DigitalOcean Holdings Stock

[Audio Gap] of Citi's TMT Conference. My name is Mark Zhang, and I help -- I'm part of the Citi Software Equity Research team. Today, we have the pleasure of hosting DigitalOcean CEO, Paddy Srinivasan; and CFO, Matt Steinfort, all both to my left. Welcome to the conference, and thank you so much for attending, gents.

Well, thank you for hosting us, Mark.

Yes, absolutely. So I think maybe let's backtrack a little bit. Paddy, it's been, call it, 3 years and change since you've taken over the helm as CEO. At the time when you joined, it's been -- DigitalOcean looked a lot different than what it was today. At the time, I think Paperspace was the big AI acquisition and the way that DigitalOcean approached AI. But since then, a lot has changed. Maybe take us through the journey of -- number one, what you saw in DigitalOcean coming into the role day 1? And how you sort of saw this as from a strategic standpoint, how you approach AI? And obviously, there's various ways to play AI and company's role in AI. So what made you take the bet of the path that you're taking on today?

Yes. Yes. So thank you, Mark. It's a great question to get started with. So I'm coming up on 3 years, not quite there yet, but in the world of AI, that it's almost an era. So when I came in, I mean, the first thing that really attracted me to DigitalOcean was I've been working on and off on developer platforms right from my early days at Microsoft for close to 30 years now. So that pedigree and the DNA of the company which was relentlessly focused on catering to the needs of developers of all sizes and shapes in different sizes of companies was really attractive.

And the second thing that was amazing about DigitalOcean, which still never ceases to surprise me is the strength of the product to be the primary customer acquisition engine. So those were the 2 things that I knew were the strengths of the company coming in. But then one of the first things we had to do was to make a decision on what kind of company we want to be in AI. So the big bet that we took in retrospect about 2 years ago was, do we want to focus on training or inference or both? And we made a very strategic decision to say that training was more on the hardware side. And our hypothesis was inference needs a lot of software. And at that time, it was not that evident, but now it is fairly common knowledge that inference is not only a lot heavier dependency on software, but inference is also a more durable workload.

It is a workload that resembles cloud patterns. It is a production workload, and it is typically consumed when you have product market fit, and it is eventually built right back to customer usage for the most part. So that is the bet that we made. And along the way, we had to make many other bets, right? So for example, we had to make the bet that software needs to evolve to serve the needs of inference in a very different way than what training was. We also made a bet that we have to nail software first and then build scale. So that is the inflection point we are in as a company. We feel very confident given that we have built our inference stack, working hand-in-hand with customers.

That's one of the luxuries we have that many other Neoclouds don't have is having a front row seat in co-development, co-innovation with leading-edge AI natives that push the boundary of inference gives us the ability and the confidence that we are building the right thing. We are catering to the needs of the most demanding AI native customers. So we feel we are on a really good path to a great software platform now we are addressing the scale or we are building scale to support the software. A lot of Neoclouds or most of the Neoclouds, I would say, went after scale first because they were focused on training and are trying to cobble together a software platform.

And the other major decision that we continue to make is I fundamentally believe that technology platforms, especially developer-oriented infrastructure platforms cannot be assembled. They have to be built from the ground up, right? Of course, you can have a couple of acquisitions here and there on small feature gaps. But fundamentally, it is really, really hard to cobble together. This is not an application stack like Salesforce. There's a reason why you don't see Azure, AWS or Google Cloud really go in an acquisition spree to build their platform, right? So developer platforms, especially in AI infrastructure have to be built from the ground up, and that's -- the platform we call as AI native cloud has 5 layers, and it is beautifully integrated into a single pane of glass, provides all the control pane required to build and scale inference applications.

So a lot of small decisions that have put us in a very good position of strength, I would say, because we have some of the frontier inference workloads running on our platform.

Sure. That's terrific. Thanks for that overview. Speaking of scale and where you are in the position you are today, it's sort of like this opportunity to really inflect going forward. The inferencing engine is seeing just incredible traction, reaching calling, 6, 000-plus logos within a matter of 6 months. Can you -- I guess, like obviously, there's the tech advantage and the software that you provide on it. But can you maybe speak to what has been the driver of the success there, whether from a product market fit, the tech differentiation, product marketing? And how -- I understand like there's many use cases here for AI and inferencing. How are your -- like AI natives using basically like properly using your products today?

