So let me take it. Let me pass it one by one. Yes, you can count the number of gigawatts we are outlining, not everyone because we are focused on, to be fair, with 6 customers, 4 of them are just simply going to be huge. And we outlined -- we walk through with you guys, their journey into deploying XPUs.

As I said, it's a journey Google has been doing for the last 10 years. Others, OpenAI last 2, 3 years, Meta last 3 years. Each one is doing it differently, and we like to take you guys through it. But more importantly, what it means in the deployment at XPU for -- for the -- for these 3 years, '26, '27, '28.

And so we laid -- you're right, Stacy, you add up the gigawatts. So we deploy, we are showing where they are headed. One we didn't tell you specifically when we came up to the final number is how many of these gigawatts do actually come out for actual deployment because deployment doesn't just include getting the chips out there before you get the chips out there, you've got to get the data center shell in place and ready for literary production. And we're talking about fiscal years. So what we're saying is, as you add up the gigawatts, we're not saying that over the next 2 years, '27, '28, then there are 30 gigawatts that will go into production, and therefore, we ship it. We think we judge it conservatively to be somewhat less, but we see the demand that if they can get it up in place and our products, and we can ship those chips regs in place, it will be 30 gigawatts between the customers we have, the 6 customers we have.

Question is, will all 30 come into production within this 2-year fiscal years. And we are giving you a judge number of $115 billion of chips to these guys in fiscal '27, another $230 billion in '28, which would add up, I know, to about $350 billion. So another way of saying is we believe pretty -- with a pretty high degree of confidence, we won't ship $350 billion of AI semiconductors to these customers in the next 2 years. That's the best way to look at it. It doesn't necessarily mean 30 gigawatts need to have been deployed within that period.

Turning to your more next interesting question, right. As I said, when you do an XPU, it's not only performed as well, if not better for your particular LLM workloads for each of our customers. It's also half the cost. In fact, less than half the cost. And that's -- you made my point exactly.

One moment for our next question. That will come from the line of Ben Reitzes with Melius.

This is [ Jack Ader ] on for Ben. Could you shed a little bit more light on the maximum off-balance sheet risk for the backstop agreements. Your last Q showed that the first tranche had a maximum exposure of about $29 billion. Is this about the order of magnitude for each of the Anthropic and OpenAI gigawatts over the next few years? And can you add a little more color as well on the residual value and the risk associated?

Thanks, Jack. Listen, we don't have anything to announce today on residual value guarantees or backstop. So there's nothing new to add. So the numbers that you outlined around what we've already done remain true. And as I mentioned in my prepared remarks, we're going to evaluate any strategic financing really on a deal-by-deal basis. And we expect that any that we do in the future is they're going to have unique features and they're going to be tailored specifically to the lab and to the investor needs. So I can't give you an overarching look at like what's the max and what each one is going to look like. But we will tell you at the right time, we have nothing to announce today.

One moment for our next question. That will come from the line of Vivek Arya with Bank of America.

Amie, I just wanted to clarify any impact on gross margins in '27 and '28, given this XPU mix and rise in memory cost.

And then, Hock, you mentioned the 2 frontier labs will be your largest AI customers, even though Google continues to say they are supply constrained. So why aren't they taking more. And in many cases, these frontier labs depend on land, power shell from their cloud service providers, sometimes NVIDIA-related entities. So how much of their use of silicon is truly their choice versus what their CSP or funding partner dictate? So I'm just curious what gives you the certainty that they will give Broadcom, right, that specific business in '27 or '28 when so much of their funding depends on other cloud service providers who might have a different view when it comes to the choice of silicon in that specific facility.

I know the best way to answer that -- a couple of ways I have of answering that and a couple of points. Don't forget, the rate -- with the rate of growth these startups. And we're talking about -- you know we have 6 customers, and we are creating this financial vehicles, as Amie described them for just 2 of them. Not all. The other 4 customers of ours are financially secure, stable enough to be able to fund it themselves and we're happy for them to do that and get them up to where they need to with technology, shared technology, which we have in plentiful supply.

With these 2 guys, Anthropic and OpenAI, I mean this is a thinking of I mean, you have 2 geniuses in the middle of [indiscernible] Mongolia, say, and they need to go to college to fulfill where they want to. So we do what we can to help them. And part of it is creating a -- creating sources of financing to help these companies with the leading edge frontier models in the world, be able to play in the same playing field and be able to offer these great technology products to the world.

And that's simply what we're doing. It's a great investment for us when you think about it that every gigawatt of compute they deploy, they could achieve $30 billion of ARR annual revenue per gigawatt. That's a hell of a business model. So for us as a great investment to focus on doing. So to put that simply, that's how we look at this, a simple thing, it makes economic sense for Broadcom to invest and enable these guys.

Now these kinds are going to be hyperscalers in their own right. So what they do is they want first-party compute capacity just as they create their own -- they want to create silicon that is very cost, performance optimized, and they want to run their own data centers, eventually first batch, not even eventually, they want to run it as fast as they can.

Short term, you're right. They are using cloud services, third-party services to deploy their models. Long term, we see these guys to be no different from a hyperscaler and will run their own data centers and be first party to offer AI generative APIs, assess models to the world. So we see that happening. It's not speculation. It's actually happening, and we are in the midst of enabling that.

And I'm happy to take your gross margin question. First, I want to just make sure I correct what I said before. Just to be super crisp. Gross margin for our Semiconductor Solutions segment was approximately 67% in Q4 -- or Q3. And to your question, gross margin really reflects revenue mix between both Semiconductors and Infrastructure Software. And then within Semiconductors, it also is -- the product mix is reflected as well as increasing memory content. And so as you can see, as XPUs become a larger proportion of our total revenue mix, it impacts our margin. We guide 1 quarter at a time. So we'll tell you each quarter how -- what our margin is going to look like.

Yes. It's not the first time we have told you guys that because stop focusing on gross margin is what we're saying, look at where matters, operating margin at the bottom -- at the end of the day because the growth in revenue far out surpasses the growth in OpEx, operating spending to support that growth in revenue. So we have a lot of accretion operating leverage, we call it, on operating margin. So we expect to be able to sustain operating margin even as mix of products dilute the gross margin.

One moment for our next question. And that will come from the line of Tom O'Malley with Barclays.

This one is for Hock and Charlie. So you guys talked a bit about the Tomahawk Ultra, you're ramping a variety of basics over the next couple of years. How should we think about the attach rate of Tomahawk Ultra to the AC that you're developing? I would assume that when a customer goes down the road with you for an ASIC, they would decide to use your networking as well. Maybe some kind of forecast as to how many of those customers are using your Tomahawk Ultra? How are you using Ethernet just because I assume your penetration with your accelerators is going to lead to a lot of success on the networking side. Any help there would be great.

Go ahead, Charlie.

Thank you, Hock. Yes, on the attach rate for scale up, today, we're actually seeing customers who build XPUs with us, deploy either Tomahawk 6 used to deploy Tomahawk 5. Now they're going to Tomahawk 6. And some of them are going to Tomahawk Ultra. So if you look at where we're seeing XPUs and even GPUs. We're seeing scale-up solutions adopt both Tomahawk 6 and Tomahawk Ultra and the beauty and the reason why they're doing this, it's as Hock articulated earlier on, it's Ethernet-based, which means it's open, anybody can connect to it, especially with the standards that we have collaborated with the entire industry on. So at this point in time, we're seeing it pretty much deployed in both XPU clusters and in some GPU clusters.

One moment for our next question. That will come from the line of Will Stein with Truist Securities.

Congrats on the good results and the very impressive longer-term outlook you gave. Hock, hope you can tell us a little bit about the constraints as it relates to land power and shell. You mentioned this a little bit earlier, but I wonder to what degree the outlook is sensitized to potential constraints in that area?

Good question, Will. Of course, we have to be realistic and when we provide our outlook, we not just look at -- just look at what our customer, our end customer, 1 of those AI frontier models, just asks of us for compute capacity in the form of chips or in some cases, even in the form of racks. We are very engaged with each of them on how much or which -- how much LPS, land, power and inside -- systems they have before we actually believe it will happen because this -- some -- this part on site power shell has a long lead time. The construction project, you correctly indicate that. So we do that. And we do that on development analysis with our customer, and we reflected in the forecast outlook we're giving you today. .

One moment for our next question, and that will come from the line of Joshua Buchalter with TD Cowen.

With XPU, I think you have $35 billion of the financing secured. As we think about the 10 gigawatts for Anthropic and the 5 for OpenAI in fiscal 2028, are you expecting most or all of this to be financed through the XPV? And can you give us sort of any help on the time line and hurdles required to secure financing as we look forward to those deployments getting secured?

Let me start broadly and Amie will give you specifics. Not everyone, not every site will be the same. Because don't forget, next 2 years, we'll see things happen between -- from -- with Anthropic and OpenAI. And it's an open secret Anthropic is well on its way to an IPO. And with that, its investment credit will change. OpenAI, we still don't -- have less clarity, but that's different. And so with that, let me pass it to Amie.

I mean I think just to double down on that, I think in every scenario, our partners were going to be looking for financing from lots of different sources. And in some cases, we may step in and help out it, if it makes sense to us and if it's within our framework. And in other cases, we may not. But we're going to look at it case by case and evaluate the specific needs.

One moment for our next question, and that will come from the line of Jim Schneider with Goldman Sachs.

I was wondering if you can maybe address some of the other constraints that are impacting your outlook. You mentioned land power shell. Do you think that is the largest of the constraints that you face heading into fiscal '27? And specifically, can you address any other supply chain constraints, whether it be memory substrates or other components that could be impacting your outlook? And to what extent you could expect those to get ameliorated?

Thank you, Jim. That's a hell of a question. And you're right on all counts. It's a multidimensional issue that to get an AI data center deployed, yes, we now are all more sensitive about -- especially at scale and the scale -- we are now delivering at scale to our customers, XPUs, AI data center built upon XPU computing accelerators.

Land, power and shell, as I indicated, to question with Will, is a big concern. It more than a big concern, it dictates specific timing of when this capacity gets deployed and be available. But very much in the mix that we have to think about with our customers. And it's a joint collaborative effort, not just 1 direction. It's -- you said it correctly, leading-edge silicon because for 1 to produce -- to get the chip produced and availability of those chips and the time it comes in.

Then even deeper than that, we talk about substrates, which is specific issues, which is leading us to build our own substrate capacity at scale in our factory in Singapore. Jointly, we won our partners. And of course, we all know about memory, HBM memory. And beyond HBM memory, the system memory that goes into AI servers, which we don't supply necessarily, but our customers have to secure too. So all this is a multidimensional problem coming from various sources. And then depending on different times, each might become a bottleneck. And so it's a constant interesting challenge as we work this through, which is why in some ways, I'm so glad we have only 6 customers to deal with.

Our next question that will come from the line of Vijay Rakesh with Mizuho.

Hock and Amie, just a quick question. Thanks for giving the visibility of fiscal '27 and fiscal '28 AI revenues. Just a quick question. As you go through subsequent generations of XPU, can you talk to how your dollar per gigawatt should improve into the subsequent XPU generations? And also, I saw your CapEx went up. Just wondering if you're adding capacity on the [indiscernible] side.

Well, Charlie, you want to take the CapEx issue?

We -- so on the CapEx side, as we've been sharing with you, Hock and I, for several quarters, we continue to invest in our factories. Substrate is what Hock talked about, which was actually going in production shortly. But also our EML CW and VCSEL factories, our Indium phosphide factories, both in the U.S. as well as in Singapore, we're actually more than tripling them year-on-year. So we've already expanded the capacity for this year, and we're increasing it significantly for the next 2 years, and that's part of what you actually are seeing in CapEx.

And this one I indicated in my remarks, which is we demand for lasers, whether it's EML lasers, CW lasers is fast surpassing supply out there in the industry. So we are doing a path to double down actually on capacity, and we have a fairly substantial share of this market to us in our interest to enable this ecosystem of growth.

On your earlier question, a gigawatt -- dollars per gigawatt, it's very interesting what you're saying because keep in mind something that is very interesting, too, which is, you're right, with every generation of XPU or GPU, the performance increases. And therefore and the silicon goes for the leading edge, more expensive to produce. And so ASPs go up for XPU and per GPU. But keep in mind, as they become higher performance, the power increase per chip, per XPU, per GPU, which means that less of it of a more advanced XPU in 1 gigawatt.

So what we are seeing per gigawatt is in the range of less than $30 billion, $20 billion to $30 billion per gigawatt. And we expect that to be very sustaining in that level because the power of each chip goes up. So even it increased the price of the chip, the content in dollars is relatively stable, just except the fact that there's going to be a lot more gigawatts out there. But the dollars per gigawatt will remain in the $20 billion to $30 billion level. of our content, but we have seen and we expect to see an acceleration in the number of gigawatts just to address the exponential compute demand required by our customers.

That is all the time we have today for Q&A. I would now like to turn the call back over to Ji Yoo for any closing remarks.

Thank you, Cherie. This quarter, Broadcom will be presenting at the Goldman Sachs Communacopia and Technology Conference on Tuesday, September 8. Broadcom currently plans to report its earnings for the fourth quarter and fiscal year 2026 after close of market on Wednesday, December 9, 2026. A public webcast of Broadcom's earnings conference call will follow at 2:00 p.m. Pacific. That will conclude our earnings call today. Thank you all for joining. Cherie, you may end the call. .

Thank you all for participating. This concludes today's program. You may now disconnect.

Welcome to Broadcom Inc.'s Second Quarter Fiscal Year 2026 Financial Results Conference Call. At this time, for opening remarks and introductions, I would like to turn the call over to Ji Yoo, Head of Investor Relations of Broadcom Inc.

Thank you, operator, and good afternoon, everyone. Joining me on today's call are Hock Tan, President and CEO; Charlie Kawwas, President, Semiconductor Solutions Group; and Ram Velaga, President, Infrastructure Software Group. Also joining is Kirsten Spears, Chief Financial Officer. As we announced, Kirsten will be retiring June 12, and today, we have joining us our incoming Chief Financial Officer, Amy Tiner. Thank you, Kirsten, for your leadership over the past 12 years.

Broadcom distributed a press release and financial tables after the market closed, describing our financial performance for the second quarter fiscal year 2026. If you did not receive a copy, you may obtain the information from the Investors section of Broadcom's website at broadcom.com. This conference call is being webcast live and an audio replay of the call can be accessed for 1 year through the Investors section of Broadcom's website.

During the prepared comments, Hock and Kirsten will be providing details of our second quarter fiscal year 2026 results, guidance for our third quarter of fiscal year 2026 as well as commentary regarding the business environment. We'll take questions after the end of our prepared comments. Please refer to our press release today and our recent filings with the SEC for information on risk factors that could cause our actual results to differ materially from the forward-looking statements made on this call.

In addition to U.S. GAAP reporting, Broadcom reports certain financial measures on a non-GAAP basis. A reconciliation between GAAP and non-GAAP measures to the extent possible is included in the tables attached to today's press release. Comments made during today's call will primarily refer to our non-GAAP financial results.

I will now turn the call over to Hock.

Thank you, Ji. Well, in Q2 revenue was a record $15 billion as we grew 79% year-on-year. Driving this growth was AI semiconductor revenue at a record $10.8 billion, 143%... sorry.

Well, let me restart, guys. Thank you, Ji. Thank you, everyone, for attending joining today. In our fiscal Q2 2026, total revenue reached a record $22.2 billion, up 48% year-on-year, above our guidance on strength in AI semiconductors. Q2 operating margin was a record 67%, and adjusted EBITDA was a record 69% of revenue, which was above our guidance. Even as our revenue scales up massively driven by AI, our operating and EBITDA margins remain strong and stable.

Turning to semiconductors. Q2 revenue was a record $15 billion, as I said before, as we grew 79% year-on-year. Driving this growth was AI semiconductor revenue at a record $10.8 billion, up 143% year-on-year and above our outlook. Networking represented almost 40% of our Q2 AI revenue. Demand for XPUs and networking is simply insatiable. During the quarter, bookings for AI semiconductors were over $30 billion against the $10.8 billion we shipped.

In the second half of 2026, we expect AI semiconductor revenue to double from the first half we shipped last -- from the ship this year. Consistent with this trend in Q3, we expect AI semiconductor revenue to accelerate to $16 billion, up over 20 -- to up over 200% year-on-year. For the full year 2026, we expect to achieve AI semiconductor revenue of $56 billion, up approximately 180% from fiscal 2025. Now we expect this momentum to continue into fiscal year 2027, and reiterate our AI semiconductor revenue guidance to be in excess of $100 billion. We expect AI semiconductor revenue growth to continue in fiscal 2028 based on the following initiatives with our 6 core customers.

As you are aware, with Google, we announced in April that we entered into a long-term agreement to develop and supply multiple generations of TPUs and AI networking. Our relationship continues to be strategic and very substantial as we continue to deliver vastly superior technology and execution compared to other alternatives. This ability to provide differentiated value to Google ensures that our business will sustain and grow for the foreseeable future.

For Anthropic. As you know, for 2026, we are providing access to Broadcom TPU-based compute of over 1 gigawatt. In April, we entered into an agreement to enable Anthropic to access another 5 gigawatts of next-generation TPU-based compute beginning in 2027. For OpenAI, we have delivered silicon, and we are on track for production late 2026. We have a contractual commitment to deploy 1.3 gigawatts in 2027 as part of the larger 10 gigawatts that we by 2029 agreement we announced last year.

For Meta, in April, we announced a partnership to deliver multiple generations of MTIA XPUs. And under this agreement, we expect to deploy 3 gigawatts through the end of 2028. The initial order for 1 gigawatt, which includes XPUs and our networking has been received and we'll start delivering in the second half of 2027. For our other 2 customers, we expect shipments to begin late 2026 and accelerate into 2027. To date, we have received purchase orders totaling $6 billion.

While we have significant IP and execution leadership in XPUs, networking is key to building scalable XPU and GPU clusters. And here in networking, with at least one generation of technology and product leadership. For scale up within racks, we enable direct attached copper based on an industry-leading 200G and 400-gig serdes, driving co-packaged copper with Ethernet and PC Express switches. For scale-out between racks, we have been shipping the industry's only 100 terabit Ethernet switch, the -- for over a year. We will now be taking out our next-generation 200 terabit switch this quarter. And in CPOs, which is co-package optics 1.6-terabit DSPs, CW and EML lasers, we are the de facto standard in the industry. To extend AI and that cost us across data centers, we remain the industry leader with our Jericho3 and Jericho4 fabric solutions, enabling the world's largest deployments of and multiple hyperscalers.

Our strategic vision is to bring together Broadcom's leading technology and investor partners with the strongest balance sheet to deliver at scale, sufficient compute capacity at the lowest cost and power for the leading AI frontier labs, including Anthropic and OpenAI. To deliver this vision, we are creating the AI XPU platform with Apollo and Blackstone and other leading investors to deploy more than 20 gigawatts of compute capacity through 2028. The first tranche of this platform valued at $35 billion is, in fact, currently been launched by Apollo.

Now turning to non-AI semiconductors. Q2 revenue of $4.2 billion was up 6% year-on-year. Bookings during the same period exceeded $6 billion, which is a clear indication we on the path towards a full cyclical recovery. Broadband, server storage, and enterprise networking together were up, partially offset by a seasonal decline in wireless. Consistent with this trend in Q3, we forecast non-AI semiconductor revenue to be approximately $4.5 billion, up 12% from a year ago.