Yes, it's a good question. So it has been 4 months since we launched. We went into GA. It's probably slightly less than 4 months, and we have a lot more than 6,000 customers at this point. And it goes back to our strength in product-led growth, right? We -- of the 6,000 that we announced in August, maybe there are a handful that we acquired through our direct sales motion. A vast majority, like 99% of those customers came to us by the virtue of our product lending itself to a try and then buy kind of model.

So it is just organic word-of-mouth developer adoption. And the reason why the product market fit was just jumping off the page is a couple of reasons. One, there's just a lot of latent demand for inferencing, right? The whole world -- anyone that is writing software now is building it as AI native, agent native, it needs tokens. It's very inference hungry. So that's number one.

Number two is -- there is -- so there's a difference between when we say we have inference and when a Neocloud says they have inference, it's like really apples and pineapples. They sound kind of similar, but they don't taste anything like each other, right? So when we say inferencing, it is managed inferencing, where we say, we will give you an endpoint for an API and you can define what kind of tokens you want, what is your SLA, what is your throughput? What is the quality of service you're expecting? And you leave the rest to us, how to manage the infrastructure, where to deploy it, how to load balance it? All of that stuff, what kind of hardware or accelerator it is running on, that is our headache. You don't worry about it. You want certain quality of service and certain type of tokens, we will deliver that to you 24/7.

When you talk about running inferencing on a Neocloud infrastructure, it is basically saying, okay, I need -- I'll give you a certain class of accelerators like MI-355 or GB300. And then you take over from there. You install the drivers, you build the development pipeline. You install the models and you manage the life cycle of the whole thing. So both will work, but for you to take a piece of hardware and run inferencing at scale, it takes a significant amount of heavy lifting that I believe only a handful of companies can do at this point especially when you start talking about advanced reasoning models like Kimi K3, GLM 5.3, these are near frontier open weight models. K3 has 2.8 trillion parameters, of which 200 billion are active nodes. So you have to load up the model and keep 200 billion parameters in memory and the KV cache, it's a 1 million context window.

So you need to be able to load the whole thing, have KV cache for 1 million context window and manage it and manage the throughput and do the offloading of the cash from memory to disk and disk back to memory, it is not easy. That's why when you go to open router, you look at -- all the inference providers, you will see dozens and dozens of them, 90% of them are serving flash models, which are very, very small models, which are great for some use cases, but not great for near frontier tasks.

So when you start getting into the near frontier type of model performance, it the general industry trend, you asked about use cases more. The use cases are becoming more and more sophisticated, right? To serve those, you need real frontier reasoning models like Fable and Astra and K3 and GLM 5.3. And a vast majority of AI native workloads prefer to have managed inferencing rather than just raw metal infrastructure where they have to manage the whole thing. So when we talk about the inference demand is unbounded, what we mean is the managed inference demand, like the demand for consuming tokens, not demand for consuming actual infrastructure.

So it's very different when we talk about our inference platform. It comprises of many, many different individual building blocks, not just a token delivery service, but there are a lot of other services that make up our inference engine.

Got you. Got you. And I know you speak to the demand coming from more sophisticated workloads and more specifically test. I know obviously, like coding is a very sort of like top of mind test that we always think about when we think about AI. But anything sort of like cutting edge and emerging like test that you see AI native performing that could be basically getting to an inflection point? Or anything else that gives you conviction of long-term durable demand.

Yes. So coding is the obvious one that we are seeing like coding today versus coding 12 months ago, completely different, it feels like, right, from a practitioner point of view. But we are starting to see a lot of different adoption in use cases like generative media, very advanced use cases in real enterprise companies, where we have a lot of AI native customers that offer generative media technology to their customers. And the end customer is typically a large enterprise or digital agencies.

So the way you build digital campaigns, the way you build digital assets from ads to even full feature movies to anything in between, like even Hollywood production support, all of these things are transforming. And we see a significant amount of demand and pull for generative media models. We are starting to see a lot of our customers build go-to-market applications, which are agent native. So we are seeing a lot of companies starting to deliver contact center software that is agent first, right? Customer support software that is agent first. We are starting to see gaming companies that are becoming AI-centric or AI native. So we are starting to see many of these other, I call them like capital A agent software.

So when we talk about agents, it's like a coding task has multiple subagents that are doing autonomous tasks. That's, I call it, like small A agents. And then you have capital A agents, which are labor replacement workflows, right? Multi-day, multi-turn, advanced reasoning. So you're starting to see agents that are replacing significant amount of human labor in many walks of life. So we are starting to see an emergence of multitude of use cases, personal productivity. Like it seems like earlier this week was news from Meta a couple of weeks ago, it was Grok Bot.

And then in between, there's another start-up that is getting a lot of attention called Instinct. So we're starting to see personal productivity agents really exploding. So there are at least half a dozen -- repeat workloads we are seeing. When I say repeat, I mean, there are multiple customers running their workloads on us on some of these micro verticals.

Interesting, very compelling use cases that are emerging. I think just given that, call it, there's a various multitude of products and features on the inferencing engine, how are customers -- can you maybe just walk through the life cycle of -- from the beginning to sort of like at production, how are customers using inferencing engine from a feature module standpoint?

Yes. It's a great question. So what we are observing just to -- and obviously, this is just a canonical example because this is what I'm distilling from watching a lot of AI native companies. When they start, almost everyone says, okay, I need to find out serious companies. They want to find out whether they have product market fit, right? The easiest way to do that is they pick either the latest Anthropic model or an OpenAI GPT model and go after finding product market fit. So that's typically how companies started because even 6 months ago, open rate models were 6 to 9 months behind the frontier models.

Today, the open weight models, the near frontier models are barely 2 to 3 months behind the open rate models. So what I'm observing is a lot of companies as they get product market fit, a couple of things happen. One is they start observing the cost of goods sold line item, which is the token consumption just starts getting out of control. So that's number one.

Number two is there is a lot of awareness in the market now that if you have an open weight model, it is a lot easier for you as an AI native company and you as an enterprise that is adopting that product to have an opportunity to own your intelligence, own your rate leading to own your intelligence, right? So for these 2 reasons, I see the equation flip in terms of token consumption from closed source models to open weight models. So I think we talked about the fact that we flipped literally from 25/75 to 75/25 in favor of open weight models, and it's only accelerating, right? And this is not just us industry trends. You can look at any of the AI aggregators, you'll see similar statistics.

So in terms of a typical life cycle, so they start moving to open weight models. And then one of the first things that happen is they want to take the open weight model like a Kimi K3, and there are famous case studies like Cursor, it is famously on Kimi, right? DoorDash is on Kimi. So what they typically do is they do fine-tuning to make the model more domain-specific. And then they do reinforcement learning, which is a fancy way of saying, I can fine-tune the model to be more specific to -- we are here in New York, very Wall Street specific. But then reinforcement learning makes it even more customer-specific, so I can train it on the usage patterns of Citi every night.

So the next morning, the pipeline works in a way that after the reinforcement learning, they do real-time agent eval to make sure that the model passes the val test. So the performance for those use cases that it has been reinforced to learn becomes much better the next morning and then deploy that. We are also starting to see -- I mean, these are all modules that they leverage from inference providers. So the other thing they do is we are starting to see many workflows use a mixture of models.

So for certain speech to text, for example, a certain class of models may be best price for performance versus reserving a heavy-duty model like K3 for expensive but high reasoning tasks, for example. So they use model routing. We introduced another feature called model synthesis, which is for the same task, you can have multiple models work on it real time and in just a few milliseconds return back the answer and our model synthesis module will stitch together the answer. And the user will not even know whether some parts of it came from K3, other parts came from a flash model, right?

So we do that to preserve the cost performance equation and it is all controllable by the builder of the AI product, and then we serve the answer. And then comes -- once you start serving it, then you need to have the ability to monitor and manage the whole life cycle of a model deployment. So these are all different aspects of the token consumption through our inference engine. And then comes the agent aspect of it. So last week, we announced an agent platform, which breaks down the -- an agent is an autonomous piece of code that does work, right? So typically, these agents are short-lived, but they perform tasks over multiple hours or even multiple days, but in short births of a few milliseconds at a time.

So we introduced a new cloud primitive called Sandboxes, which are compute cycles that are switched on and off in a matter of a couple of hundred milliseconds. And the agent performs a task, goes down, wakes up a game, has all the memory, does a little bit more work and then so on and so forth. And then the ability to have a swarm of agents all orchestrating and working together to accomplish that. So these are all the capabilities of our platform.

Yes, that's a very comprehensive platform from where you were 3 years ago. But can you maybe speak to the adoption rate and penetration of this sort of comprehensive platform, whether it's from model routing to agent management, what's the penetration? And what's the sort of product adoption? And what's the opportunity here going forward?