In summary, we expect Q3 semiconductor revenue to be $20.5 billion, up 124% year-on-year.

Let me turn to Infrastructure Software segment. Q2 software revenue of $7.2 billion was up 9% year-on-year, in line with our guidance. Bookings continue to be strong as we sustain ARR growth of 17% year-over-year. For Q3, we forecast software revenue to be approximately $8.9 billion, up 31% year-on-year. We just released VMware Cloud Foundation 9.1, focused on improving infrastructure efficiency, security and support for enterprise AI inferencing workloads. With strong server demand globally, the deployment of VCF9.1 for on-prem cloud computing is extremely strong, driving robust revenue growth. This release adds heterogeneous compute support across GPUs and CPU architectures, including AMD, Intel and NVIDIA platforms, enabling customers, enterprise cloud customers to run AI Kubernetes and traditional virtualized workloads on a common private cloud environment.

So to sum it up for Q3 2026, we expect our consolidated revenue to grow to $29.4 billion, up 84% year-on-year. We expect operating margin to be stable at approximately 67% of revenue and adjusted EBITDA to be at approximately 68% of revenue.

And with that, let me turn the call over to Kirsten.

Thank you, Hock. Let me now provide additional detail on our Q2 financial performance. Consolidated revenue was a record $22.2 billion for the quarter, up 48% from a year ago. Gross margin was 77.1% of revenue in the quarter down 230 basis points year-on-year as semiconductor became a larger proportion of our product mix. Consolidated operating expenses were $2.2 billion, of which $1.6 billion was R&D. Q2 operating income was a record $14.9 billion, up 52% from a year ago. Note that even with the decline in gross margin, operating margin increased 200 basis points year-over-year to 67.3% as operating expenses remained relatively flat. Adjusted EBITDA of $15.2 billion or 69% of revenue was above our guidance of 68%.

Now a review of the P&L for our 2 segments, starting with semiconductors. Revenue for our Semiconductor Solutions segment was a record $15 billion with growth accelerating to 79% year-on-year driven by AI. Semiconductor revenue represented 68% of total revenue in the quarter, and AI semiconductor revenue represented 49% of total revenue.

Gross margin for our Semiconductor Solutions segment was approximately 70%. Operating expenses of $1.2 billion reflected increased investment in R&D for leading-edge AI semiconductors and represented 8% of revenue. Semiconductor operating margin of 62% was up 460 basis points year-on-year, reflecting our strong operating leverage.

Now moving on to infrastructure software. Revenue for infrastructure software of $7.2 billion was up 9% year-on-year and represented 32% of revenue. Gross margin for infrastructure software was 93% in the quarter and operating expenses were $1 billion in the quarter. Q2 software operating margin was up 310 basis points year-on-year to approximately 79%.

Moving on to cash flow. Free cash flow in the quarter was a record $10.3 billion and represented 46% of revenue. We spent $231 million on capital expenditures. We ended the second quarter with $19.6 billion of cash compared to $14.2 billion in the prior quarter. We ended the second quarter with inventory of $4.3 billion as we continue to secure supply to support strong AI demand. Our days of inventory on hand were 86 days in Q2 compared to 68 days in Q1 in anticipation of accelerating AI semiconductor growth in the second half of the year.

Turning to capital allocation. In Q2, we paid stockholders $3.1 billion of cash dividends based on a quarterly common stock cash dividend of $0.65 per share.

Now moving to guidance. Our guidance for Q3 is for consolidated revenue of $29.4 billion, up 84% year-on-year. We forecast semiconductor revenue of approximately $20.5 billion, up 124% year-on-year. Within this, we expect Q3 AI semiconductor revenue of $16 billion, up over 200% year-on-year. We expect Q3 infrastructure software revenue of approximately $8.9 billion, up 31% year-on-year.

Moving on to margins. As the proportion of AI revenue significantly grows in Q3, we expect Q3 consolidated gross margin to be down to approximately 74%. This decline in gross margin does not represent a structural change in semiconductor margin. Rather, it reflects product mix between semiconductors and infrastructure software. Regardless of the impact to gross margin, we expect Q3 operating margin to be 67%, which is flat quarter-on-quarter, demonstrating our strong operating leverage. We highly recommend that investors model semiconductor and infrastructure software margins separately to properly reflect the impact of changes in total revenue mix going forward. We expect the non-GAAP tax rate for Q3 and fiscal year 2026 to be approximately 16% due to the impact of the global minimum tax and the geographic mix of income compared to that of fiscal year '25. In Q3, we expect a non-GAAP diluted share count to be approximately 4.94 billion shares, excluding the impact of potential share repurchases.

That concludes my prepared remarks, Operator, please open up the call for questions.

[Operator Instructions] And our first question will come from the line of Harlan Sur with JPMorgan.

Thanks for all your support, Kirsten and Amy, welcome to the team. First, just a house -- quick housekeeping item, Hock. On this fiscal year, AI sort of 2x growth second half over first half. That will put AI revenues over $60 billion with sequential growth in fiscal Q4, but you gave us this $56 billion number, which is only like 1.5x half-over-half growth with AI actually being down sequentially. So if you could just help us kind of square the numbers there? And then for my real question, back in December of last year, you talked about in the AI backlog in the next 18 months, $73 billion. Market sort of took that number, spread that linearly over 6 quarters. But we know that the backlog is always more front-loaded over the first 4 quarters, right? And sure enough, like you're going to deliver around 80% or more of that backlog in this fiscal year or first 4 quarters.

Just given the strength of all our programs, the broadening of the customer base, accelerating year-over-year trends in your AI shipments, all the multi-gigawatt partnerships that you just articulated today, which is -- most of it, which is set to start to fire next year. Is it fair to assume that your 18 months AI backlog second half of this year to first half of fiscal '27 sits at $200 billion or better?

That's a very complicated set of number questions. To begin with last time with '26, doing a math basically 2x to 2x. The first half, we ship in total AI revenue, something in the range of $19 billion. You're going to be precise. So -- and if we do what I indicate in 2x that in the second half, you get to pretty much in the range of what we're talking about, which is around $56 billion, Harlan. So that number is still very, very -- does tie up very well.

Now your bigger question on the second half, which you've gone in a very detailed analysis of this. Yes. we keep the momentum going as we expect to see in 2027, what we will see in 2027, continued growth of the level we're talking about. And if you drive on that basis of what we're seeing here, almost 2x -- 2,000 -- in the range of 2x, what 2026 will be. I think you will easily see that 2027 will very easily $100 billion in 2027, which is pretty much what we indicated last quarter, and we are continuing to say that it will be over $100 billion in 2027. So in that sense, if anything else, it might be based on what we're doing very much on track, if not stronger. But -- we're not trying to guide you every quarter what '27 would be like. So we basically say we continue to be in excess of $100 billion in '27. But it is on the same trajectory as we are seeing in the back half of '26.

One moment for our next question. And that will come from the line of Blayne Curtis with Jefferies.

Hock, I wanted to ask you, inter-quarter, you had that 8-K with the long-term agreement with Google. I think, obviously, you're probably not going to tell me what the total value is there, but I think there's a lot of concern about share within that customer. I was just kind of curious your -- now that you have this agreement, maybe you could speak to a little bit more in terms of your confidence. And if there's upside to that customer, is it a fixed amount? Or is there share? Is there any way you can kind of add some color to that agreement that came out?

Well, it's -- it's a very, very strong agreement, and it basically reflects the strength of the partnership we have, simply because of the products we do, the multigeneral products, and the intellectual property we deploy into this whole program.

To answer your question specifically, it's a commitment that is very substantial in dollars, very, very substantial amount of dollars. Now we also accept the fact that while we like to win every design in that program. We also accept the fact that given the growth of consumption of and development and consumption of AI compute even by our partner, Google, that we fully expect that there will be some diversity of sources for them. But our commitment from them is a very substantial dollar mark.

One moment for our next question, and that will come from the line of Ross Seymore with Deutsche Bank.

Congrats to both Kirsten and Amy. Question on the gross margin side of things. I know, Kirsten, you talked about it going down due to the mix dynamic within the semis versus the software side. Given the strength in the software side in the quarter, seems like the gross margin is falling a little bit harder. So behind the scenes, can you just talk a little bit about what the drivers within semis are? Is that the XPU versus the networking side of things? And is that trend likely to continue next year? Are there rack scale versus chip scale, all those sorts of dynamics? Any color you could give on that would be helpful.

Yes. Certainly, as our semiconductor business growth, just to reiterate, on a consolidated basis relative to our software business, you're going to have a decline in margins, right, a bit, right? You'll have compression. But remember that we're -- it's accretive on because we have strong operating leverage, right? So our operating margins will stand up a bit over time to that.

Within semiconductors, we've always said our ASICs, TPUs, some of the wireless business has lower margins. So as the TPUs continue to accelerate, there'll be pressure overall on margins. But the connectivity side, the AI networking side of the business has very rich margin. So it will offset it somewhat as we go.

I mean, Ross, as Kirsten said in her remarks. Structurally, the semiconductor margins remains very stable and very solid. It's the mix, particularly mix between software and the non-AI to the very, very high -- rapidly growing AI semiconductor that is just diluting gross margin.

And the Rack versus chip side of things? Is that all clarified now? .

No racks. It's chip.

Chip business only.

We own chips.

One moment for our next question. And that will come from the line of Ben Reitzes with Melius.

Appreciate it. Wanted to ask about 2027, Hock. With regard to previously, you talked about the TAM being and well, it's kind of a longer-term question actually. You've talked about the TAM being 10 to 20 and whatnot. It seems that one of your competitors talked recently about the TAM per gigawatt going up a lot as we go throughout the decade. And it seems it wasn't just due to infrastructure. It was due to the compute and networking components and other things. Perhaps you're familiar with the comment that Jensen made where the overall infrastructure is going from something around 50 something towards 100 and the compute content going way up. Are you seeing the same thing as you go throughout the long term? Is that potentially being an accelerator what you've already outlined in terms of your TAM per gigawatt, and how are you thinking about that?

Sure. Well, I think the accelerating part, if you talk about power, realize one thing, is the dollars per gigawatt, the content dollars of racks per gigawatt is not accelerating that much because you are creating chips that have -- each individual chips are driving higher and higher power -- so you're driving less chips, though the ASP of each chip is going up in price. So dollars per gigawatt -- billions of dollars per gigawatt is relatively stable. But the number of gigawatts will keep -- will accelerate as I think some of our remarks indicate. And that's what we are seeing. -- that amount of gigawatts required compute capacity, as mentioned by number of gigawatts is growing very fast. And we're seeing that -- and we are seeing that particularly to the point where for even two of our customers that we're talking about, which is Anthropic and OpenAI, for which we are creating this platform to enable them to run sufficient compute power. We're talking about capacity, as mentioned by gigawatt power that are way ahead of what we fully -- what we have expected, say, 6 months ago. And that's just these guys.

We don't talk about the consumption beyond the Plant XPU platform we have announced here from our other customers, which is Google-owned internal workloads, matters, workloads and any other case and the other 2 customers that we have. So fold that in, and you're talking about gigawatts in totality, you ask about '27 or '28, that will continue to grow. We expect, in fact, '28 to be a substantial growth from what we are forecasting in '27.

One moment for our next question. And that will come from the line of Timothy Arcuri with UBS.

I wanted to ask you about supply and kind of your ability to get incremental volume of wafers and HBM. As I look at some of your competitors, I mean, they're kind of able to drop $20 billion out of thin air and get incremental wafer supply. So I'm wondering, do you feel pretty good about like if a customer comes to you, are you able to get upside in terms of wafers and HBM and -- and are you beginning to consider maybe using other foundries to add more optionality to your supply?

Getting supply is not just about dropping money them though that does a -- No, we are very comfortable that we have been able to secure supply of the types you mentioned about 4 needs, '26, '27. Working on '28 and '29 right now.

Right. But if a customer comes to you and wants incremental supply, are you able to go to your suppliers and get it the way that it seems like some of your competitors are.

Customers have been coming to us incrementally over the last few months. We expect that to continue. And by and large, yes.

One moment for our next question. And that will come from the line of Stacy Rasgon with Bernstein Research.

Hock, you gave some gigawatt shipment targets for next year for your various customers. I just want to know, are those any different? Do they contemplate any change from what you said last quarter, where I think you said that was like close to 10 gigawatts, you'd be shipping in '27. And can you just sort of help us shape the year? It sounded to me like you expected that to be more back half loaded in '27 given the shape of the ramps. But most importantly, is there any change that is it more gigawatts or less gigawatts or the same gigawatts versus what you were suggesting last quarter?

Well, good question. Yes, for '27, we indicated about 10 gigawatts shipment in '27. There's still very much intact. They will be shipping -- we're planning to ship 10 gigawatts in '27. And that nothing has changed. Back half loaded to that extent, yes, and which really provides an interesting trajectory into '28 with this back half trajectory. So '28, we expect a lot more gigawatts.

One moment for our next question. And that will come from the line of Jim Schneider with Goldman Sachs.

I was wondering if you could comment a little bit on the profile of your networking business, Hock, as we head through fiscal '26 and '27, about 40% of AI revenue this quarter. Would you expect that to sort of fall back down as some of these custom ramps ramp into the end of the year into early next year? Or would you sort of expect to stay at the upper end of that range? And maybe talk about when you see some of the optical and CPO revenue becoming meaningful?

There's a hell of a great question, Jim. It's just very difficult to answer because of how do we predict them. There are quite a few moving parts there, one of which is to start with is as more as a volume as more and more of our customers turn to XPUs. Obviously, XPUs uses very much our -- a lot of our networking components across the board. So that's great for us. And that drives increase in consumption. But it also means that we have been able to sell networking to non-XPU footprints. So that part of it would dilute the growth rate. And this 40% I consider as a very -- well, I'm almost a situation where stars are aligned where we are shipping a lot of networking to non-XPU, while -- the growth of XPU obviously allowing us to grow this networking business to our XPUs and we get to 40%. But I see that as probably as high as that percentage of total AR revenue would go. Not the first time I indicated that the more expected percentage as a share of total AI revenue for networking would be closer to around 30%.

One moment for our next question, that will come from the line of Tom O'Malley with Barclays.

So I noticed with the most recent deal with Anthropic that you guys are using Broadcom chips as a backstop for the deal. Do you expect more deals to come like this in the future? And then as you start to see the AI environment, is there any way you're thinking about financing in the future? Are you going to continue to do it with chips? Or anything that you can offer on that?

Can you repeat that question, especially at the front end I didn't quite get what you're saying here, and I don't want to answer it the wrong way.

Sorry, Hock. Essentially, most recent deal with enthoropic is being backstopped by Broadcom chips. Do you think that in the future, you will see more deals done this way? And then any comments on the future financing of deals with the large AI models.

I have to correct you on that. deal -- then I'll deal with Anthropic and that we basically talk about basically release disclosed in our 8-K recently as the deal we did with traffic is we use our GPU chips that we developed to provide the compute capacity to Anthropic. We won that it wasn't backstop in that sense. We were the one providing the chips to Anthropic. We were the one providing the compute capacity Anthropic.

One moment for our next question. That will come from the line of C.J. Muse with Cantor Fitzgerald.

I guess, Hock, in recent years, you've talked about really focusing your efforts on very large XPU platforms. And I'm just curious, we're seeing many kind of XPU attached derivatives across interconnect, storage, other. And I'm wondering if there's any sort of programs there that are more niche that are whetting your appetite?

Well, no, I don't think so. I think our business model is actually very, very straightforward, which is we are developing XPUs, custom accelerators for use buying our customers who are pretty much all LLM developers, whether it's for training or inference. We are also creating a portfolio of critical components to enable these XPUs and even GPUs to be clustered and from better performance. And that continues to be the model we do, which is we provide chips technology in the form of chips, whether they be computer accelerators, we call XPUs, or networking chips that cluster them together, be it switches, PCI Express connectors, DSPs, lasers, NICs and routers. And that's very much still the model we employ in semiconductors. And we still, as you can see, our financial model and our -- the program we go to drive towards a chip business model through the technologies we provide.

What we are doing to enable some of these LLM players to be able to get access to the volume of compute capacity, the large gigawatt capacity they mean to scale up their models is we are, as I announced here today, creating in partnership with guys with the best balance sheets around vehicle to basically have this chips funded for these LLM players who otherwise might have difficulty getting access to our technology, which provides them with the lowest power and the lowest cost.

One moment for our next question, that will come from the line of Atif Malik with Citi.

I have a question on infrastructure software business. Are you guys seeing any impact of AI -- Agentic AI on your software growth and renewals? And if you can just talk about some sort of long-term growth for that business?

Well, we're not seeing it, if anything else, as I reported, the high volume of core count of CPUs, selling together with GPUs is driving some accelerated growth of our VMware business. And as you can see, in Q3, we're seeing an accelerated growth, and we expect that to continue again for the next multiple quarters as this demand picks up. Long term, given the kind of products we do in infrastructure software, very, very close to literally the hardware, the basically the hypervisor, which is where our products are, we do not expect to see any impact on software products.

And one moment for our next question. That will come from the line of Edward Snyder with Charter Equity Research.

This is very interesting because the gigawatts that you laid out for the different customers. It's very clear that the 2 that are offering kind of CSPs, don't want to see, but consumer versions of AI, Anthropic and OpenAI have very large gigawatt commitments in the out years. I know part of that is catch up because they've just started late where your oldest customers has been doing these for quite some time. But even part of Google's is offering cloud services to other folks, too. So are we seeing a shift here? I know that initially, a lot of the XPUs and the AI services were through the hyperscalers with their own customer workloads. We've talked about that ad nauseam. And now you're seeing -- I finally hit the enterprises and you've seen cloud take off with the programming, which is sweeping everybody. Can we expect then that there's going to be this big second wave of demand that's driven as AI starts seeing enterprises and consumer getting access to it or finding usable tools because the number you're saying here are significantly different for the 2 classes of customers?

Well, that's a very interesting target. And you may be very well right, that enterprise consuming AI is still relatively at an early stage of the game. But having said that, what we're also seeing is a lot of what the enterprise is consuming on tokens. They are buying a lot of these tokens from the platforms, the product plan, API platform API, they pull from the platforms of the same name of customers we talked about, which is Anthropic, OpenAI, Gemini. These are on the large guys, they're pulling it from. And that's where, I think, a high -- substantially, most of these tokens consumption are tied to those LLMs. And these LLMs guys as they productize their frontier models, whether it be ChatGPT 5.5 or Gemini 3.5, comes back to the same end demand on compute capacity we provide to all these guys. And so even the growth of enterprise demand that we're now starting to see as enterprise starts to consume AI, AI tokens, for their own workloads for their own productivity users as consumers do. They are buying from the same source, the same few guns. And that's one driving this very large, I call it, in scale growth in compute capacity that we are experiencing, and we see that continuing to happen now through 2027 and what we're seeing now through 2028 as well. So this is getting to be quite a sustainable and steepening trajectory of demand.

So if I could, doesn't this change the dynamic of what we talked about before, you talked about Lot 7 -- about so customers for your XPUs, -- but this is actually happening. We've already seen a Google offering cloud services for GPUs. It opens up XPU access to all those smaller companies that don't kind of meet the criteria for doing the own ASICs that you could partner with to Broadcom's technology through these platforms. Is it why would that not be the case?