Yes. So as you go more and more away from the bare metal infrastructure, we see the adoption of the primary front door to the platform is now moving slowly, but very, very clearly moving away from, hey, can you just give me a GPU to, hey, I want tokens of this flavor becomes the front door. But once they come to our platform for tokens, then they start expanding to all of the features that I talked about. They start building agents, deploying agents. Once you deploy agents, you have to monitor them. Agents are very data hungry. So you need storage, you need databases to persist.

So then it starts proliferating. In the August earnings call, Matt and I talked about the fact that our 100,000-plus AI customers are attaching core cloud at a 70% clip. So pretty much everyone that comes for AI is attaching core cloud. In another way of saying it, AI is becoming a demand gen engine for our cloud services. And that is even before we announced our agent platform. So that's only going to become more and more acute. So from a unit economics point of view, the higher up the stack you go, it ceases to be GPU economics and it starts looking more like software economics, right? So we feel like we are just starting to scratch the surface.

As these AI native applications become more and more sophisticated, I think what is the origin of tokens? Or where does the token demand come from? It comes from agents. It comes from modern applications and applications need to be executed somewhere. They need to be -- they need to orchestrate. They need to be fed data. They need to persist information. So you need storage and databases. So we are starting to see that flywheel take effect where companies come for tokens, expand into agents, agents are data hungry and agents need to be orchestrated, so you need CPUs. And the more agents you have, the more token hungry they become, so they consume more tokens, and that's the flywheel we described in the last earnings.

Yes. No, absolutely. It's very encouraging to see the momentum of that flywheel. Now maybe like just a quick one on the flywheel aspect. How sticky does a customer become once they're on this flywheel? And sort of like what's the -- maybe like ARR or revenue uplift that you see from a customer that's bought into this flywheel of, hey, like I came in for AI. Now I'm buying your core cloud. What's sort of the upsell opportunity here?

You just look at the ARR per megawatt that we generate today, it's a function of the blend of services we have. We have a small amount of bare metal, right? It's -- I think at the last earnings, it was about 15% of our AI customer ARR. Most of our AI customer ARR is inference services and core cloud pull-through. And so when you think of, okay, what do you see in the industry on a total ARR per total megawatt from the Neocloud, it's still in the high single digits and maybe it's getting to like 10 or 11. They may announce deals at higher rates. That's on newer technology.

But if you look at it on an embedded base, it's $9 million to $10 million -- if you look at us, we're generating in Q2, it was $22 million per megawatt. And on an incremental basis, it's 30%, 40% higher than what you're seeing out of the Neoclouds. That's all because of the higher value-added services, the inference services, the core cloud pull-through that are both stickier and higher margin. So we think that there's a tremendous amount of upside there, particularly as the inference engine is really only a couple of months old, and we're just starting to see the economics there. And the big lever that I don't think the market fully grasps yet is when you sell inference services and you sell tokens, it becomes a price and yield optimization game.

It's not a, hey, did you get $4 per hour on a GPU versus $3 or did you sign a long-term contract or a short-term contract to take advantage of surge kind of capacity pricing. it's all about how many tokens can you generate from the same amount of infrastructure. And the more you can generate, the better you can deliver those tokens, the higher you can drive that price. So without adding incremental megawatts, there's still a price lever that enables you to drive up your ARR per megawatt quite a bit.

Got you. Got you. That's terrific. And I think related to that question or to that thought, what do you see as the greatest torque here? Obviously, like you mentioned a lot of just token consumption and pricing. Where do you see most of the torque? Is it going up the inferencing stack or selling more core compute? What's sort of the thoughts?

Yes, there are 3 main drivers. One is the technologies that are coming out, the latest generations of GPUs generate more tokens per megawatt. They're more expensive. So the CapEx per megawatt is higher. But the return that you can generate on that CapEx is very similar. So you're generating more tokens potential for the same amount of megawatts. So that's kind of the first driver. The second driver, as you said, is the more you can get core cloud attached and you can get higher layer services and not sell bare metal that has a lift in ARR per megawatt as well, which is material. And we're still in very early stages of that. And then the third is the token optimization that I was describing. Because if you think about it, if you sell a GPU on a long-term contract, bare metal or even GPU as a Service, someone's buying 100% of that GPU.

So they're paying you some number of dollars per hour. That's what everybody quotes in the industry, but they're using 100% of that box, whether they use it or not, they're paying for it. When you switch to selling by token, they're only paying for what they consume. And so you'd say, okay, well, I can charge them a lot more, which is good. So you get the price is materially higher but you're only using the infrastructure to provide tokens when people are consuming them, which right now is largely kind of North American business hours, right? So you assume you're only getting 60%, something like that utilization, but you're charging enough more that it's better than selling it directly as a GPU or GPU as a service.

So then the art is in, well, how do you drive that utilization up? How do you get that 60% to 70% to 80%. And for that, you need to do things that are creative. You need to pull traffic in from other parts of the world. to offset the time zone kind of differences. You can do things like batch inferencing, where people are running agents overnight to scrape all the news and to prepare kind of summary packages for them. You can schedule that stuff in the off hours. And so it becomes a -- it's like airline seating. It's a price and utilization optimization game, which is a very different muscle to have when you're in this space. And what your business is built on is I sell long-term contracts to a handful of customers and they just use it. That's a very different muscle than having been a consumption-based cloud for over a decade, always constantly thinking about how do you maximize consumption-based utilization on a fixed set of infrastructure.

Got you. Got you. No, that all makes sense. And I think just to round out this topic, putting this all together, where can we see ARR per megawatt get to from this '22 that we are currently at today?

Yes. We don't guide to it. What I would say is all of the things I just described are positive and should be incremental to the ARR per megawatt that we've been able to generate. And so that 13 that we had said on an incremental basis was based on Q4 of '25 numbers, and it was based on the mix of inference services in bare metal and Core Cloud at the time. And since then, we've -- those percentages have improved. The services we've launched are giving us additional levers. Pricing has increased. But just like the older generations of either H100, H200 pricing is going up, not down. So we're very optimistic about our ability to continually drive that yield higher.

Got you. Got you. And then I think we obviously went through the software side of the growth story. Maybe we'll move to the hardware side of the growth story. You guys are 4x your capacity within a manner of, call it, 3 years with, call it, 80 megawatts of incremental capacity coming on through '28. Where does the greatest execution risk lie here? And how do you -- how are you thinking of in terms of sourcing hardware, getting the rack space, getting yourself into data centers? Speak to us about the process and the risk of going forward.

I think we've been able to successfully navigate -- it's a very competitive market, and there's a lot of kind of interest in data center space and GPU capacity. I think we've done a very good job focusing on working with Tier 1 data center providers, which is a little bit different than some of the other approaches in the industry. We work with the Equinixes and databanks and QTS and Tier 1 data center providers that are -- have been building and operating high-quality data centers for a long period of time. that's enabled us to turn up our data centers on time and even ahead of schedule, all 3 of the data centers we turned on in '26. We're on time or ahead of schedule. And that's been going well.

We're also sourcing incremental space pretty effectively. Again, we're a little bit unique from a customer standpoint for the data center providers. The Tier 1 -- the top of the Tier 1 data center providers can sell to hyperscalers or investment-grade customers all day long. There's tons of demand for that. They're not as interested in selling to the Neoclouds. They have a very different credit profile and risk profile and customer concentration. When they look at us, they see someone that's like, okay, I can sell to a hyperscaler, but I know what yield I'm going to get. I know what kind of terms I'm going to get they can sell to us, or slightly -- we're not investment grade likely at this point, but we're not that far off. And they see it and get a better yield from these guys.

They've got a massively diverse set of customers and it's another way of playing the AI trade that doesn't carry the risk that they might carry with some other folks. And so we've been pretty successful about taking down incremental capacity. From a GPU standpoint, we have tremendous relationships with both NVIDIA and AMD, and that's been a great kind of tailwind for us as much as anything.

Got you. And Matt, maybe can you just quickly run through some of the P&L impacts of the upcoming -- the build-out from a gross margin standpoint, free cash flow standpoint, when should we see sort of metrics begin to trough and we inflect back up? Speak to some of the timing dynamics there.

Yes. And we put out a supplement at one point to try to explain this, I think it was earlier this year. When you add capacity, it has pressure on gross margins and EBITDA margin to a lesser degree, because when you bring on a new data center, you get hit with the lease expense right away, you take the equipment, particularly we tend to finance our equipment, we pay for it over time, take the depreciation as soon as it's shipped to you. So there's a little bit of headwind on the front end. But the ramp is pretty good from a revenue and an EBITDA standpoint, and the operating leverage that we're generating because we're not adding people or OpEx at the same rate that we're growing revenue.