Well, I guess the answer to that is it's possible, but the reality of the whole issue is this -- the compute capacity, most of the AI generated provided in the form of SaaS models. APIs are pulled from the cloud, whether they're from Bedrock, Vertex, Azure or the first-party from the -- it is still provided in the cloud. So most of all that demand at the end of the day in terms of compute capacity, which is what we're doing, comes from those few large frontier model developers and the products they generate to supply to consumers and enterprises globally.

Source of demand comes from those frontier model labs who are developing the products, which consumer enterprise like you and us and our companies are consuming. And what we're doing is providing that capacity to that demand source as opposed to going to a company or a bank and trying to provide them XPUs and then they're having to try to build it run create a software stack to then write applications and run it themselves. I'm sure there are a few enterprises doing now, but there are not many. It's early stage in that whole game. Right now, the -- most of the demand are coming from the frontier model players who are creating products things, as I said, like code assistance, like our engineering verticals, which are really coming from the same source of guys who are doing all those few guys doing the frontier model. It's not really coming from 100,000 companies directly trying to buy XPUs or for the method GPUs is...

One moment for our next question. And that will come from the line of Joe Moore with Morgan Stanley.

You talked about $30 billion of AI bookings in the quarter, which is a lot, I guess, relative to this quarter and next quarter shipments. Can you talk about the dynamic? Why is there so much backlog now? Or is there -- you sort of said you can react to upside with supply, just why so many bookings this quarter relative to revenue?

Well, that's -- that's a huge demand of compute -- see a lot of large is few 6 customers now being realized that lead time to get compute, you need lead time. You need to be thoughtful. And it's not just asking for wafers to get the chips on memory to ensure that HBMs are available or DRAM is available, they're also talking about, "Hey, I got to have the power, the power shell." So all this is planning ahead. And what we are seeing the bookings that come in is not for immediate delivery some hope to have. But the reality, they all accept is they need to align quite a few other things in place before they can deliver. But they are placing their orders early and they're placing their orders down, and they are placing orders in fairly huge demand, which basically gives us a lot more visibility than we normally otherwise would have in semiconductors.

Our visibility runs all the way to 2028 right now. Three months ago, I can tell you, visibility run pretty much '27. Today, it runs '28, and that's -- and that's a big part of the reason why we are creating this XPV platform as really the platform to plan, to build up this cap -- put in place such capacity for those large frontier model customers of ours who are seeing, as you guys are seeing in some of the financials, they are telling you and there are any experience you have, which is driving a huge amount of consumption of tokens from those compute capacity we are giving them. We have the benefit now of a lot of lead time. and we're planning that. And it's not because of shortage of our components is also the other elements that need to be put in place, which particularly relates to power and connection to into an infrastructure globally through America, at least, that enables inference to be distributed through to consumers and enterprises throughout the country. So it's -- we're just getting a lot of lead time.

We do have time for one final question, and that will come from the line of Joshua Buchalter with TD Cowen.

In the past, you've talked about sort of $15 billion, $20 billion per gigawatt of compute. And given the 10 that you implicit implied is what you'll be doing next year, it implies a much larger number than $100 billion. You've also mentioned that the value per gigawatt does vary per project. So I guess, how should we think about the evolution of your revenue per gigawatt over time as I would expect, on one hand, the -- pricing to increase on programs you're already shipping, but also there are other projects that are entering the model.

Our revenue -- our content by gigawatt will increase, but it's simply our content from the fact that our comps will XPU will go up in price very dramatically, particularly when you not only put SRAMs into it, you start putting a long -- you start putting embedding CPU core into the XPUs and making those ships basically multi-die with lots of HBM. So the trajectory of content increase will go up. It just doesn't go up every month, every 6 months or every quarter, but it will follow one generation to the next, that the content will grow per gigawatt.

I would now like to turn the call over to Ji Yoo, Head of Investor Relations for closing remarks.

Thank you, operator. Broadcom currently plans to report its earnings for the third quarter of fiscal year 2026 after close of market on Wednesday, September 2, 2026. A public webcast of Broadcom's earnings conference call will follow at 2:00 p.m. Pacific Time. That will conclude our earnings call today. Thank you all for joining.

Sheri, you may end the call.

This concludes today's program. Thank you all for participating. You may now disconnect.

Welcome to Broadcom Inc.'s First Quarter Fiscal Year 2026 Financial Results Conference Call. At this time, for opening remarks and introductions, I would like to turn the call over to Ji Yoo, Head of Investor Relations of Broadcom Inc.

Thank you, operator, and good afternoon, everyone. Joining me on today's call are Hock Tan, President and CEO; Kirsten Spears, Chief Financial Officer; Charlie Kwwas, President, Semiconductor Solutions Group; and Ram Velaga, President, Infrastructure Software Group.

Broadcom distributed a press release and financial tables after the market closed, describing our financial performance for the first quarter fiscal year 2026. If you did not receive a copy, you may obtain the information from the Investors section of Broadcom's website at broadcom.com. This conference call is being webcast live, and an audio replay of the call can be accessed for 1 year through the Investors section of Broadcom's website.

During the prepared comments, Hock and Kirsten will be providing details of our first quarter fiscal year 2026 results, guidance for our second quarter of fiscal year 2026 as well as commentary regarding the business environment. We'll take questions after the end of our prepared comments. Please refer to our press release today and our recent filings with the SEC for information on the specific risk factors that could cause our actual results to differ materially from the forward-looking statements made on this call.

In addition to U.S. GAAP reporting, Broadcom reports certain financial measures on a non-GAAP basis. A reconciliation between GAAP and non-GAAP measures is included in the table attached to today's press release. Comments made during today's call will primarily refer to our non-GAAP financial results. I will now turn the call over to Hock.

Thank you, Ji. And thank you, everyone, for joining us today. In our fiscal Q1 2026, total revenue reached a record $19.3 billion and that's up 29% year-on-year and exceeding our guidance on the back of better-than-expected growth in AI and semiconductors. This top line strength translated into exceptional profitability with Q1 consolidated adjusted EBITDA, hitting a record $13.1 billion which is 68% of revenue.

These figures demonstrate that our scale continues to drive significant operating leverage. Now we expect this momentum to accelerate as our custom AI XPUs hit the next phase of deployment among our 5 customers. So looking ahead to next quarter, Q2 '26, we're guiding for consolidated revenue of approximately $22 billion, which represents 47% year-on-year growth. Let me now give you more color on our semiconductor business.

In Q1, revenue was a record $12.5 billion as year-on-year growth accelerated to 52%. This robust growth was driven by AI and semiconductor revenue, which grew 106% year-on-year to $8.4 billion, way above our outlook. In Q2, this momentum accelerates, and we expect semiconductor revenue to be $14.8 billion, up 76% year-on-year. Driving this is AI revenue growth, which will accelerate very sharply to 140% year-on-year to $10.7 billion.

Now our custom accelerator business grew 140% year-on-year in Q1. This momentum continues in Q2. The ramp of custom AI accelerators across all our 5 customers is progressing very well. For Google, we continue our trajectory of growth in '26 with strong demand for the seventh-generation Ironwood TPU. In 2027 and beyond, we expect to see even stronger demand from next generations of TPU.

for Anthropic, we are off to a very good start in 2026 for 1 gigawatt of TPU compute. And for '27, this demand is expected to surge in excess of 3 gigawatts of compute. Our XPU franchise, I should add, extends beyond GPUs. Now contrary, the recent analyst reports Meta's custom accelerator MTIA roadmap is alive and well. We're shipping now. And in fact, for the next generation XPUs, we will scale to multiple [ gigawatts ] in '27 and beyond.

Rounding off for customers 4 and 5, we see strong shipments this year and which we expect to more than double in 2027. We also now have a sixth customer. We expect OpenAI deploying in volume, their first-generation XPU in 2027 and over 1 gigawatt of compute capacity. Let me take a second to emphasize our collaboration with these 6 customers to develop AI XPUs is deep, strategic and multiyear. We bring to the partnerships, each of them unmatched technology in service, silicon design, process and acknowledging, advanced packaging and networking to enable each of these customers to achieve optimal performance for their differentiated LLM workloads.

We have the track record to deliver these XPUs and high volumes and an accelerated time to market with very high yields. And beyond technology, we provide multiyear supply agreements as our customers scale up deployment of their compute infrastructure. Our ability to ensure supply in these times of constrained capacity in leading-edge wafers, in high-bandwidth memory and substrates ensures the durability of our partnerships. And we have fully secured capacity of these components for '26 through '28.

Consistent with the strong outlook for XPUs, demand for AI networking is accelerating. AI networking revenue grew 60% year-on-year and represented 1/3 of total AI revenue. In Q2, we project AI networking to accelerate a lot more and grow to 40% of total AI revenue. We are clearly gaining share in networking. Let me explain in scale out, our first-to-market Tomahawk 6 switch at 100 terabit per second as well as our [ 200G SERDES ] are capturing demand from hyperscalers, whether they use XPUs or GPUs. This lead will extend in '27 with our next-generation Tomahawk featuring double the performance.

Meanwhile, in scale up as cluster sizes and our customers expand, we are uniquely positioned to enable these customers to stay on direct attached copper through our 200G SERDES as we next step up to 400G SERDES in 2028, our XPU customers will likely continue to stay on direct attach copper. And this is a huge advantage as the alternative of going to optical. It's more expensive and requires significantly more power, reflecting the foregoing factors our visibility in 2027 has dramatically improved.

Today, in fact, we have line of sight to achieve AI revenue from chips, just chips in excess of $100 billion in 2027. We have also secured the supply chain required to achieve this. Now turning to non-AI semiconductors. Q1 revenue of $4.1 billion was flat year-on-year, in line with guidance. Enterprise networking, broadband server storage revenues were up year-on-year, offset by a seasonal decline in wireless.

In Q2, we forecast non-AI semiconductor revenue to be approximately $4.1 billion, up 4% from a year ago. Let me now talk about our infrastructure software segment. Q1 infrastructure software revenue of $6.8 billion was in line with our guidance, it was up 1% year-on-year. For Q2, we forecast infrastructure software revenue to be approximately $7.2 billion, up 9% year-on-year.

VMware revenue grew 13% year-on-year. Bookings continue to be strong and total contract value booked in Q1 exceeded $9.2 billion sustaining an ARR, which is annual recurring revenue growth of 19% year-on-year. Let me reinforce that this growth in our infrastructure software business reflects our focus and investments in foundational infrastructure. And our infrastructure software is not disrupted by AI. In fact, VMware Cloud Foundation, VCF is the essential software layer in data centers, integrating CPUs, GPUs, storage and networking into a common high-performance private cloud environment as the permanent abstraction layer between AI software and physical chips, silicon, VCF cannot be disintermediated or replace.

It allows enterprises in fact, to scale complex generative AI workloads effectively with agility that hardware alone cannot provide. We are confident that the growth in generative and agentic AI will create the need for more VMware not less. So in summary, let me put it all together for Q2 2026, we expect consolidated revenue growth to accelerate to 47% year-on-year and reach approximately $22 billion, and we expect adjusted EBITDA to be approximately 68% of revenue. So with that, let me turn the call over to Kirsten.

Thank you, Hock. Let me now provide additional detail on our Q1 financial performance. Consolidated revenue was a record $19.3 billion for the quarter, up 29% from a year ago. Gross margin was 77% of revenue in the quarter. Consolidated operating expenses were $2 billion, of which $1.5 billion was R&D. Q1 operating income was a record $12.8 billion, up 31% from a year ago.

Operating margin increased 50 basis points year-over-year to 66.4% on favorable operating leverage. Adjusted EBITDA of $13.1 billion or 68% of revenue was above our guidance of 67%. Now let's go into detail for our 2 segments. Starting with Semiconductors. Revenue for our Semiconductor Solutions segment was a record $12.5 billion, with growth accelerating to 52% year-on-year driven by AI. Semiconductor revenue represented 65% of total revenue in the quarter.

Gross margin for our Semiconductor Solutions segment was up 30 basis points year-on-year to approximately 68%. Operating expenses of $1.1 billion reflected increased investment in R&D for leading-edge AI semiconductors and represented 8% of revenue. Semiconductor operating margin of 60% was up 260 basis points year-on-year, reflecting strong operating leverage.

Now moving on to infrastructure software. Revenue for infrastructure software of $6.8 billion was up 1% year-on-year and represented 35% of revenue. Gross margin for infrastructure software was 93% in the quarter and operating expenses were $979 million in the quarter. Q1 software operating margin was up 190 basis points year-on-year to 78%.

Moving on to cash flow. Free cash flow in the quarter was $8 billion and represented 41% of revenue. We spent $250 million on capital expenditures. We ended the first quarter with inventory of $3 billion as we continue to secure components to support strong AI demand. Our days of inventory on hand were 68 days in Q1 compared to 58 days in Q4 in anticipation of accelerating AI semiconductor growth.

Turning to capital allocation. In Q1, we paid stockholders $3.1 billion of cash dividends based on a quarterly common stock cash dividend of $0.65 per share. During the quarter, we repurchased $7.8 billion or approximately 23 million shares of common stock. In total, in Q1, we returned $10.9 billion to shareholders through dividends and share repurchases.

In Q2, we expect the non-GAAP diluted share count to be approximately 4.94 billion shares, excluding the impact of potential share repurchases. We ended the first quarter with $14.2 billion of cash. Today, we are announcing our Board of Directors has authorized an additional $10 billion for our share repurchase program effective through the end of calendar year 2026. Now moving on to guidance. Our guidance for Q2 is for consolidated revenue of $22 billion, up 47% year-on-year.

We forecast semiconductor revenue of approximately $14.8 billion, up 76% year-on-year. Within this, we expect Q2 AI semiconductor revenue of $10.7 billion, up approximately 140% year-on-year. We expect infrastructure software revenue of approximately $7.2 billion, up 9% year-on-year. For your modeling purposes, we expect consolidated gross margin to be flat sequentially at 77%. We expect Q2 adjusted EBITDA to be approximately 68%. We expect the non-GAAP tax rate for Q2 and fiscal year 2026 to be approximately 16.5% due to the impact of the global minimum tax and the geographic mix of income compared to that of fiscal year '25. That concludes my prepared remarks. Operator, please open up the call for questions.

[Operator Instructions]

And our first question will come from the line of Blayne Curtis with Jefferies.

Just a clarification on the question. Just a clarification, Hock, the greater than $100 billion. I think you said AI chips. I just want to make sure you're clarifying the difference between the ASICs and networking and didn't know how rack revenue fits in there. And then the question, I think the biggest overhang on the group here is that you grew roughly double in the quarter AI. I think that's what kind of cloud CapEx is growing this year. I'm just kind of curious to your perspective. I think given the outlook that you have for '27, you should be a share gainer. I'm just kind of curious to your perspective in terms of the pessimism that investors kind of think of that the hyperscalers need to get a return on investment in this year or next year or if not the year after? I'm just kind of curious, your perspective, how you factor that into your outlook?

Well, what we see -- what we've seen over the last few months and continue to see even more is -- and it's really not so much talking about hyperscalers. Our customers, Blayne, is limited to those few players out there, and some of them are hyperscalers. Some of them are non-hyperscalers, but they all have one thing in common, which is to create LLMs, productize it and generate platforms, be it for enterprise consumption, in code assistance, agentic AI or be it for consumer subscription that we know about whatever it is, is that few prospects and many of whom are our customers now, who are creating this -- whether it's generative AI, agentic AI, but creating a platform. That's our customer. And with respect to each of those guys, we have seen far stronger and stronger demand for compute capacity for training, which is something they do need constantly.

But what is very, very interesting and surprising to us is very much for inference in order to productize the LLM, their latest LLMs they create and monetize it. And that inference is driving a substantial amount of compute capacity, which is great for us because this -- or these players, these 5, 6 customers of ours on their path to creating their own custom accelerators. And beyond that, their own design architecture of networking clusters of those customer accelerators.

So I think we're going to see demand picking up as we have heard announcements in the past 6 months. Now to clarify your first part, Blayne. When I say we forecast, we have a line of sight that our revenue in '27 will be significantly in excess of $100 billion, I'm focusing on the fact that these are pretty much all based on chips, whether they are XPUs, whether they are switch chips DSPs, these are silicon content we're talking about.

[Operator Instructions] And that will come from the line of Harlan Sur with JPMorgan.

Congratulations to the team on the strong results. There's been a lot of noise around CSPs and hyperscalers embarking on their own internal XPU TPU design efforts, right? We call it COT or customer-owned tooling. This is not a new dynamic with ASIC, right? I think the Broadcom team has been through the COT competitive dynamic before over the 30 years, right, that you've been a leader in the ASIC industry and very few of these COT initiatives have ever been successful.

Now on the AI, some of these COT initiatives are coming to the market now, but it looks like -- they're at these 2x less performance in your current generation solutions, 2X less complex in terms of chip design complexity, packaging complexity, IP. So maybe just a quick 2-part question. Hock, one for you is given your visibility into next year, do you see the COT [indiscernible] projects taking any meaningful TPU XPU share from Broadcom? And then maybe the second quick question for, Charlie, is given that Broadcom's TPU XPU programs from a performance complexity IP perspective are 12 to 18 months ahead of any of these COT programs, how does Broadcom team widen this gap further?

Well, that's a great question. And it fits into that -- I purposely took the time in my opening remarks to say that when any of our -- any I guess, hyperscaler or LLM developer tries to create become self-sufficient entirely in creating what you call customer-owned tooling or COT model, they face tremendous challenges.

One is technology. which is a technology as it relates to creating the silicon chips and particularly in XPUs that they need to do the computing and that is needed to optimize and run an inference on the workloads they produce their LLM. It's technology we talked about comes in from different dimensions. We need the best silicon design team around, you need cutting-edge, really cutting-edge SERDES, very advanced packaging. And most -- and just as much you need to understand how to network clusters of them together.

We've been doing this for 20 years -- more than 20 years in silicon and in this particular space today in generative AI. If you're trying to as an LLM player to do your own chip, you cannot afford to have a chip that is just good enough. You need the best chips that is around because you're competing against other LLM players. And most of all, you're also competing against NVIDIA, who is by no means letting down their [ gun ]. They are producing better and better chips with every passing generation.

So you have to, as an LLM, trying to establish our platform in the world, have to create chips that are better than if not competitive, not just in NVIDIA, but all the other platform players that you're competing against. And for that, you really need own belief, and we see that percent -- partner in silicon with the best technology, IP and execution around.

And very modestly, I would say, we are by far way out there. And we will not see competition in COT for many years to come. It will come eventually, but we're still a long way off because the race which we see continues. And 1 thing I add in there that is particularly unique to us, when you create the silicon, you really have to get it up and running in high volume in production very quickly, time to market. We are very, very experienced in doing that. Anybody can design a chip in a lab that works well. Can you produce 100,000 of those chips quickly at yields that you can -- that you can afford? And we don't see too many players in the world that can do that. Charlie?

I think you covered it very well, Hock .

One moment for our next question and that will come from the line of Ross Seymore with Deutsche Bank.