Our operating margins, our adjusted operating margins have been still really good. It was 24% adjusted operating margin in Q2. And we've guided to this year, we're generating cash. And so we feel pretty good about that. But to your question, will that -- when will that trough? Well, it depends on when you slow down growth. If you continue to grow, and we've guided to 35% plus by the end of this year and north of 50% plus next year, you're going to continue to bring on data centers, you're going to continue to scale, and you're going to continue to have some margin pressure and from our perspective, it's our job to take capital and earn a good return on that. And if we can continue to earn the kinds of returns that we're earning, then we should be investing, and that's the mode that we're in right now.

Got you. Got you. Last question to wrap up the session. As you reflect on the past 6 months, what are some of the important milestones that strengthened your confidence and conviction in the long-term opportunity. Maybe Matt will just start from the finance function and Paddy, wrap us up on the just strategic operations.

Yes, it's a great question. We get this a lot. People ask us, why didn't you raise your 50% plus guidance for 2027. And we'll provide more update on our outlook, probably at next earnings. But if you think of the things that are better now than then, we've signed 9-figure deals. We've taken on more data center capacity. We've launched an inference engine that gives us pricing leverage we've indicated we're going to exit this year at a higher growth rate than we were. So all of that gives us a ton of confidence in our long-term outlook.

Yes. And from my perspective, Mark, I spend my time predominantly on only 2 things. Are we building the right thing? Are we getting the right customer, right? Are we building the right thing to build a durable moat, build a durable business. And I think we absolutely are. We have a lead in software. I would say you can put us up against anyone including the hyperscalers. And I think our software platform on inferencing on agentic execution is second to none. So I feel really good about it.

Obviously, we have a lot of work to do. We have a summit coming up on October 13. We'll be taking the covers off of even more innovation at that point in time. Second thing is, are we attracting and attracting durable customers that are expanding on us. I think the numbers are proving themselves out in terms of or the growth of million-dollar customers, lack of churn in those cohorts. Those customers are attaching a lot of core services and that flywheel is spinning really hard. So if we are able to do both those things in a consistent, persistent manner, I think we are in the process of building a very, very valuable business.

Terrific. I think that's a great place to cap it off. Thank you so much, gents.

Thank you, Martin. Appreciate it.

All right. We will go ahead and kick it off. Really delighted to be here at the opening company session Day 1 Goldman Sachs Communacopia. I'm Gabriela Borges. I cover software here at Goldman. My colleague, Maura Hager, is on the stage with me as well. Delighted to have Paddy and Matt, CEO and CFO, of DigitalOcean. Thank you so much for being here.

It's wonderful to be here. It's a wonderful way to what I call start the sprint to finish the year.

Paddy, I want to rewind back to when you first came in as CEO. And at the time, the DigitalOcean strategy in AI hinged on an asset called Paperspace, which was acquired a few months before you joined the team. And at the time, the industry feedback on Paperspace was a little bit mixed. And I fast forward to today and the business that you've built on what was originally an acquisition along with the core IP of DigitalOcean is really incredible. So maybe just walk us through that. How did you go from arriving at DigitalOcean and seeing the Paperspace asset and then building it into what you have today, which is much more holistic, which is much more deep from a technology standpoint?

Thank you, Gabriela. That's a great question to set us up here. So I think the first thing we had to figure out was, what role did we want to play as an AI infrastructure provider, right? So I think the first order decision was to figure out -- in the AI infrastructure space, there were 2 broad categories. One was training, one was inference. And we made a bet, which at that time, a lot of people squinted at that decision saying, "Okay, we don't want to go after the training space. We want to go after the inference," which at that time was a little perplexing.

But in hindsight, the reason why we made the decision was, number one, self-perfective, like what are we good at? We are really good at understanding developers. We are really good at building platforms. We are really good at managing global scale infrastructure for production workloads. So that was a big part of it.

The second was inferencing. We believed back then and now everyone believes that it is the more durable workload. It is the workload that companies eventually come to when they start making money. It is the workload that the end customer is paying for, for the most part versus the VCs or your investor money. So that was the biggest decision we had to make.