In your script, you leaned a little bit more into the networking differentiation than you have in the past. So I guess kind of a short-term and a longer-term question. The short term is, what's driving that up to 40% of the AI revenues? And the longer-term question is, is that going -- that percentage mix in that $100 billion plus, is that changing now what sort of leadership do you expect to maintain in that business, whether it's scale out or scale up? And is your leadership position there helping on your XPU side as you can optimize across both the compute and the networking side.

Well, let's address the first part of that fairly complex question first, Ross. Yes, in networking, especially with the new generation of GPUs, XPUs that are coming out there. We're running at 200 gigabit SERDES out there in terms of bandwidth. And the Tomax that we introduced over 6 months ago a closer to 9 months ago. We're the only one out there. And our customers and the hyperscalers wants to run with the best networking and with the most band we found there for their clusters.

So we are seeing huge demand for this only 100 terabit per second switch out there. So that's driving a lot of demand. And couple that we're running bandwidth on scaling out optical transceivers at 1.6 terabit. We are, again, the only player out there doing DSP at 1.6 terabits. That combination is driving, I would say, the growth of our networking components, even faster than our XPUs are growing, which is already pretty remarkable.

So that's what you're seeing. But at some point, I would think these things will settle down, though. We're not slowing down the pace because, as I said, next in '27, we'll launch next-generation Tomahawk 7, 2x the performance and will probably be by far, the first out there, and then we'll continue to sustain that momentum. And -- but at the end of the day, to answer your question, yes, I expect as the composition of our total AI revenue in any quarter that we'll be ranging between probably 33% to 40% AI networking components.

One moment for our next question and that will come from the line of CJ Muse with Cantor Fitzgerald.

I'm curious, how are you thinking about the move to disaggregate [ prefill decode ] from the GPU ecosystem? And the impact of custom silicon demand. Are you seeing any potential changes in sort of the relative mix between GPUs and customer silicon?

I'm not sure I fully understand your question, CJ. Could you clarify what you mean disaggregate?

Sure. Pushing off workloads to CPX for prefill and working off [ Grok ] for decode and having that disaggregated kind of world. And does that put any pressure in terms of the demand for custom versus going with a full GPU stack.

Okay. I get what you mean. That's what disaggregation [indiscernible] what you -- in a way, what you're really saying is what, how is the architecture of accelerated GPU or XPU evolving as workload starts to evolve. And that's what we are seeing very much in particular. The one size fits all of a general purpose GPU gets you only that far. It can still keep going on because you can still run different workloads like you run, make sure of experts even though you have -- you want to run mixture of experts with sparse costs to be very effective, you hear the term, but in a GPU, you are designed for dense matrix multiplication.

So you do it with software kernels, but it's not as effective as you [ hot coated ] in silicon and make those purposely designed to be much more performing for mixture of expert workloads, say. The same applies for inference. And what that drives down to is you start to see designs of XPUs become much more customized for particular workloads of particular LLM customers of ours, and the design starts to [ depart ] from what is the traditional standard GPU design, which is why as we always indicated before. XPUs will effect will eventually be more the choice simply because it will allow flexibility in making designs that work with particular workloads,

1 for training even and one for inference. And as you say, 1 perhaps would be better at prefilling and want to be better at post-training or reinforced learning or test time scaling. You can tweak your TPUs towards the XPU [indiscernible] to a particular kind of workload LLM that you want. And we're seeing that roadmap in all our 5 customers.

One moment for our next question and that will come from the line of Timothy Arcuri with UBS.

I had just a question on sort of the puts and takes on the gross margin as you begin to ship these racks. I mean, obviously, it's going to pull the blended margin down. But I'm wondering if there's any guardrails you can give us on this. It seems like the racks are maybe 45%, 50% gross margin. So I guess, should we think about that pulling gross margin down like 500 basis points, roughly, as these racks begin to ship? And I guess part of that, HOck, is there some like floor to the gross margin below which you wouldn't be willing to do more racks?

Hate to tell you that you must be a bit hallucinated. Our gross margin is solidly at the number Kirsten report. We will not be affected by the gross one and buy more and more AI product going up. We have gotten our yields. We've gotten our cost to a point where -- the model we have in AI will be fairly consistent the models we have in the rest of the semiconductor business. Kirsten? .

I would agree with that. I think on further study, relative to even comments that I did make last quarter, the impact relative to our overall mix is actually not going to be substantial at all. So I wouldn't worry about it. .

[Operator Instructions] And that will come from the line of Stacy Rasgon with Bernstein.

I don't know if this is for Hock or Kirsten, but I wanted to dig in a little more to this substantially more than $100 billion next year. I'm trying to just count up the gigawatts. I counted, I don't know, 8 or 9, you have 3 from Anthropic, one from open AI, so that's 4. You said Meta was multiple, so east, that gets you to 6. Google, I figure should be bigger than Meta.so like at least 3, that's 9 and then you got a few others.

I just thought that your content per gigawatt was sort of, call it, a $20 billion per gigawatt range. I guess what I'm asking, is my math around the gigawatts you plan to ship in 27 correct? And how do I think about your content per gigawatt as that ships -- maybe we'll be "substantially" more than $100 billion.

Stacy, you have a very interesting perspective, and I've got to [ remind ] you for that. But you're right. You can look at it a gigawatts, which is the right way to look at it instead of dollars because that's how we sell our chips. So you have to realize we -- depending on our LLM customer, our 6 customers, sorry, not 56. The dollars per gigawatt is varies, sometimes quite dramatically. -- it does vary. But you're right. It's so far from the dollars you are talking about. And if you look at it by gigawatt in '27, we are seeing getting close to 10 gigawatts. .

Our next question that will come from the line of Ben Reitzes with Melius Research.

Hock, great to be speaking with you. I wanted to ask you about your commentary about supply visibility on those 4 major components through 2028. A, how did you do it? This is probably you're the first one to kind of go out through the '28 time frame? And secondly, after this astounding growth in 2027 for your AI business, do you have enough visibility to grow quite a bit in 2028 based on the supply that you see in that kind of commentary?

The best answer is, yes, you're right. We anticipate this sharp accelerated growth. Nobody could anticipate that rate of growth is showing, but we kind of anticipate a large part of it, I guess, longer than 6 months. We were early in being able to lock up T glass, the [ infamous ] T glass you all heard about, we were very early. We've locked up substrates. We have worked on our good partners on the rest of the stuff we talked about. And so the answer to your question is it's somewhat anticipation early and the fact that we have very good partners out there in these key components. What else can I say except that, yes. Charlie, do you want to add anything? .

Yes. Just maybe a couple of quick ones. I think you covered that piece really well. I think, Ben, the other piece that's really important as Hock said, we build custom silicon for 6 customers. We have very deep strategic multiyear engagement with them. They share with us because of this custom capability exactly what they anticipate at least over the next 2 to 3 years, sometimes 4 years.

And so because of that, that's exactly why we went and secured all the elements, Hock, talked about. And when we secure this, it requires investments with these partners, sometimes developing not just more capacity, but the right technology and capacity for that. So we have to go secure it for multiple years. And we're probably -- you're right. We're probably the first 1 to secure that up to '28 or beyond.

And can you grow in '28 with what you see in supply? Sorry to sneak that in.

Yes.

Our next question will come from the line of Vivek Arya with Bank of America Securities.

Hock, I just wanted to first clarify the Anthropic project you're doing the $20 billion or so for a gigawatt this year, how much of that is chips and how much of that is kind of rack? I just wanted to understand when you say $100 billion in chips, is there a distinction between chips versus your rack-scale projects because just that project is supposed to triple next year.

And then my question is your AI business is transitioning from kind of 1 large customer that was where you had kind of exclusive partnership to now multiple customers who are using multiple suppliers. So how do you get the visibility and the confidence about how your share will progress at these multiple customers because it's a very kind of fragmented engagement that they have across a whole range of cloud service providers and so. So what are you doing to ensure that you have solid visibility and the right market share at this fragmented set of customers who are using multiple suppliers. .

Vivek, you have to understand 1 thing about, first, as Charlie correctly put down, very nicely. We only have very few customers to be precise 6 for the volume we are driving the revenue we're driving we only have just 6, prior to the even less refilling. And number two, also I have to understand with the dollars each of them spend and the criticality of the nature of what they're embarking on.

And that's why I threw out this method has MTI, that's the A customer accelerator program. To them, as every one of my customers in this space, it's a strategic play. It's not optionality to them long term, short term, medium term is strategic, extremely strategic. They don't stop and they are very clear each of them on where they want to position these custom silicon within the trajectory of the LLM development and the trajectory of how they develop, inference for productizing those elements. That part, we have very clear visibility anything else on GPU, using new cloud -- using cloud business, these are all transactional and optionality. So you have to -- you point out very correctly, it seems very confusing. Trust me, not for us, on those customers we have. They're very strategic. They're very targeted, and they know exactly what they're building up and how much capacity they want to build up each year. And the only thing they think about is can you do it faster. Otherwise, it's very strategic and targeted on a projected road map. Anything else you see in the mix is pure I call it, opportunistic for these guys, the optionality. So it's very clear.

And on the clarification, Hock, Anthropic racks versus chips?

I'd rather not answer then, but we're okay. As Kirsten said, we're good on our dollars and margin.

Our next question that will come from the line of Tom O'Malley with Barclays.

I have 1 for Hock and one for Charlie. So Hock, I know you're very specific in particular about what you put in the [indiscernible] and you noted that customers are saying a direct attached copper through 400 gig SERDES. Is there any reason you're pointing that out in particular, especially as a leading pioneer in CPO. And then on Charlie side, as you're adding more customers here, I'd imagine customers that design ASIC with you are going to use scale-up Ethernet. Maybe talk about scale-up protocols and how you see Ethernet developing here as well...

Okay. No, unless -- I'm just highlighting the fact that with -- on networking, our technology is really very, very uniquely positioning us to help our customers and more than our customers, even customers using John purpose GPUs, not just XPUs, which is that if you are running trying to create LLM and running, creating your own AI data centers and we're designing architecting it. You truly want larger and larger domains or clusters phone and you really want to connect XPUs to XPUs directly where you can.

And the best way to do that is used directly attached copper. That's the lowest latency, lowest power and lowest cost. So you want to keep doing that, especially in scale up as long as possible. In scaling out we pass that we use optical. That's fine. While I'm talking about scaling up in a rack in the cluster domain, you really want to use direct attach corporate as long as you can. And we are still based on our technology that Broadcom has with -- especially on connecting XPU or even GPU to GPU. We can do it with copper, and we can push the envelope from 100 gig to 200G to even 400G, we have SERDES now running 400G that can drive distance on a rank to run copper. All I'm trying to say is you don't need to go run into some bright shiny objects called CPO, even as we are the lead in CPOs. CPOs will come in its time. not this year, maybe not next year, but in its time. Charlie?

Yes. No, well said Hock. And on the question of Ethernet with the debut of the cloud, Ethernet became the de facto standard in every cloud for the last 2 decades. If you look at the debut of the back-end networks, as Jack articulated, there was 2 years ago, a big fight about what protocol should be used to achieve the latency, the scale necessary on scale-out.

And the industry at the time, 24 months ago was not clear. we were clear. We were very clear, actually, about what the answer should be. And again, because of the deep engagements with our partners, they made it very clear to all of us and the industry, GPU or XPU that Ethernet is the scale out of choice. Checkmark. Today, everyone is talking about scaling out with Ethernet. Now when it comes to scale up, yes, exactly like what happened 3, 4 years ago, on scale up now, what's the right answer for this. And what we're hearing consistently and what we're seeing is the right answer is Ethernet. And as you know, last year, we've announced with multiple hyperscalers and many of our peers in the semiconductor industry that Ethernet scale up is the right choice. That's what we believe will happen. Time will tell, but a lot of the XPU designs were doing. We're being asked to scale up through Ethernet, and we're happy to enable that.

And our next question that will come from the line of Jim Schneider with Goldman Sachs.

Hock, it was helpful to hear you discuss the progress of your other full custom XP engagements outside of TPUs. As we look into next year, is it fair to assume that those are mostly targeting inference applications or not? And then could you maybe qualitatively speak to either the performance or cost advantages relative to GPUs that is giving those customers the ability to forecast in such large scale?

Thanks. It's -- most of our customers begin with inference, simply because that tends to be -- that tends to be the easiest path to stand on. Not necessarily from anything else than the fact that when you do inference, it's much -- it's less compute, but also then the question is, do you need this general purpose massive, dense matrix multiplication GPUs, when you can do it more efficiently, effectively with customs inference silicon XPUs that do the job better or just as well, much cheaper cost, lower power. And that's what we find this customer starting with.

But they are now in training and many of the XPUs are used both in training as well as inference. And by the way, they are interchangeable. Just a GPU can be used not just for training, which they are, perhaps, more perfectly suited to, but they can be used for inference. What we're seeing is our XPUs are used for both. And we are seeing that going on. But we're also seeing very rapidly more for those customers who are much more mature in the progression I talked about in their journey towards complete XPU that they will start to develop 2 chips each year, simultaneously, one for training, 1 for inference to be specialized Why? Because what we're seeing very clearly for this players, LLM players, you do the training to get -- to achieve a higher level of intelligence, smart for your LLM.

So great, you get yourself a great LLM state-of-the-art or more. Now you've got to productize it, which means inference. Well, you can then decide at that time, you've got your model going as the best because if you decide then to do your inference productization, it will take you a year at least to productize at which time somebody else is going to create an LLM better than yours. So you -- that's a leap of faith here that when you do training to create the next level of super intelligence in your LLM, you have to be investing simultaneously in inference, both in terms of the chip and the capacity. So our visibility is really coming out better and better as we find those 6 customers get more mature in the progression towards better and better LLMS. So yes, that is the trend we are seeing. It's not happening to all our 6 customers yet, but we're seeing a majority of them headed in that way right now

Thank you. One moment for our next question. And that will come from the line of Joshua Buchalter with TD Cowens.

Congrats on the results. I appreciate all the details on the expectations for deployment at specific customers. I was hoping you could just maybe reflect on how visibility has changed over the last 1 to 2 quarters that gave you the confidence to give us more details? And then on a specific one, you mentioned greater than 1 gigawatt for OpenAI in 2027, with that deal being for 10 gigawatts through 2029, that implied a pretty sharp inflection, I guess, in 2028. Is that the right way to think about it? And was that sort of always the plan?

Yes. Well, yes, this -- as you've all seen, and you all know in this generative AI race that we are in now, and I shouldn't use the word race, let's call it, progression among the few players we see here, I mean, it's a competition. Each is trying to create an LLM better than the other and more tailored for specific purpose, be it enterprise, be it consumer, be it search. Each one is trying to create it more and more. And all of that requires not just training, which is important to keep improving your LLM models. But inference for productization and monetization of your LLMs.

And we are going -- and probably call it the fact that we've been engaged with some of them now for more than a couple of years. We're getting better and better visibility as they have more and more confidence that the XPUs they are working on with us is achieving what they are getting it as they get the sense that the XPUs they are working on with the software, with the algorithms they needed, they are having more confidence that this XPU silicon is what they need.

And it gets better and better. And It's get better, we get more visibility as Charlie puts up perfectly because at the end of the day, we only have 6 guys to work on. And these 6 guys are all, as I said, look at XPUs and AI in a very strategic manner. They don't think 1 generation at a time. They think multiple generation multiple years.

And in spite of all the noise out there on what's available they think very long term on how they deploy the experience they develop with us, how they deploy in achieving better and better LLMs that they want to create and more than that, how they deploy and monetizing. So it's -- we are part of their strategic roadmap. We are not in just optionality of, oh, shall I use a GPU, shall I use it in the cloud because I need to train for 6 months. No. this is more than that. The investment these guys are making a long term, and it's great to be part of that long-term roadmap as opposed to a transactional roadmap. And the noise, as I answered an earlier question from is there's a lot of noise that makes up short-term transactions with one is long-term strategic positioning of our business and our product. And to sum it all, I think our business in XPUs is a strategic sustainable play for all the 6 customers we have today.

Thank you. That is all the time we have for Q&A today. I would now like to turn the call back over to Ji Yoo for any closing remarks.

Thank you, Sheri. Broadcom currently plans to report its earnings for the second quarter of fiscal year 2026 after the close of market on Wednesday, June 3, 2026. A public webcast that Broadcom's earnings conference call will follow at 2:00 p.m. Pacific. That will conclude our earnings call today. Thank you all for joining. Sheri, you may end the call.

This concludes today's program. Thank you all for participating. You may now disconnect.

Welcome to Broadcom Inc.'s Fourth Quarter and Fiscal Year 2025 Financial Results Conference Call. At this time, for opening remarks and introductions, I would like to turn the call over to Ji Yoo, Head of Investor Relations of Broadcom Inc.

Thank you, Sheri, and good afternoon, everyone. Joining me on today's call are Hock Tan, President and CEO; Kirsten Spears, Chief Financial Officer; and Charlie Kawwas, President Semiconductor Solutions Group.

Broadcom distributed a press release and financial tables after the market close, describing our financial performance for the fourth quarter and fiscal year 2025. If you did not receive a copy, you may obtain the information from the Investors section of Broadcom's website at broadcom.com.

This conference call is being webcast live, and an audio replay of the call can be accessed for 1 year through the Investors section of Broadcom's website.

During the prepared remarks, Hock and Kirsten will be providing details of our fourth quarter and fiscal year 2025 results, guidance for our first quarter of fiscal year 2026 as well as commentary regarding the business environment. We'll take questions after the end of our prepared comments.

Please refer to our press release today and our recent filings with the SEC for information on the specific risk factors that could cause our actual results to differ materially from the forward-looking statements made on this call.

In addition to U.S. GAAP reporting, Broadcom reports certain financial measures on a non-GAAP basis. A reconciliation between GAAP and non-GAAP measures is included in the tables attached to today's press release. Comments made during today's call will primarily refer to our non-GAAP financial results. I'll now turn the call over to Hock.

Thank you, Ji, and thank you, everyone, for joining us today. Well, we just ended our Q4 fiscal '25. And before I get into details of that quarter, let me recap the year.

In our fiscal 2025, consolidated revenue grew 24% year-over-year to a record $64 billion, and it's driven by AI semiconductors and VMware. AI revenue grew 65% year-over-year to $20 billion, driving the semiconductor revenue for this company to a record $37 billion for the year. In our Infrastructure Software business, strong adoption of VMware Cloud Foundation or VCF, as we call it, drove revenue growth of 26% year-on-year to $27 billion.

In summary, 2025 was another strong year for Broadcom. And we see the spending momentum by our customers for -- in AI, continuing to accelerate in 2026. Now let's move on to the results of our fourth quarter 2025.

Total revenue was a record $18 billion, up 28% year-on-year and above our guidance on better-than-expected growth in AI semiconductors as well as Infrastructure Software. Q4 consolidated adjusted EBITDA was a record [ $12.2 billion, ] up 34% year-on-year. So let me give you more color on our two segments.

In Semiconductors, revenue was $11.1 billion as year-on-year growth accelerated to 35%. And this robust growth was driven by the AI semiconductor revenue of $6.5 billion, which was up 74% year-on-year. And this represents a growth trajectory exceeding 10x over the 11 quarters we have reported this line of business. Our custom accelerated business more than double year-over-year, as we see our customers increase adoption of XPUs, as we call those custom accelerators in training their LLM and monetizing their platforms through influencing APIs and applications.

These XPUs, I may add, are not only been used to train and influence internal workloads by our customers, the same XPUs in some situations have been extended externally to other LLM peers, best exemplified at Google, where the TPUs use in creating Gemini, have also been used for AI cloud computing by Apple, Coherent and SSI as an example. And the scale at which we see this happening could be significant. And as you are aware, last quarter, Q3 '25, we received a $10 billion order to sell the latest DPU [indiscernible] to [indiscernible]. And this was our fourth customer that we mentioned. And in this quarter Q4, we received an additional $11 billion order from the same customer for delivery in late 2026.

But that does not mean our other two customers are using GPUs. In fact, they prefer to control their own destiny by continuing to drive their multiyear journey to create their own custom AI accelerators or XPU racks, as we call them. And I'm pleased today to report that during this quarter, we acquired a fifth XPU customer through a $1 billion order placed for delivery in late 2026.

Now moving on to AI networking. Demand here has even been stronger as we see customers build out their data center infrastructure ahead of deploying AI accelerators. Our current order backlog for AI switches exceeds $10 billion as our latest 102-terabit per second Tomahawk 6 switch, the first and only one of its capability out there. continues to book at record rates. And this is just a subset of what we have.

We have also secured record orders on DSPs, optical components like lasers and PCI Express switches to be deployed -- all to be deployed in AI data centers. And all these components combined with XPUs, bring our total order on hand in excess of $73 billion today, which is almost half Broadcom's consolidated backlog of $162 billion. We expect this $73 billion in AI backlog to be delivered over the next 18 months. And in Q1 fiscal '26, we expect our AI revenue to double year-on-year to $8.2 billion.

Turning to non-AI semiconductors. Q4 revenue of $4.6 billion was up 2% year-on-year and up 16% sequentially based on favorable wireless seasonality. Year-on-year, broadband showed solid recovery, wireless was flat and all the other end markets were down as enterprise spending continued to show limited signs of recovery. And accordingly, in Q1 we forecast non-semiconductor revenue to be approximately $4.1 billion, flat from a year ago, down sequentially due to wireless seasonality.

Let me now talk about our Infrastructure Software segment. Q4 Infrastructure software revenue of $6.9 billion was up 19% year-on-year, above -- and above our outlook of $6.7 billion. Bookings continue to be strong as total contract value booked in Q4 exceeded $10.4 billion versus $8.2 billion a year ago. We ended the year with $73 billion of Infrastructure Software backlog, up from $49 billion a year ago. We expect renewals to be seasonal in Q1 and forecast Infrastructure Software revenue to be approximately $6.8 billion. We still expect, however, that for fiscal '26 Infrastructure Software revenue to grow low double-digit percentage.

So here's what we see in 2026, directionally, we expect AI revenue to continue to accelerate and drive most of our growth. And non-AI semiconductor revenue to be stable. Infrastructure software revenue will continue to be driven by VMware growth at low double digits. And for Q1 '26, we expect consolidated revenue of approximately $19.1 billion, up 28% year-on-year. And we expect adjusted EBITDA to be approximately 67% of revenue. And with that, let me turn the call over to Kirsten.

Thank you, Hock. Let me now provide additional detail on our Q4 financial performance. Consolidated revenue was a record $18 billion for the quarter, up 28% from a year ago. Gross margin was 77.9% of revenue in the quarter, better than we originally guided on higher software revenues and product mix within semiconductors.

Consolidated operating expenses were $2.1 billion, of which $1.5 billion was research and development. Q4 operating income was a record $11.9 billion, up 35% from a year ago. Now on a sequential basis, even as gross margin was down 50 basis points on semiconductor product mix, operating margin increased 70 basis points sequentially to 66.2% on favorable operating leverage.

Adjusted EBITDA of $12.2 billion or 68% of revenue was above our guidance of 67%. This figure excludes $148 million of depreciation. Now a review of the P&L for our 2 segments, starting with Semiconductors.

Revenue for our Semiconductor Solutions segment was a record $11.1 billion with growth accelerating to 35% year-on-year, driven by AI. Semiconductor revenue represented 61% of total revenue in the quarter. Gross margin for our Semiconductor Solutions segment was approximately 68%. Operating expenses increased 16% year-on-year to $1.1 billion on increased investment in R&D for leading-edge AI semiconductors. Semiconductor operating margin of 59% was up 250 basis points year-on-year.

Now moving to Infrastructure Software. Revenue for Infrastructure Software of $6.9 billion was up 19% year-on-year and represented 39% of total revenue. Gross margin for Infrastructure Software was 93% in the quarter compared to 91% a year ago. Operating expenses were $1.1 billion in the quarter, resulting in Infrastructure Software operating margin of 78%. This compares to operating margin of 72% a year ago, reflecting the completion of the integration of VMware.

Moving on to cash flow. Free cash flow in the quarter was $7.5 billion and represented 41% of revenue. We spent $237 million on capital expenditures. Days sales outstanding were 36 days in the fourth quarter compared to 29 days a year ago. We ended the fourth quarter with inventory of $2.3 billion, up 4% sequentially. Our days of inventory on hand were 58 days in Q4 compared to 66 days in Q3 as we continue to remain disciplined on how we manage inventory across the ecosystem. We ended the fourth quarter $16.2 billion of cash, up $5.5 billion sequentially on strong cash flow generation. The weighted average coupon rate in years to maturity of our gross principal fixed rate debt of $67.1 billion is 4% and 7.2 years, respectively.

Turning to capital allocation. In Q4, we paid stockholders $2.8 billion of cash dividends based on a quarterly common stock cash dividend to $0.59 per share. In Q1, we expect the non-GAAP diluted share count to be approximately 4.97 billion shares, excluding the potential impact of any share repurchases.

Now let me recap our financial performance for fiscal year 2025. Our revenue hit a record $63.9 billion with organic growth accelerating to 24% year-on-year. Semiconductor revenue was $36.9 billion, up 22% year-over-year. Infrastructure Software revenue was $27 billion, up 26% year-on-year. Fiscal 2025 adjusted EBITDA was $43 billion and represented 67% of revenue. Free cash flow grew 39% year-on-year to $26.9 billion.

For fiscal 2025, we returned $17.5 billion of cash to shareholders in the form of $11.1 billion of dividends and $6.4 billion in share repurchases and elimination. Aligned with our ability to generate increased cash flows in the preceding year, we are announcing an increase in our quarterly common stock cash dividend in Q1 fiscal 2026 to $0.65 per share, an increase of 10% from the prior quarter. We intend to maintain this target quarterly dividend throughout fiscal '26, subject to quarterly board approval. This implies our fiscal 2026 annual common stock dividend to be a record $2.60 per share, an increase of 10% year-on-year.

I would like to highlight that this represents the 15th consecutive in annual dividends since we initiated dividends in fiscal 2011. The Board also approved an extension of our share repurchase program, of which $7.5 billion remains through the end of calendar year 2026.

Now moving to guidance. Our guidance for Q1 is for consolidated revenue of $19.1 billion, up 28% year-on-year. We forecast Semiconductor revenue of approximately $12.3 billion, up 50% year-on-year. Within this, we expect Q1 AI semiconductor revenue of $8.2 billion, up approximately 100% year-on-year. We expect Infrastructure Software revenue of approximately $6.8 billion, up 2% year-on-year.

For your modeling purposes, we expect Q1 consolidated gross margin to be down approximately 100 basis points sequentially, primarily reflecting a higher mix of AI revenue. As a reminder, consolidated gross margins through the year will be impacted by the revenue mix of Infrastructure Software and Semiconductors and also product mix within Semiconductors.

We expect Q1 adjusted EBITDA to be approximately 67%. We expect the non-GAAP tax rate for Q1 and fiscal year 2026 to increase from 14% to approximately 16.5% due to the impact of the global minimum tax and shift in geographic mix of income compared to that of fiscal year 2025.

That concludes my prepared remarks. Operator, please open up the call for questions.

[Operator Instructions] Our first question will come from the line of Vivek [ Arya ] with Bank of America.

Just wanted to clarify, Hock, you said $73 billion over 18 months for AI, that's roughly $50-ish billion plus for fiscal '26 for AI. I just wanted to get -- make sure I got that right.

And then the main question, Hock, is that there is sort of this emerging debate about customer-owned tooling, your ASIC customers potentially wanting to do more things on their own. How do you see your XPU content and share at your largest customer evolve over the next 1 or 2 years?

Well, to answer your first question, what we said is correct that as of now, we have $73 billion of backlog in place secured of XPUs, switches, DSPs, lasers for AI data centers that we anticipate shipping over the next 18 months. And obviously, this is as of now, I mean, we fully expect more bookings to come in over that period of time. And so don't take that $73 billion as that's the revenue that we ship over the next 18 months. We're just saying we have that now and in that bookings has been accelerating.

And frankly, we see that bookings not just in XPUs, but in switches, DSPs, all the other components that go into AI data center. We have never seen bookings of the nature that what we have seen over the past 3 months, particularly with respect to Tomahawk 6 switches. This is one of the fastest-growing products in terms of deployment that we've ever seen of any switch products that we put out there. It is pretty interesting and partly because it's the only one of its kind out there at this point at 102 terabits per second. And that's that exact product needed to expand the clusters of the latest GPU and XPUs out there.

So that's great. But as far as what is the future XPU is your broader question, my answer to you is don't follow what you hear out there as gospel. It's a trajectory. It's a multiyear journey. And many of the players and not too many players doing LLM wants to do their own custom AI accelerator for very good reasons. You can put in hard way if you use a general purpose GPU, you can only do in software and -- kernels and software. You can achieve performance-wise so much better in the custom purposed design, hardware-driven XPU.

And we see that in the TPU and we see that in all the accelerators we are doing for our other customers, much, much better in areas of [indiscernible] call, training, inference, reasoning, all that stuff.

Now what that means that over time, they all want to go do it themselves, not necessarily. And in fact, because the technology in silicon keeps updating keeps evolving. And if you are an LLM player, where do you put your resources in order to compete in this space, especially when you have to compete at the end of the day against merchant GPU who are not slowing down in the rate of evolution. So I see that as this concept of customer tooling is an overblown hypothesis, which frankly, I don't think will happen.

A moment for our next question. And that will come from the line of Ross Seymore with Deutsche Bank.

I want to go to something you touched on earlier about the TPUs going a little bit more to like a merchant go-to market to other customers. Do you believe that's a substitution effect for customers who otherwise would have done ASICs with you? Or do you think it's actually broadening the market? And so what are kind of the financial implications of that from your perspective?

So that's a very good question, Ross. And what we see right now is the most obvious move it does is it goes -- the people who use TPUs, the alternative is GPUs, merchant basis as the most common thing that happens. Because to do that substitution for another custom, it's different. To make an investment in custom accelerator is a multiyear journey. It's a strategic directional thing. It's not necessary a very transactional or short-term move.

Moving from GPU to TPU is a transactional move, going into AI accelerator of your own is a long-term strategic move and nothing would deter you from them to continue to make that investment towards that end goal of successfully creating and deploying your own custom AI [indiscernible]. So that's the motion we see.

And that will come from the line of Harlan Sur with JPMorgan.

Congratulations on the strong results, guidance and execution. Hock, again, I just want to -- I just want to sort of verify this, right? So you talked about total AI backlog of $73 billion over the next 6 quarters, right? This is just a snapshot of your order book like right now. But given your lead times, I think customers can and still will place orders for AI in quarters 4, 5 and 6. So as time moves forward, that top [indiscernible] number for more shipments in the second half will probably still go up, right? Is that the correct interpretation?

And then given the strong and growing backlog, right, the question is, does the team have 3-nanometer, 2-nanometer wafer supply, colos, substrate, HBM supply commitments to support all of the demand in your order book. And I know one of the areas where you are trying to mitigate this as in advanced packaging, right? You're bringing up your Singapore facility. Can you guys just remind us what part of the advanced packaging process the team is focusing on with the Singapore facility?

Well, to answer your first simpler question, Harlan, you're right. You can say that $73 billion is the backlog we have today to ship over the next 6 quarters. You might also say that given our lead time, we expect more orders to be able to be absorbed into our backlog for shipments over the next 6 quarters.

So taking that we expect revenue -- a minimum revenue one way to look at it of $73 billion over the next 6 quarters, but we do expect much more as more orders come in for shipments within the next 6 quarters. Our lead time depending on the particular product, it is -- can be anywhere from 6 months to a year.

On -- with respect to supply chain is what you're asking, critical supply chain on silicon and packaging, yes, that's an interesting challenge that we have been addressing constantly and continue to. And with the strength of the demand and the need for more innovative packaging, advanced packaging because you are talking about multi-chips in creating every custom accelerator now. The packaging becomes a very interesting and a technical challenge.

And building our Singapore fab is to really talk about partially in-sourcing those advanced packaging. We believe that we have enough demand, we can literally in-source not from the viewpoint of not just costs, but in a viewpoint of supply chain security and delivery. We're building up a fairly substantial facility for packaging, advanced packaging in Singapore, as indicated, purely for the purpose to address the package advanced packaging side. Silicon-wise, now we go back to the same pressure source in -- with TSMC. And so we keep going for more and more capacity in 2-nanometers, 3-nanometers, and so far, we do not have that constraint. But again, time will tell as we progress and as our backlog builds up.

One moment for our next question. The next question will come from the line of Blayne Curtis with Jefferies.

I wanted to ask, with the original $10 billion deal you talked about, a rack sale, I just wanted to -- with the follow-on order as well as the fifth customer, can you just maybe describe how you're going to deliver those? Is it an XPU? Or is it a rack? And then maybe you can kind of just walk us through the math and kind of what the deliverable is? Obviously, Google uses own networking. I'm kind of curious, too, would it be a copy exact of what Google does, now that you could talk to it to -- name? Or would you have your own networking in there as well?

That's a very complicated question, Blayne. Let me tell you what it is, it's a system sale. How about that? It's a [ real ] system sale. We have so many components beyond XPUs, customer accelerators in our AMI system -- in AI system, any I said them used by hyperscalers that yes, we believe it begins to make sense to do it as a system sales and be responsible. But be fully responsible for the entire system, or rack, as you call it. I think people are understanding as a system sales better.

And so on this customer number 4, we are selling it as a system with our key components in it. And that's no different than selling a chip. We certify a final ability to run as part of the whole selling process.

One moment for our next question. And that will come from the line of Stacy Rasgon with Bernstein.

I wanted to touch on gross margins and maybe it feeds into a little bit the prior question. So I understand why the AI business is somewhat dilutive to gross margins. We have the HBM pass-through. And then presumably with the system sales that will be more dilutive. And you hinted at this in the past, but I was wondering if it could be a little more explicit. As this AI revenue starts to ramp, as we start to get system sales, how should we be thinking about that gross margin number, say, we're looking out 4 quarters or 6 quarters? Is it low 70s? I mean could it start with the 6 at the corporate level?

And I guess I'm also wondering I understand how that comes down, but what about the operating margins? Do you think you get enough operating leverage on the OpEx side to keep operating margins flat? Or do they need to come down as well?

I'll let Kirsten give you the details, but enough for me to broadly high level explain to you, Stacy. Good question, phenomenal. Is -- you don't see that impacting us right now, and we have already started that process of some systems sales. You don't see that in our numbers. But it work. And we have said that openly. The AI revenue has a lower gross margin than our -- obviously, the rest of our business including software, of course. But we expect the rate of growth of -- as we do more and more AI revenue to be so much that we get the operating leverage on our operating spending that operating margin will deliver dollars that are still high level of growth from what it has been. So we expect operating library to benefit us at the operating margin level, even as gross margin will start to deteriorate high level.

No, I think Hock said that fairly. And the second half of the year when we do start shipping more systems the situation is straightforward. We'll be passing through more components that are not ours. So think of it similar to the XPUs where we have memory on those XPUs and we're passing through those costs. We'll be passing through more cost within the rack.

And so those gross margins will be lower. However, overall, the way Hock said it, gross margin dollars will go up, margins will go down, operating margins -- because we have leverage operating margin dollars will go up, but the margin itself as a percentage of revenues will come down a bit. But we're not -- I mean, we'll guide closer to the end of the year for that.

One moment for our next question. And that will come from the line of Jim Schneider with Goldman Sachs.

Hock, I was wondering if you might care to calibrate your expectations for AI revenue in fiscal '26 a little bit more closely. I believe you talked about acceleration in fiscal '26 off of the 65% growth rate you did in fiscal '25. And then you're guiding to 100% growth for Q1. So I'm just wondering if the Q1 is a good jumping off point for the growth rate you expect for the full year or something maybe a little bit less than that?

And then maybe if you could separately clarify whether your $1 billion of orders for the fifth customer is indeed OpenAI, which you made a separate announcement about.

Wow, there's a lot of questions here. Let me start off with '26. Our backlog is very dynamic these days, as I said, it is continuing to ramp up. And you're right. We originally 6 months ago said, maybe year-on-year AI revenues would grow in '26, 60%, 70%. Q1 we double. And Q1 '26 today, we're saying it double. And we're looking at it because all the thresholds keeps coming in, and we give you a milestone of where we are today, which is $73 billion of backlog to be shipped over the next 18 months. And we do fully expect, as I answered the earlier question, for that $73 billion over the 18 months decreased growing. Now moving is a moving number as we move in time. But it will grow.

And it's hard for me to pinpoint what '26 is going to look like precisely. So I'd rather not give you guys any guidance, and that's why we don't give you guidance, but we do give it for Q1. Give it time, we'll give it for Q2. And you're right, is in -- to us, it is an accelerating trend. And my answer is likely to be an accelerating trend as we progress through '26. I hope that answers your question.

One moment for our next question. And that will come from the line of [ Ben Reitzes ] with Melius Research.

I wanted to ask, I'm not sure if the last caller said something on it, but I didn't hear it in the answer was I wanted to ask about the OpenAI contract. It's supposed to start in the second half of the year and go through 2029 for 10 gigawatts. I'm going to assume that, that's the fifth customer order there. And I was just wondering if you're still confident in that being a driver, are there any obstacles to making that a major driver and when you expect that to contribute and your confidence in it.

You didn't hear the answer from my last caller Jim's question because I didn't answer it. I did not [indiscernible] and I'm not answering it either. It's the fifth customer, and it's a real customer and it will grow. They are on a multiyear journey to their own XPUs. And let's leave it at that.

As far as the open AI view that you have, we appreciate the fact that it is a multiyear journey that will run through '29 as our press release with OpenAI showed 10 gigawatts between '26 -- more like '27, '28, '29, Ben, not '26. It's more like '27, '28, '29, 10 gigawatts. That was the OpenAI discussion. And that's, I call an agreement and alignment of where we're headed with respect to various respective and valued customer OpenAI. But we do not expect much in '26.

One moment for our next question. And that will come from the line of C.J. Muse with Cantor Fitzgerald. .

I guess, Hock, I wanted to talk about custom silicon and maybe speak to how you expect content to grow for Broadcom generation to generation. And as part of that, your competitor announced CPX offering, essentially accelerator for an accelerator for massive context windows. I'm curious if you see a broadening opportunity for your existing 5 customers to have multiple XPU offerings.

Thank you. No, yes, it's -- you hit it right on. I mean the nice thing about a custom accelerator is you try not to do 1 size fits all and generationally. Each of these 5 customers now can create the XPU customer accelerator for training and inference in [indiscernible] basically it's almost 2 parallel threads going on almost simultaneously for each of them. So I have had plenty of versions to deal with. I don't need to create any more version. We've got plenty of different content out there just on the basis of creating these customer accelerators.

And by the way, when you do customer accelerators, you tend to put more hard way in that unique differentiated versus trying to make it work on software and creating kernels into software. And that's very tricky too. But thinking about the difference where you can create in hardware, those [indiscernible] call data routers, versus the dense matrix multipliers, all in 1 -- same [ chip ]. And that's what many of -- just 1 example of what creating customer accelerators is letting us do. Or for that matter, a variation in how much memory capacity or memory bandwidth from -- for the same customer from chip to chip, just because even in inference, you want to do more reasoning, first decoding versus something else like [indiscernible].

So you literally start to create different hardware for different aspects of how you want to train or influence and run your workloads. It's a very fascinating area, and we are seeing a lot of variations and multiple chips for each of our customers.

One moment for our next question. And that will come from the line of Harsh Kumar with Piper Sandler.

Yes, Hock and team, first of all, congratulations on some pretty stunning numbers. I've got an easy 1 and a more strategic one. The easy 1 is, you guide in AI Hock and Kirsten is calling for almost $1.7 billion of sequential growth. I was curious maybe you can talk about the diversity of the growth between the 3 existing customers? Is it pretty well spread out, all of them growing? Or is 1 sort of driving much of the growth.

And then Hock, strategically, 1 of your competitors bought a photonic fabric company recently. I was curious about your take on that technology and if you think it's disruptive or you think it's just [indiscernible] at this point in time.

I like the way you address this question because the way that you address the question to me is almost hesitant. Thank you. I appreciate that.

But on your first part, yes, we are driving growth and begin to feel like this thing never hit and it's a real mixed bag of existing customers and on existing XPUs. And I'll be pick of it, it's XPUs that we're seeing. And that's not to slow down the fact that as I indicated in my remarks and commented on the demand for switches, not just Tomahawk 6, Tomahawk 5 switches, the demand for our latest 1.6 terabit per second DSPs that enables optical interconnects for scale out particularly, it's just very, very strong. And by extension demand for the optical components and laser [indiscernible] just going nuts. All that come together.

Now all that is small relatively lesser dollars when it comes to XPUs, as you probably guess, I mean to give you a sense, maybe let me look at it on a backlog side. Of the $73 billion of AI revenue backlog over the next 18 months, I talked about, maybe $20 billion of it is everything else. The rest is XPUs. Hope that gives you a sense of what the mix is. But the rest is still $20 billion, that's not small by any means. So we value that.

So when you talk about your next question of silicon photonics, and as a means to create basically much better, more efficient, lower power interconnects in not just scale out, but hopefully, it scale up. Yes, I could see a point in time in the future when silicon photonics methods is the only way to do it.

We're not quite there yet. But we have the technology and we continue to develop the technology, even at each time we develop it first for 400 gigabits bandwidth, going on to 800 gigabit bandwidth, not ready for it yet. And even with the product -- and we're now doing it for 1.6 terabit bandwidth to create silicon photonics switches, silicon photonics interconnects, not even sure it will get fully deployed because engineers -- our engineers, our peers and the peers we have out there was somehow trying to find a way to still do -- try to do scale up within a rack in copper as long as possible and in scale out in nonpluggable optics.

The final, final straw is when you can do it well in pluggable optics. And of course, when you can do it even in copper, then you're right, you go to [indiscernible] and it will happen, and we're ready for it, just saying not anytime soon.

One moment for our next question. That will come from the line of Karl Ackerman with BNP Paribas.

Hock, could you speak to the supply chain resiliency and visibility you have with your key material suppliers, particularly [indiscernible] as you not only support your existing customer programs but the 2 new custom compute processors that you announced in your quarter. I guess what I can get at is you also happen to address the very large subset of networking and compute AI supply chains. You talked about record backlog. If you were to pinpoint some of the bottlenecks that you have, the areas that you're aiming to address [indiscernible] from supply chain bottlenecks, what would they be? And how do you see that ameliorating into '26?

It's across the board, typically. I mean it's -- we are very fortunate in some ways that we have the product technology and the operating business lines to create multiple key leading-edge components that enables today's state of the AI data centers. I mean our DSP, as I said earlier, is now at 1.6 terabits per second. That's the leading edge connectivity for bandwidth for this -- for the top on top of the [indiscernible] XPU and even GPU. And we intend to be that way. And we have the lasers, EMLs, VCSELs, [indiscernible] that goes with it.

So it's fortunate that we have all this and the key active components that go with it. And we see it very quickly, and we expand the capacity as we do the design to match it. And long -- this is a long answer to what I'm trying to get at, which is I think we are any of these data center suppliers of the system regs, not counting the power shell and all that. Now that starts to get beyond us. On the power shell and the transformers and the gas turbines.

If you just look at the rack, the systems on AI, we probably have a good handle on where the bottlenecks are because sometimes we are part of the bottlenecks, which we then want to get it [indiscernible] to resolve. So we feel pretty good about that through 2026.

One moment for our next question. That will come from the line of Christopher Rolland with Susquehanna.

Just first, a clarification and then my question. And sorry to come back to this issue. But if I understand you correctly, Hock, I think you were saying that OpenAI would be a general agreement, so it's not binding maybe similar to the agreements with both NVIDIA and AMD.

And then secondly, you talked about flat non-AI semiconductor revenue, maybe what's going on there? Is there still an inventory overhang? And what could -- what do we need to get that going again? Do you see growth eventually in that business?

Well, on the non-AI semiconductor, we see broadband literally recovering very well. And we don't see the others -- no, we see stability. We don't see a sharp recovery that is sustainable yet. So I guess, given -- a couple more quarters. But we don't see any further deterioration in demand. And it's more, I think, maybe to -- AI is sucking the oxygen a lot out of enterprise spending elsewhere and hyperscaler spending elsewhere. We don't see getting any worse. We don't see it recovering very quickly with the exception of broadband. That's a simple summary of non-AI.

With respect to OpenAI, without diving into it, I'm just telling you on 10 gigawatt announcement is all about. Separately, the journey with them on the custom [indiscernible] progresses at a very advanced stage and will happen very, very quickly. And it's -- and it will have a committed element to this whole thing. And then we -- but what I was articulating earlier was the 10 gigawatt announcement. And that 10 gigawatt announcement is an agreement to be aligned on developing 10 gigawatts for OpenAI over '27 to '29 time frame. That's different from the XPU program we're developing with them.

And we do have time for one final question, and that will come from the line of Joe Moore with Morgan Stanley.

Great. So if you have $21 billion of rack revenue in the second half of '26, I guess do we stay at that run rate beyond that? Are you going to continue to sell racks? Or does that sort of that type of business mix shift over time? And I'm really just trying to figure out the percentage of your 18-month backlog that's actually full systems at this point.

Well, it's an interesting question. And on that question basically comes to how much compute capacity is needed by our customers over the next, as I say, over the period beyond 18 months. And your guess is probably as good as mine based on what we all know out there, which is really what they relate to. But if they need more, then you see that continuing even larger. If they don't need it, then probably it won't. But as of -- what we are trying to indicate is that the demand we are seeing over that period of time right now.

I would now like to turn the call back over to Ji Yoo for any closing remarks.

Thank you, operator. This quarter, Broadcom will be presenting at the New Street Research Virtual AI Big Ideas Conference on Monday, December 15, 2025. Broadcom currently plans to report its earnings for the first quarter of fiscal year 2026 after close of market on Wednesday, March 4, 2026. A public webcast of Broadcom's earnings conference call will follow at 2:00 p.m. Pacific.

That will conclude our earnings call today. Thank you all for joining. Operator, you may end the call.

This concludes today's program. Thank you all for participating. You may now disconnect.

Okay. Good afternoon, everybody. Welcome to the Goldman Sachs Communacopia and Technology Conference. I'm Jim Schneider, the Semiconductor analyst here at Goldman Sachs. It's my real distinct pleasure to welcome Broadcom and CEO, Hock Tan, on the stage today. Welcome, Hock.

Thank you.

Thanks for being here. I think it's fair to say that your business portfolio has really evolved radically over the past several years. Company is much bigger today, partly because the software business is much bigger due to VMware, but also really just the Semiconductor business is so much bigger now due to AI. If you just sort of step back and think about the top 2 or 3 priorities for Broadcom over the next few years, what are they? And then if we sit down here again in 5 years, what do you think is going to be the one thing that surprises investors the most?

Well, I'm not sure what will surprise you guys the most, but I'll answer that first part. Right now, front and center for us as we see it, is this huge opportunity in AI, in AI compute to be precise. I mean, we are seeing very, very interesting, very, very strong demand for AI compute. And we are driving the business model. We're driving resources of this company to addressing the specific needs from, to be honest, a very narrow group of customers.

If you think philosophically about AI for a second, maybe help us understand your vision personally for where AI is going as a technology. Why does the world need AI? How useful do you think it's going to be? Is it going to live up the amount of investment that the investment community is putting into it today?

I hate this question you know, na? Because I'm not like that. I'm not a cheerleader. I try not to be anyway. It's hard not to hear. No with -- I think AI is here at least to stay for a while because I think we have discovered what AI could do. It's this huge sophisticated algorithms that enables us based on history to so impressively predict what would happen next. [ I thought on -- I look at AI as, ] even generative AI, now many of you probably say it's more than that, but it is pretty good.

And because of that, it's a great tool to be used to basically make us feel more intelligent and probably it does. But where we are, coming back to your first question on Broadcom is, we're very -- the key part, a lot of all this is we're not writing the algorithms to create the LLMs. We make no to such pretrained model that does it. What we're doing is we enable it because you can write and create the best algorithms in the world, and they exist today and they keep improving. But if you don't have the ability to create compute, to manifest it as compute, all of it is just nice equations and algorithms on whiteboard is not practical.

So -- and the manifestation of it is that compute, which is where our know-how in semiconductors, especially in being able to create a very effective compute vehicle, compute engine is what makes the difference. And that underpins where we are. And what we're seeing for the next 3 years is accelerating demand for that compute capacity. In that narrow group of customers, we focus our attention on, Jim, is those guys trying to creating those LLMs out there. In many ways, this LLM is another term for these people -- these players trying to create super intelligence and by so doing it, over time, create the new platforms of tomorrow.

Yes. Are there any use cases for AI that excite you most personally in terms of utility, monetization or otherwise?

It's great for you to be able to create a nice poem for my wife. I'm sure it does for you, too. That's a use case that excite me.

Excellent. Maybe you want to turn it back to Broadcom's business for a second, but maybe if there's a...

I told you I hate this kind of question.

Let's talk about your business. You just guided your fiscal '26 AI revenue to grow on, previously in the 60% range. On the earnings call last week, you had -- you announced that you expect this growth to accelerate materially as now you've converted another customer who's focused on inference. Maybe help us understand the broader trend you're seeing among customers here. Is the demand intensity higher for inference or for training right now?

Well, I think, it's hard to take a snapshot at any point and say it is. I think what we see here among the seven "players that we have." We consider both customers and others that consider prospects. One thing I see in common is they are the ones investing to create best LLMs out there. And it's a constant road map. It's a constant journey for them. And that creation of intelligence improvement comes only from creating better LLMs, which comes from research in training.

Training, create the intelligence, inference, monetize the intelligence. So you tell me which happens more. It's a function, I think, at that point in time when one of those players we engage with decide, "Hey, it's time to create -- scale to create monetization." So they're going to inference. Otherwise, if you're just going for super intelligence, just focus on research, that's training. And I think both happens. And both are at very high level of investment usage at this point.

You mean -- give us a sense for 2026, how big is the acceleration beyond the 60% you've already called for? And more importantly, do you think that growth rate is even faster in '27 for you?

I'd rather not talk about '26, '27 because I don't give that guidance. Just to give you a sense, yes, we did indicate 60%, '26 growth rate, which was the growth rate of '25. It looks like we're accelerating beyond that in '26. We put a milestone. We gave you a little clues all over the place just to confuse you. But we gave you a milestone that says -- and that was done about a year ago, earnings call, that we see a served available market, available market, not a forecast of $60 billion to $90 billion. Remember that, right?

So now you ask me this question. Let me throw you something that might -- I mean you think about. As you know, I signed up for an extension of my contract to 2030. With this extension, of course, I get some incentives. One of that incentives that tied to it is tied to -- so you guys know AI revenue for Broadcom, very simple. And I need to hit to achieve max on my incentive for 2030, 5 years from now, give you a sense, AI revenue of exceeding $120 billion, okay? Today, '25, our AI revenue is $20 billion. But that gives you a sense of our belief how strong the demand for compute is in this race towards AI, generative AI, super intelligence.

That's a good CAGR, Hock.

Well, that is a question whether you get there.

Very good. It's great to hear that -- the detail. Just one last question on this AI revenue piece. You've talked previously about your three existing customers. You've added a fourth, and you've talked about three additional prospects you have in the pipeline. Maybe just kind of give us a sense about, do you see a universe of potential customers outside those four plus three prospects? Or is it pretty much those seven is what you see as the potential opportunity?

For Broadcom, we're driving a business model just on those guys. These are guys who are creating those LLMs and who knows how long these guys will survive, whether they will go through the whole race. But what we see now is about seven guys doing it. And beyond that, we do not see those as our market at all.

Okay. Fair enough. Maybe just go a little deeper on some of the customer side of things. I think in the past, you said that you view companies that have their own LLM capabilities to be the most valid important customer prospects. Separately, I think you've also said that you want to pursue largest volume opportunities available in the market. So maybe give us -- investors a sense of how do you diligence the different customer prospects? Can you discriminate customers between ones that have promise and the ones that are a little bit less interesting to you? How do you make that determination?

I think we kind of made it simple now these days. Those customers who are doing -- really doing LLMs, that's our customer. That's it. We'll stop looking at whether they're big or small.

Okay. I think it's easy. Relative to competitive landscape, I think, the conventional wisdom out there in the investment community is probably that custom ASICs are good for internal workloads and merchant GPUs are more broad-based, and they're good for more of the external workloads as well as internal workloads. I think we've seen more than a few headlines of the advantage of one of your largest customers on the custom ASIC side and what's happening there.

So I'm kind of curious, do you see that conventional wisdom shifting at all? Or do you see it changing? Do you see a point in time where enterprises are going to have more software capabilities and actually are able to use ASICs to some of the major CSPs?

I don't think we look at it that way. And to answer your question, I don't think enterprises, at least in the foreseeable future would ever want to consider developing the core technology for them to enable AI computing. Rather, the way we have looked at this market consistently is really in generative AI that we see today, it's two broad simple segments. There's one segment, as I mentioned, of these few players and some of them are hyperscalers, some of them are not. They're just, for one of a better word, super start-ups in AI. I know some of you call them as labs, perhaps it is.

These are the guys who are really making that investment in LLMs, which entail huge amount of R&D, particularly in respect to training, which is just -- not just creating that algorithms to create those LLMs and make it work, but to also create the ability and spend on the ability to train huge amounts of data to improve on their journey towards super intelligence as we -- for one of a better word. That's one category.

And then we have a second category of customers who are almost like the rest of us. I call it enterprise. They probably include sovereign. They include public cloud guys. But at the end of the day, they go back to enterprise out there. There's thousands and thousands of companies out there who are looking at AI as a very interesting tool to improve the way they run their business to basically achieve mostly productivity gains, which is truly -- and trying to find use case for it and doubling most of them POCs right now on AI tools and trying to run AI and whether they get that ability through renting GPUs or buying some and running on-prem, these are that category.

And that category of market are largely inference, truly inference. They talk about training. It's limited. Post training, perhaps, test time scaling, but it's all inference. It's looking for use case and trying to get a return on investment, mostly on productivity. And these guys are going to stay largely on merchant silicon, merchant GPU because they're not going to create XPUs, write software stacks to make it work and figure out all these things, who cares? All they want is get their models running and a return on investment.

These guys, I do not see, at least in the near term, ever going beyond merchant GPU. There's 10,000 of them perhaps today, each spending $10 million a year, maybe more. That's a big size, $100 billion market. In contrast, we see these few guys creating LLM, very few. But they are able to spend today some $30 billion a year on AI compute. So you have $100 billion here and you're maybe $100 billion, $200 billion there. We focus on one market, and they're very distinct from the other market.

Very fair. And then also serving the custom compute segment. Can you maybe talk about how you add value there and the defensibility of your position? In other words, talk about some of the core IP that you add or you augment your customers' design with the value you provide through the manufacturing process. But to the point, when investors ask what -- how defensible is your position? And are you at risk of your customers, your biggest customers going straight to a foundry with their solution? How do you answer that?

It's the same way why we are -- where we are in the Semiconductor business over the years. It's engineering. It's deep engineering, advanced engineering IP, especially in the semiconductor space, some part of it in the software space. It's about having a lot more intellectual property in creating your chips at the end of the day, but it's even beyond that. Thanks for that question. People think about AI compute as the GPU or the XPU, which is what we call the actual multiplier engine that's used to do matrix multiplication and regression.

But it's much more than that because, as you know, each single GPU or XPU, no matter how powerful we all make it. And we can make them very powerful by cramming more and more multipliers, [indiscernible] multipliers into this 800 square millimeter piece of silicon. And you do that by going deeper and deeper submicron process technology so you can cram more and go even one step further by then -- instead of making it one die in one package or chip, you do multi-dies. We have reached a point where the latest product we're doing, we have actually three dies in one package. That's how much we have crammed in.

So we can make this GPU or XPU in our case, super, super good in doing this multiplication and matrix regression. But that's not enough. You don't create train -- you are not able to do this generative AI training of this huge database to create this billion, billion parameters models with one single GPU or XPU, no matter how hard you try. You need to create a cluster of them.

And the more complex your LLM becomes, the larger your clusters have to be. Then you face yourself a totally different problem. How do you create all this matrix multiplication across, say, 100,000 of the GPU simultaneously. And then you go to 1 million. That's even a bigger headache. That's one of the -- that is, in fact, I consider the biggest technology challenge in computing on generative AI as you progress your models higher. And that comes in the other aspect, networking. How do you connect them? How do you have the bandwidth to connect them? How do you schedule the workloads, orchestrate it so that they can all run in parallel. This is more so as much software as the networking side, which is hardware.

I think this is one of the biggest challenge towards progression, making massive progress on being able to create super intelligence. How do you get these huge workloads that you're supposed to run on these huge models and be able to train a huge amount of data on a 1 million GPU or XPU simultaneously and converge on a solution. And that's part of it is slowly coming to a surface that is networking, that might be the biggest problem because in generative AI, I said that before, I'll say it again, the network becomes the computer, not any single GPU or XPU.

We all tend to focus on that. It's way beyond that. Developing XPU, GPU and when trying to come up, it's hard, a lot of intellectual property required, but truly, it's easier than doing the networking. That's where the challenge starts. And yes, that's how we differentiate ourselves. We come up with a better mousetrap, better technology and just outrun the competition, no different than what we're doing in all other areas in semiconductors where we have been very successful.

Yes. And that's a perfect segue into the next question I am going to ask about networking because clearly, that's a mainstay franchise for the company historically. Talk about -- I think you've previously talked about in these AI networks, I think you said scale-up opportunity is about 5 to 10x larger in content relative to scale out. How do you think about the adoption curve for Ethernet and scale up? And just sort of talk about your sort of overall product portfolio and how it gets there?

Ethernet is going to happen inevitably because you don't want to have to create new protocols, which are closed protocols for -- to replace something that is frankly available, tested, proven for the last 20, 30 years and where most of the operators, designers with the hyperscalers who are the leaders in developing are very familiar with. It is and it comes to a bigger picture of what you -- where you're getting at is disaggregation. There's more with XPUs. We are in Broadcom in a way, disaggregating the XPU or GPU from the networking side. That's a choice.

You can choose to have whatever GPU, XPU you prefer and you can choose whatever networking switching you prefer. And that disaggregation is eventually going to optimize by being able to pick the best of breed of each, your ability to do generative AI computing and leads to eventually best performance in reaching the LLMs outcomes that you want to drive to. It's less a platform. It's really about disaggregating, which is exactly what hyperscalers have when they create cloud computing, public cloud computing, reach -- achieve towards that direction.

They disaggregate hardware, software, CPU from networking, everything. I think you will see in generative AI, that's the path is headed now on those platform LLM guys I talked about. When it comes to enterprise, however, I think it's still very OEM-based simply because the enterprise don't have the interest nor capability to create that need to disaggregate and optimize. The footprint are not big enough. But those guys doing the LLM, those few guys I talked about, for us, we're dealing with seven. They will disaggregate, they will optimize.

And then if you think about your outlook for the Networking business, obviously, some -- however much AI grows, how do you think about the networking business tracking that envelope faster, slower, the same?

I tend to think it will -- a lot of that will be tied again back to that narrow group of customers. We talk about the big hyperscalers and those super startups. And the growth will gross from no other reason than it's a move to Ethernet, which is one thing. But more than that, that the ability as you go to larger and larger clusters, the importance of scaling up within a rack becomes super important.

And when you do that level of scale up within rack for AI, generative AI computing, this is where the matrix multiplication truly happens in your model, you're talking about massive bandwidth. You're talking of bandwidth now soon that will go to -- and this is about not just GPU to GPU, XPU to XPU connection, you're talking about memory sharing across GPU. And you know how much memory these GPUs have. So you really want to have a bandwidth that goes to 100 terabit per second and not stick at the 28 we are seeing today in the copper racks that is you want to go optical. You want to be able to connect not just 72 GPUs to each other. You want to connect 512, even to 1,024 XPUs to XPUs.

That's the scale you want to drive through, the bandwidth you want to drive through which will create very much faster convergence to the solution -- to the training solution outcome you're looking for. That's the road map we're driving towards. And we're not talking years away. We're talking like within the next year or 2, '26, '27, that will start to happen. The product is out there, and the technology is available. It's a question of the deployment at this point.

Very good. One question I get a lot of times is on the Co-Packaged Optics and their role in the industry going forward. Maybe how ready is that technology, in your opinion, for mainstream deployment or large customer deployment? And how do you think the adoption curve is going to go for CPO?

I love to hype you guys on things. The latest is Co-Packaged Optics. It's just optics, okay? I don't know why you need to call it Co-Packaged Optics, except it's silicon photonics. Now the world is still getting ready for silicon photonics. A big part of the reason is what I said earlier, when you do scaling up to a scale larger and larger because the clusters get bigger and bigger and you're an LLM player, so you do that.

If you're not an LLM player, you're a little enterprise and you just run one rack or maybe no more than 36 GPUs, you don't need any of that. Run direct attach copper, you're done. So these big LLM guys trying to run 100,000, 200,000, 500,000 XPU or GPU. Then you want, as I said earlier, scaling up in a rack, you want to make it optical because by doing it optical, you can connect directly 512 compute nodes, XPU or GPU to each other, much better than anything else you can do.

And potentially, you go one step further on bandwidth at your -- [ direct ] power switch, you can go to 1024, but that's optics, moving away from the copper we are seeing today. Now copper will still be around 2026 mostly. But by 2027, it's all going to be optical. Then people talk about what optical solution do you need? So people talk about, "Hey, one thought is called silicon photonics." And one manifestation of that is Co-Packaged Optics, where you integrate the active components in a fiber optic interconnect into the silicon, whether it's a GPU silicon or the switch silicon, you do both. That's called co-package opting. And that's a dream.

This is really a good dream because you reduce power, we know by 40%, you do it. All that is great. And we -- and at Broadcom, we already have the technology done 3 years ago. The issue is this optics, typically optical interconnect because it's a lot mechanical, have anywhere from 5% to 8% failure rate. So when you have pluggable optics that you guys hear about, when it fails, you just unplug it and put in a new one, pluggable.

When you do co-package optics or silicon photonics, you integrate your GPU, that expensive $40,000 GPU into the optics that fails at 5%, 8%, you got a problem. So the question is, which we have been studying, when you create silicon photonics, does the integrated solution have the failure characteristic of silicon, which is like 0.1% barely or optics, which is 5% to 8%. That data, we're still collecting. And we have been collecting for the last 2, 3 years. I'd like to believe it will have the characteristics of silicon rather than optics. But again, as we test through it, we'll figure it out. So I'm sorry, long answer to your short question.

No, it's quite okay, quite okay. And then finally, I think you talked about competitively, your belief is that Ethernet is going to win in the end. Maybe talk about over what time frame you think that occurs? And just kind of like any of the factors that would slow that down from happening?

The easiest answer is as among the few hyperscalers we deal with, the ones who are doing LLMs, they are the ones who are also seeing that it's an existential need to create their own custom accelerators, the XPUs. As they replace -- as they start to deploy XPU in a steady, progressive manner like all things over the next 5 years, you'll see Ethernet come into play. So there's a direct correlation back to disaggregation, the disaggregation out of integrated platform [racks ] to one that's more disaggregated with XPUs and Ethernet switching, interconnects, one that is kind of more optimized for the other.

Yes. I'd be remiss if I didn't touch on your software portfolio for a moment...

I knew you were getting to that.

Maybe just talk broadly. I mean you've obviously done a great job of acquiring companies in the enterprise, where you have very sticky customer bases, very -- driving very high margins over time for that business. Maybe talk about sort of the durability of your software franchise as you see it, whether it be in VMware or all the products you've had up until now on the existing business? And then maybe just kind of address your appetite for M&A in the software space? Or are you 100% focused on the AI opportunity at this point?

I love the Software business we have acquired because we are very careful what we buy. And the whole game and the whole thing in software is you hit it right on, durability, sustainability. And the key thing to it is you have to invest. Unlike what you may think, we actually invest a lot in the software we have. If nothing else, to make it be better for the customers who are using it. And not only that, proper investing in the technology to make it better, you invest in support.

Enterprises will always break software. When they do, you want to make it very resilient so that they can get it back up fast. So support is important. And that's why we make a lot of investment that in services and support and keeping it durable. If you do that, we accept the fact that we don't need our Software business to grow dramatically, just sustain and grow single digits. Just by doing that and not trying to just push to grow, you make very good margin. And you've seen that in the way we announced it, very good profitable margin, but a stable kind of business.

And here's the funny part about answering your first part of it, non-AI and software, great businesses of ours contribute a big part of our revenue and earnings. But in a matter of a few years, my AI revenue will exceed a combination of both. [ It isn't ] a really a matter of 1 or 2 years, it will exceed.

So from our viewpoint, with this big opportunity and the dramatic growth I have indicated that we are going through, we're kind of looking at it and say, you don't need to buy anything else, you don't need to invest in anything else to accrete to your revenue nor earnings. Just keep investing in AI. And our view for the next few years, that's where we are.

That's a great place to end it. Thank you very much, Hock, for being here. We really appreciate it.

Thank you.

Welcome to Broadcom Inc.'s Third Quarter Fiscal Year 2025 Financial Results Conference Call. At this time, for opening remarks and introductions, I would like to turn the call over to Ji Yoo, Head of Investor Relations of Broadcom Inc. Please go ahead.

Thank you, Sheri, and good afternoon, everyone. Joining me on today's call are Hock Tan, President and CEO; Kirsten Spears, Chief Financial Officer; and Charlie Kawwas, President, Semiconductor Solutions Group.

Broadcom distributed a press release and financial tables after the market closed, describing our financial performance for third quarter of fiscal year 2025. If you did not receive a copy, you may obtain the information from the Investors section of Broadcom's website at broadcom.com. This conference call is being webcasted live and an audio replay of the call can be accessed for 1 year through the Investors section com's website.

During the prepared comments, Hock and Kirsten will be providing details of our third quarter fiscal year 2025 results, for our fourth quarter of fiscal year 2025 as well as commentary regarding the business environment. We'll take questions after the end of our paired comments. Please refer to our press release today in our recent filings with the SEC for information on the specific factors that could cause our actual results to differ materially from the forward-looking statements made on this call.

In addition to U.S. GAAP reporting, Broadcom reports certain financial measures on a non-GAAP basis. A reconciliation between GAAP and non-GAAP measures is included in the tables attached to today's press release. Comments made during today will primarily refer to our non-GAAP financial results.

I will now turn the call over to Hock.

Thank you, Ji. And thank you, everyone, for joining us today. In our fiscal Q3 2025, total revenue was a record $16 billion up 22% year-on-year. Our revenue growth was driven by better-than-expected strength in AI semiconductors and our continued growth in VMware. Q3 consolidated adjusted EBITDA was a record $10.7 billion, up 30% year-on-year. Now looking beyond what we are just reporting this quarter, with robust demand from AI, bookings were extremely strong.

And our current consolidated backlog for the company hit a record $110 billion. Q3 semiconductor revenue was $9.2 billion as year-on-year growth accelerated to 26% year-on-year. And this accelerated growth was driven by AI semiconductor revenue of $5.2 billion, which is up 63% year-on-year and extend the trajectory of robust growth to 10 consecutive quarters. Now let me give you more color on our XPU business which accelerated to 65% of our AI revenue this quarter.

Demand for custom AI accelerators from our 3 customers continue to grow as each of them journeys at their own pace towards compute self-sufficiency. And progressively, we continue to gain share with these customers. Now further to these 3 customers, as we had previously mentioned, we have been working with other prospects on their own AI accelerators. Last quarter, one of these prospects release production orders to Broadcom. And we have accordingly characterized them as a qualified customer for XPUs. And, in fact, has secured over $10 billion of orders of AI regs based on our XPUs. And reflecting this, we now expect the outlook for fiscal 2026 AI revenue to improve significantly from what we had indicated last quarter.

Turn to AI networking. Demand continued to be strong because networking is becoming critical as LLM continue to evolve in intelligence, computer classes who grow bigger. The network is the computer, and our customers are facing challenges as they scale to classes beyond compute nodes. For instance, Scale Up, which we all know about, is a difficult challenge when trying to create substantial bandwidth to share memory across multiple GPUs or XPUs within rate. Today's AI rate scales up a mere 72 GPUs at 28.8 terabytes second bandwidth using a proprietary MVLink.

On the other hand, earlier this year, we have launched Tomahawk 5, we've opened AI with open Ethernet, sorry, which can scale up 512 compute nodes for customers using XPUs. Moving on to scaling out across rigs. Today, the current architecture using 51.2 terabit per second requires 3 tiers of networking switches. In June, we launched Tomahawk 6 and our Ethernet-based 102 terabit per second switch, which flattens the network to 2 tiers. Resulting in lower latency, much less power. And when you scale the clusters beyond a single data center footprint, you now need to scale computing across data centers. And over the past 2 years, we have deployed our Jericho3 Ethernet router with hyperscale customers to just do this. And today, we have launched our next-generation Jericho4 Ethernet fabric routed with 51.2 terabit per second deep offering intelligent congestion control to handle classes beyond 200,000 compute nodes crossing multiple data centers. We know the biggest challenge to deploying larger clusters of compute for generative AI will be in networking.

And for the past 20 years, Broadcom has developed for Ethernet networking is entirely applicable to the challenges of Scale Up, scale out and scale across in generative AI. And turning to our forecast. As I mentioned earlier, we continue to make steady progress in growing our AI revenue. For Q4 2025, we forecast AI semiconductor revenue to be approximately $6.2 billion up 66% year-on-year.

Now turning to non-AI semiconductors. Demand continues to be slow to recover. In Q3, revenue of $4 billion was -- fed sequentially. While broadband showed strong sequential growth, enterprise networking and service storage were down sequentially. Wireless and industrial were flat quarter-on-quarter as we -- in contrast, in Q4, driven by seasonality, we forecast non-AI semiconductor revenue to grow low double digits sequentially and to approximately $4.6 billion. Broadband, service storage and wireless are expected to improve, while enterprise working remains down quarter-on.

Now let me talk about our infrastructure software segment. Q3 Infrastructure software revenue of $6.8 billion was up 17% year-on-year, above our outlook of $6.7 billion as bookings continued to be strong during the quarter. We booked in fact, total contract value over $8.4 billion during Q3.

But here is what I'm most excited about. After 2 years of engineering development by over 5,000 developers, we delivered on a promise when we acquired VMware. We released VMWare Cloud Foundation version 9.0 a fully integrated cloud platform, which can be deployed by enterprise customers on-prem or carry to the cloud. It enables enterprises to run any application workloads, including AI workloads on virtual machines and on modern containers. These provides the real alternative to public cloud.

In Q4, we expect infrastructure and software revenue to be approximately $6.7 billion, up 15% year-on-year. And in summary, continued strength in AI and VMware will drive our guidance for Q4 consolidated revenue to approximately $17.4 billion. up 24% year-on-year, and we expect Q4 adjusted EBITDA to be 67% of revenue.

And with that, let me turn the call over to Kirsten.

Thank you, Hock. Let me now provide additional detail on our Q3 financial performance. Consolidated revenue was a record $16 billion for the quarter, up 22% from a year ago. Gross margin was 78.4% of revenue in the quarter better than we originally guided on higher software revenues and product mix within semiconductors.

Consolidated operating expenses were $2 billion, of which $1.5 billion was research and development. Q3 operating income was a record $10.5 billion, up 32% from a year ago. On a sequential basis, even as gross margin was down 100 basis points on revenue mix. Operating margin increased 20 basis points sequentially to 65.5% on operating leverage. Adjusted EBITDA of $10.7 billion or 67% of revenue was above our guidance of 66%. This figure excludes $142 million of depreciation.

Now a review of the P&L for 2 segments, starting with semiconductors. Revenue for our Semiconductor Solutions segment was $9.2 billion, with growth accelerating to 26% year-on-year driven by AI. Semiconductor revenue represented 57% of total revenue in the quarter. Gross margin for our Semiconductor Solutions segment was approximately 67%, down 30 basis points year-on-year on product mix. Operating expenses increased 9% year-on-year to $951 million on increased investment in R&D for [ Ledge ] AI semiconductors. Semiconductor operating margin of 57% was up 130 basis points year-on-year and flat sequentially.

Now moving on to infrastructure software. Revenue for infrastructure software of $6.8 billion was up 17% year-on-year and represented 43% of revenue. Gross margin for infrastructure software was 93% in the quarter compared to 90% a year ago. Operating expenses were $1.1 billion in the quarter, resulting in infrastructure software operating margin of approximately 77%. This compares to operating margin of 67% a year ago, reflecting the completion of the integration of VMware.

Moving on to cash flow. Free cash flow in the quarter was $7 billion and represented 44% of revenue, we spent $142 million on capital expenditures. Day sales outstanding were 37 days in the third quarter compared to 32 days a year ago. We ended the third quarter with inventory of $2.2 billion, up 8% sequentially in anticipation of revenue growth next quarter. Our days of inventory on hand were 66 days in Q3 compared to 69 days in Q2 as we remain disciplined on how -- inventory the ecosystem.

The third quarter within $10.7 billion of cash and $66.3 billion of gross principal debt. The weighted average couponing and years to maturity of our $65.8 billion in fixed debt is 3.9% and 6.9 years, respectively. The weighted average interest rate and years to maturity of our $500 million at floating rate, debt is [ 4.7% ] and 0.2 years, respectively.

Turning to capital allocation. In Q3, we paid stockholders $2.8 billion of cash dividends based on a quarterly common stock cash dividend of $0.59 per share. In Q4, we expect the non-GAAP diluted share count to be approximately 4.97 billion shares excluding the potential impact of any share repurchases. Now moving to guidance. Our guidance for Q4 is for consolidated revenue of $17.4 billion, up 24% year-on-year. We forecast semiconductor revenue of approximately $10.7 billion, up 30% year-on-year. Within this, we expect Q4 AI semiconductor revenue of $6.2 billion, up 66% year-on-year. We expect infrastructure software revenue of approximately $6.7 billion, up 15% year-on-year.

For your modeling purposes, we expect Q4 consolidated gross margin to be down approximately 70 basis points sequentially, primarily reflecting a higher mix of XPUs and also wireless revenue. As a reminder, consolidated gross margins through the year will be impacted by the revenue mix of infrastructure software and semiconductors and product mix within semiconductors. We expect Q4 adjusted EBITDA to be 67%. We expect the non-GAAP tax rate for Q4 and fiscal year 2025 to remain at 14%.

I will now pass the call back to Hock for some more exciting news.

I don't know but exciting, Kirsten, but I do. I thought before we move to questions, I should share an update. The Board and I have agreed that I will continue as the CEO of Broadcom through 2030 at least. These are exciting times for Broadcom and I'm very enthusiastic to continue to drive value for our shareholders.

Operator, please open up the call for questions.

[Operator Instructions]. And our first question will come from the line of Ross Seymore with Deutsche Bank.

Thank you for sticking around for a few more years. So I just wanted to talk about the AI business and specifically the XPU. When you said you're going to grow significantly faster than what you had thought a quarter ago, what's changed? Is it just the impressive prospect moving to a customer definition. So that $10 billion backlog that you mentioned? Or is it stronger demand across the existing 3 customers? Any detail on that would be helpful.

I think it's both, Ross. But to a large extent, is the fourth customer that we now add on to our roster, which we will ship pretty strongly in 2026, beginning of 2026, I should say. So a combination of increasing volumes from existing 3 customers and we moved through that very progressively and steadily and the addition of a fourth customer with immediate and fairly substantial demand really put our really changes our thinking of what '26 would be starting to look like. Thank you.

[Operator Instructions]. That will come from the line of Harlan Sur with JPMorgan.

Good afternoon. Congratulations on the well executed quarter and strong free cash flow. I know everybody is going to ask a lot of questions on AI, Hock. I'm going to ask about the non-AI saving business. If I look at your guidance for Q4, it looks like the non-AI sing business is going to be down about 7%, 8% year-over-year in fiscal '25, if you hit the midpoint of the Q4 guidance.

Good news, the negative year-over-year trends have been improving through year. In fact, I think you guys are going to be positive year-over-year in the fourth quarter. You characterized it as relatively close to the cyclical bottom, relatively slow to recover. However, we have seen some green shoots of positivity, right, broadband, server storage and enterprise networking, you're still driving the DOCSIS upgrade in broadband cable. You've got on upgrades in China and the U.S. in front of you, enterprise spending on network upgrades is accelerating. So near term, from the cyclical bottom, how should we think about the magnitude of the cyclical upturn. And given your 30- to 40-week lead times, are you seeing continued order improvements in the non-AI segment, which would point you to continued cyclical recovery into next fiscal year?

Well, if you take a look at that non-AI segment, I mean you're right, from a year-on-year Q4 guidance, we are actually up, as I say, slightly couple 1% or 2% from a year ago. It's not much really to shout about at this point. And the big issue is the puts and takes and the put and takes and the bottom of all this is other than seasonality that we perceive, if you look at the short term, without looking year-on-year by looking sequentially, we see in things like wireless and we even start to see some seasonality in server storage these days. We don't -- it kind of all washes out so far.

The only consistent trend we've seen over the last 3 quarters that is moving up strongly is broadband. And nothing else, if you look at it from a cyclical point of view, seems to be able to sustain an uptrend so far. I don't think it's -- but as a whole, they are not getting worse, as you pointed out, Harlan, but they are not showing a V-shaped recovery as a whole, that we would like to see and expect to see in cyclical semiconductor cycles. The only thing that gives us some hope is broadband at this point, and it is recovering very strongly. But then it was a business that was most impacted in the short downturn of '24 and early '25. So again, one take that with a grain of salt.

But the answer to you for you is non-AI semiconductor is kind of slow to recover, as I said. And Q4 year-on-year is maybe low single digit is the best way to describe it at this point. So I'm expecting to see more of a U-shaped recovery in non-AI, and perhaps by late mid '26, late '26, we'll start to see some meaningful recovery. But as of right now, not clear.

Are you starting to see that in your order trend in your order book just because your lead times are like 40 weeks, right?

We are. But we've been track before we are -- the bookings are up, and they are up year-on-year in excess of 20%. Nothing like what AI bookings to lie but 23% is still pretty good, right?

[Operator Instructions]. That will come from the line of Vivek Arya with Bank of America.

Best wishes for the next part of your tenure. My question is on, if you could help us quantify what is the new fiscal '26 AI guidance. Because I think the last call, you mentioned fiscal '26 could grow at the 60% growth rate. So what is the updated number? Is it 60% plus the $10 billion that you mentioned? And sort of related to that, do you expect the custom versus networking mix to stay broadly what it has been this past year or evolve more towards custom. So any quantification and this networking versus customer mix would be very helpful for fiscal '26.

Okay. Let's answer the first part first, if I could be so bold as to suggest to you, when I -- last quarter when they said, "Hey, the trend of growth of '26 will mirror that of '25which is 50%, 60% year-on-year. That's really all I said. I didn't -- but of course, it comes up 50%, 60% because that's what '25 is.

All I'm saying, if you want to put another way of looking at what I'm saying, which is perhaps more accurate is we're seeing -- the growth rate accelerates as opposed to just remain steady at that 50%, 60%. We are expecting and seeing 2026 to accelerate more than the growth rate we see in '25. And I know you love me to throw in the number at you, but we are not supposed to be giving you a forecast for '26. But best way to describe it, it will be fairly material improvement.

And the networking versus custom?

Good point. Thanks for reminding me. As we see -- and a big part of this driver of growth will be XPUs, at the risk of repeating what I said in my remarks, comes from the fact that we continue to gain at our 3 original customers. They have to, they're on their journey and each passing generation they go more to XPUs. So we are gaining share from these 3. We now have the benefit of an additional fourth significant customer. I should say, fourth and very significant customer. And that combination will mean more exposed. And as I said, at the rate as we create more and more experience among 4 guys, Networking, we get the networking with this 4 guys, but now the mix of networking from outside this 4 guys will now be a small bet diluted via smaller share. So expect actually networking percentage of the pool to be a declining percentage going into '26.

[Operator Instructions]. And that will come from the line of Stacy Rasgon with Bernstein Research.

I was wondering if you could parse out this $110 billion backlog. Did I hear that number right? Could you give some color on the makeup of it -- like how far does that go -- and like how much of that $110 billion is AI versus non-AI versus software?

Well, yes, Stacy, we generally don't break up back on digital to give you a sense of how strong the business is as a whole for the company, and it's largely driven buying AI in terms of growth. Software continued to add on a steady basis. And non-AI, as I indicated, has grown double digits. Nothing compared to AI, which has grown very strongly. Give you a sense, perhaps fully 50% of it at least is semiconductors.

Okay. And it's fair to say that semiconductor piece, it's going to be much more AI than non AI.

Right.

[Operator Instructions]. And that will come from the line of Ben Reitzes with Melius.

I appreciate it. Hock, congrats on being able to guide to the AI revenue well above 60% for next year. So I wanted to be a little greedy and ask you about maybe '27 and the other 3 customers or so. How is the dialogue going beyond these 4 customers? In the past, you talked about having 7 now added a fourth of production. And then there were 3. Are you hearing from others. And how is the trend going maybe with the other 3, maybe beyond the '26 into '27 and beyond? How is that momentum, you think, going to shape up?

And you are definitely greedy and definitely overthinking this for me. Thank you. But yes, that's -- as asking for a subjective qualification and frankly, I don't want to give that. I'm not comfortable giving that because sometimes, we stumbled in the production in fairly -- in time frames that fairly unexpected surprisingly. Equally, it could get delayed. So I'd rather not give you any more color on prospects, then let's tell you these prospects are real prospects and continue to be very closely engaged towards developing each of their own expense with every intent going into substantial production like the fall we have today a customer.

Yes. You still think that 1 million units by goal for these 7 though, is still intact?

For the 3, I said, now therefore, that's still -- only for the customers, for the prospects, no comments, I'm in no position to judge on that. But for our 3, 4 customers now, yes.

[Operator Instructions]. And that will come from the line of Jim Schneider with Goldman Sachs.

Hock, I was wondering if you could give us a little more color, not necessarily on the prospects, which you do have in the pipeline, but how you view the universe of additional prospects beyond the 7 customers and prospects you've already identified. Do you still see there being additional prospects that would be worthy of a custom chip. And I know you've been relatively circumspect in terms of the number of customers that are out there and the volume that they can provide and selective in terms of the opportunities you're interested in. So maybe frame for us the additional prospects as you see them beyond the 7.

That's a very good question. And let me answer it in a fairly broad basis. As I said before and perhaps repeat a bit more -- we're very -- we look at this market in 2 broad segments. The one is simply the guys, the parties, the customers who develop their own L&M. And the rest of the other market I consider is collectively lump as enterprise. That is markets that run -- that will run AI workloads for enterprise, whether it's on-prem or GPU, XPU or whatever as a service, the enterprise. We don't address that market, to be honest. We don't. That's because that's a hard market for us to address, and we are not set up to address that.

We instead address this LLM market. And as I said many times before, it's a very few narrow markets. Few players driving frontier models on consistent on a very accelerated trend towards super intelligence for one -- plagiarizing the term of someone else. But you know what I mean. And they are the guys who would invest, who need to invest a lot initially, my view on training, training ever larger and larger clusters of ever more capable accelerators. But also as for these guys, they got to be accountable to shareholders or accountable to being able to create cash flows that can sustain their path. They start to also invest in influence in a massive way to monetize these -- models. These are the players we work with. These are individually people, players who spend a lot of money on a lot of compute capacity just that they are only so few of them. And we have -- I have indicated identified 7, 4 of which now our customers, 3 continues to be prospects we engage with. And we're very picky and careful, I should say -- who qualifies under that, and I indicated they are building a platform or have a platform and are investing very much on leading LLM models. And we're ever and I think -- that's about it. We may see one more perhaps as a prospect. But again, we are very thoughtful and careful about even making that qualification. But right now, for sure, we have 7. And that, for now, it's pretty much what we have.

[Operator Instructions]. And that will come from the line of Tom O'Malley with Barclays.

I wanted to ask the on the Jericho4 commentary. NVIDIA talked about the XGS switch and us talking about scale across. You're talking about Jericho4 sounds like this market is really starting to develop. Maybe you could talk about why you see a material uplift in revenue there? And why it's important to start thinking about those type of -- as we move more towards inferencing.

Great. Well, thank you for picking that up. Yes, scale across is the near term now, right train, scale up, which is within the rent within which computing within the rank Scale Up, doing across ranks about in the data center. But now when you get to clusters that are -- I'm not 100% sure where the cutoff is, but say above 100,000 GPU or XPUs that you're talking about probably in many cases because of limitation of power shell, that the data that you don't do 1 single data center footprint site to hand to sit with over 100,000 of those XPUs in one side -- may not easily available land may not be complete -- so many -- some outcomes, most of all our customers that we see create multiple data center sites close at hand, not far away within range 100 kilometers. It's kind of the level, but be able to then put in homogenous XPUs or GPUs in this multiple location 3 or 4 and network across them so that they behave, in fact, a single cluster. That's the coolest part and that technology would require us because of distant deep buffering very intelligent congestion control is technology that exists for many, many years in the likes of the telcos of AT&T and Verizon doing network routing. Except this is for even somewhat more trickier work growth about the same. And we've been shipping that to a couple of hyperscalers over the last 2 years as Jericho3 as the scale of these clusters and the bandwidth required for AI training expands, we now launched this Jericho4, 51 terabit per second to handle more bandwidth. But same technology, we have tested, proven for the last 10, 20 years, nothing new. Don't need to create something new for that. It's running an Ethernet and very proven, very stable. And as I said, last 2 years, under Jericho3, which around 256 connections on no compute nodes. We've been selling to a couple of our hyperscale customers.

[Operator Instructions]. And that will come from the line of Karl Ackerman with BNP Paribas.

Hock, have you completely converted your top 10,000 accounts from vSphere to the entire -- Cloud Foundation virtualization stack. I asked because I think last quarter, 87% of accounts had adopted that, and that's certainly a marked increase versus less than 10% of those customers who bought the entire suite before the deal. And I guess as you address that, what interest level are you seeing with the longer tail of enterprise customers adopting BCF? And are you seeing tangible cross-selling benefits of your merchant semiconductor storage and networking business as those customers adopt VMware?

Okay. To answer your first part of the question. Yes, pretty much virtually way over 90% has bought VCF. Now I like to -- I'm careful about choice of what because we have sold them on it and they bought licenses to deploy it doesn't mean they are fully deployed. Here comes the other part of our world. which is to take these 10,000 customers or a big chunk of them who have taken more about their vision of a private cloud on-prem and working with them to enable them to deploy and operate it successfully on the infrastructure and on-prem. That's the hard work over the next 2 years that we see happening. And as we do it, we see expansion across the IT footprint on VCF private cloud running on the data set within their data center. That's a key part of it. And we see that continuing. And that's the second phase of my VMWare story.

First phase is convince its people to convert from perpetual subscription, and so doing purchase VCF. Second phase now is make that purchase, they made on VCF very create the value they look for in private cloud, on their premise, on their IT data centers. That's what's happening. And that will sustain for quite a while because on top of that, we will start selling advanced services security, disaster recovery, even AI running AI workloads on me. All that is very exciting.

Your second question is, is that able to enable me to sell more hardware? No. Well, it's quite independent. In fact, as they virtualize their data centers, we consciously accept the fact that we are commoditizing the underlying hardware in the data center, commoditizing services, commoditizing, storage, commoditizing even networking. And that's fine. And by so commoditizing were actually reducing cost of investments in hardware in data centers for enterprises.

Now beyond the largest 10,000, are we seeing a lot of success. We are seeing some. But again, 2 reasons why we do not expect it to be as successful. One is the value, the TCO, as they call it, that comes from it will be much less. But the more important thing is the skill sets that needs to not just deploy that you can get services and also to help them, but to keep all putting it might not be something that they can take on. And we shall see. This is a -- this is an area we're still learning. And it'd be interesting to see -- well, VMware has 300,000 customers. We see the top 10,000 as making has been people and makes a lot of sense, derive a lot of value in deploying private cloud using VCF. We now are looking at whether the next 20,000, 30,000 midsized companies see the same way. Stay tuned. I'll let you know.

[Operator Instructions]. And it comes on the line of CJ Muse with Cantor Fitzgerald.

I was hoping to focus on gross margins. I understand the guide down 70 bps, particularly with software lower sequentially and greater contributions from wireless and XPUs. But to hit that [ 17 ] spot 7, I either have to model semiconductor margins flat, which I think would be lower or software gross margins to 95% up 200 bps. So can you kind of help me better understand kind of the moving parts there to lap only a 70 bp drop?

Yes. I mean GPUs will be going up along with wireless, as I said on the call, and our software revenue will be coming up just a bit as well.

It means we...

XPUs, yes. Our part is typically our heaviest quarter, right, of the year for wireless. You have wireless and GPUs with generally lower margins, right? And then our software revenue coming.

[Operator Instructions]. And that will come from the line of Joe Moore with Morgan Stanley.

Great. In terms of the fourth customer, I think you've talked in the past about -- customers 4 and 5 were more hyperscale and 6 and 7 work more like the LLM makers themselves. Can you give us a sense for -- if you could help us categorize that? I thought that's fine. And then the $10 billion of orders. Can you give us a time frame on that?

Okay. Yes. No. It's towards at the end of the day, all 7 LLMs. Not all of them have a current -- has a huge platform we're talking about, but one could imagine eventually all of them will have or create a platform. So it's hard to differentiate the -- but coming back -- coming on the second and the delivery of the $10 billion, I'll probably be in -- around, I would say, the second half of our fiscal year 2026. I would say, to be even more precise, likely to be Q3 of our fiscal '26.

Okay. Q3, it starts or Q3, what time frame does it take to deploy $10 billion?

End in Q3.

[Operator Instructions]. And that will come from the line of Joshua Buchalter with TD Cowen.

I was hoping you could provide some comments on momentum for your first scale up Ethernet and how it compares with UA Link and PCIe solutions out there. how big of a -- how meaningful is it to have at the Tomahawk Ultra product out there with a lower latency? And how meaningful do you think to scale up the Ethernet opportunity could be over the next year as we think about your AI networking business?

Well, that's a good question. And we, ourselves, are thinking about that, too, because to begin with Ethernet, our Ethernet solutions are very disaggregated from the AI accelerators, anybody does. It's separate. We treat them as separate. Even though you're right, the network is a computer. We have always believed that Ethernet is open source. Anybody should be able to have choices, and we keep it separate from XPU. But the proof of the matter is, for our customers who use the XPU, we develop, we optimize our networking switches and other components that relate to being able to network signals in any cases, hand-in-hand with it. In fact, all this experience have developed with interface that handles Ethernet very, very much so. So that's in a way with XPUs with our customers, we are openly enabling Ethernet as a networking protocol of choice very, very openly. And it will not be our Ethernet switches. It could be any other but somebody else Ethernet switches that does it. It just happens to be where the lead in this business, so we get that. But beyond it, especially when it comes to a close system of GPUs, we see less of it, except in the hyperscalers, where the hyperscalers are able to architect the GPUs clusters very separate from the networking side, especially in scale out. In which case, on those hyperscalers, we sell a lot of these Ethernet switches that are scaling out. And we suspect when it goes to scaling across now even more Ethernet that are disaggregated from the GPUs that are in the place. As far as experience are concerned, for sure, it's all Ethernet.

[Operator Instructions]. That will come from the line of Christopher Rolland with Susquehanna.

Congrats on the contract extension, Hack. So yes, my questions are about competition, both on the networking side and the ASIC side. You kind of answered some of that, I think, in the last question. But do you view any competition on the ASIC side, particularly from U.S. or Asian vendors? Or do you think this is decreasing? And on the networking side, do you think UA link or PCIe even has a chance of displacing Sue in 2027 when it's expected to ramp.

Thank you for embracing Sue. I did expect that to come up. And I appreciate that -- well, you know I'm biased to be honest. But it's so obvious I can't help us being biased because Ethernet is well proven. Ethernet is so known to the engineers, the architects that sits in all these hyperscalers, developing, designing AI data center, their AI infrastructure. It's the logical thing for them to use, and they are using it, and they are focusing on it. and the development of separate individualized protocol, frankly, is beyond my imagination why they bought Ethernet is there. It's been well used. It's proven it can keep going up.

The only thing people talk about latency, especially in scaling up hence, the emergence of NVLink. And even then, as I indicated, it's not hard for us, and we are not the only one who can do that, a few others in Ethernet can do it in the switches. You can just tweak the switches to make the latency super good, better than NVLink, better than InfiniBand less than 250 nanoseconds easily. And that's what we did. So it's not that hard. And perhaps as my say that because we have been doing it as the Ethernet has been around the last 25 years at length. So it's there, the technology that's only to go and create some cooked up protocols that now you have to bring people around. Ethernet is the way we go. And we definitely have competition too because it's an open source system.

So I think Ethernet is a way to go. And for sure, in developing experience for our customers, all this experience with the agreement of customers are main compatible interface with denim and not some fancy other interface that one has to keep going as bandwidth increase. And we -- and I assure you, we have competition, which is one of the reasons why the hyperscaler is like Ethernet . It's not just us. They can find somebody else for whatever reason, they don't like us. and we're open to that. It's always good to do that. It's an open source system and there are players in that market, not any core system.

Switching on to XPU competition. Yes, we hear about competition and all that. It's just -- it's a competition that is an area that we always see conditions and our only way to secure our position is we try to invest and now innovate anybody else in this game. We've been fortunate to be the first one, creating this XPU model of ASICs on silicon. And we also have been fortunate to be probably one of the largest IP developers of semiconductor out there, things like serialize -- the serialize sets been able to develop the best packaging been able to design things that are very low power. So we just have to keep investing in it, which we do to outrun the competition in this space. And I believe we're doing a fairly decent job of doing it at this point.

And we do have time for one last question, and that will come from the line of Harsh Kumar with Piper Sandler.

Congratulations on all the exciting AI metrics, and thanks for everything you do for Broadcom and sticking around. My question is you've got 3 to 4 existing customers that are ramping as the data centers for AI clusters get bigger and bigger, it makes sense to have differentiation, efficiency, et cetera, therefore, the case for XPUs. Why should I not think that your XPU share at these 3 or 4 customers that are existing will be bigger than the GPU share in the longer term?

It will be -- it's a logical conclusion, as, you're correct. And we are seeing that step by step. As I say, it's a journey. It's a multiyear journey because it's multigenerational because -- this experience don't stay still either. I'm doing multiple versions, at least 2 versions, 2 generation versions for each of these customers we have. And with newer generation, they increase the consumption, the usage of the XPU as they gain confidence as the model improves, they deploy it even more. So that's the logical trend that XPUs will keep in these few customers around whereas they successfully deployed and their software stabilizes the software stack the libraries that sits on these chips stabilizes and proves itself out. They will have the content to keep using a higher and higher percentage of their compute footprint in their own XPUs for sure, and we see that. And that's why I think we progressively gain share.

Thank you, Hock.

Thank you. I would now like to turn the call back over to Ji Yoo, Head of Investor Relations, for any closing remarks.

Thank you, Sheri. This quarter, Broadcom will be presenting at the Goldman Sachs Communacopia and Technology Conference on Tuesday, September 9 in San Francisco, and at the JPMorgan U.S. AllStar Conference on Tuesday, September 16 in London. Broadcom currently plans to report its earnings for the fourth quarter and fiscal year 2025 after close of GIT on Thursday, December 11, 2025. We public webcast of Broadcom's earnings conference call will follow at 2:00 p.m. Pacific.

That will conclude our earnings call today. Thank you all for joining. Sheri, you may end the call.

This concludes today's program. Thank you all for participating. You may now disconnect.