And then there were a lot of other decisions. Number one, we were fortunate to have an incredible talent density for building platforms. That -- and over the last 18 months, we have added to that talent pool in a big, big way. So that's one. The second is, we have a phenomenal luxury of direct customer interaction and direct customer feedback. Because when you're building a platform, it's really hard to build it in a lab or build it with 4 or 5 very deep customers.

Because usually, that kind of takes you in a way that doesn't lend itself to building a broad platform. So having the luxury of now 680,000 customers is a luxury that not many companies have, right? And then you fast forward to where we are right now, I think there is -- if you take a step back and think about the platform that we have built, we actually made a couple of other important decisions.

One is, we decided to build for the most part. And we had a couple of tuck-in acquisitions here and there. But for the most part, we built a platform because my fundamental belief is platforms cannot be stitched together. You cannot assemble. This is not an application portfolio like Salesforce. There is a reason why Azure, Google Cloud, AWS did not have a lot of bolt-on acquisitions. Platforms, by definition, need to be built from the ground up and needs to be integrated top to bottom, right? So that was one.

The -- another important decision was the order of operations. The sequencing really matters. We built software first. Now we are adding scale, right? A lot of companies went for scale first and now are building software. We'll see where we all end up, but we like our chances and our order of operations.

Finally, I would say, if you take a step back, Cloud 1.0 was built to cater to applications that were built, deployed and managed by humans for the most part, right? Even the applications were servicing humans. But now the cloud that we need to build caters to applications that are built by agents. Agents are deploying these applications. Agents are monitoring and observing it. And the cloud needs to be built for agents versus humans. And we call that the AI-native cloud. So it's -- that's how I would summarize the last 3 years of our journey.

Let me pick your brain for a couple of questions here on the health of the inference market. We get questions where folks will look at coding. And obviously, coding has been one of the big agentic use cases, maybe customer experience to a lesser extent. And folks will say, "Well, where does it go from here?" So give us some insight, what are the types of things that customers are building? And you've commented a little bit on, well, we're actually starting to see real monetization versus just VC subsidies that are burning credit. Talk a little bit about that dynamic.

Yes. So when you look at our customers and what they're doing, of course, coding is a big part of the whole inference ecosystem for a number of reasons, right? It's very structured. You can -- there is a huge corpus of ground truth data you can feed into models. So there's a lot of reasons why coding has really taken off. And coding is also the fundamental building block for many other things where you can actually build PowerPoint slides or you can build interactive applications using coding as a building block.

So there's no -- relatively no surprise there, and we have a lot of customers that do that as well. But if you look at some of the other emerging micro verticals, generative media, not just from a model perspective, but there are a lot of companies that are reinventing how digital ads are produced, inventing even like full-length feature films, changing the workflows of movie production. So there's a lot of action there. It is also with OpenClaw and Hermes and other agent harnesses personal productivity is seeing a Grok Bot.

I was very pleasantly surprised. I've been using it for the last 10 days. It's an amazing product. So all these personal productivity harnesses, and now I just heard about this company called Instinct.

Instinct, over the weekend.

Yes, yes. So there are a lot of these personal productivity agent harnesses that are taking shape. I would say, we are also just starting to see the go-to-market workflows getting reshaped, right? Like, for example, the one -- very famous one is customer outreach and demand gen. And we ourselves are piloting a few different things, customer experience and contact center, which is my old space. Obviously, a lot of repetitive work. So we are starting to see a lot of these things.

But if you take a step back and think about what our customers are doing, most customers start their journey with closed-source models, right? Because you need to understand whether you have a product market fit. So the best way to do that is, "Hey, give me the most expensive, most advanced models. Let me prove that I have a business."

And then once you are kind of sensing and smelling that product market fit, typically 2 things happen, right? One, you start -- you have a real CFO and you start looking at the cost of goods sold and you're like, "Wait a second, if we -- the more we scale, the more this business model doesn't make sense." So they start looking at open-weight models. The second thing is then you start thinking about, "Hey, are we just one feature update from being completely disintermediated by the closed-source models?" So the whole concept of owning your intelligence comes into play.

So companies that are getting to the post-product market fit are using more and more open-weight models, and especially as the near frontier space gets pushed, like K3 was seminal in how advanced it was when it came out. A couple of weeks ago, we saw GLM-5.3, Qwen 3.8, and the list goes on and on. I think the distance between absolute frontier and near frontier is collapsing every week. So we are starting to see companies move more and more towards that. And these companies are also starting to create more and more agentic workflows. So that's just starting.

So if I look at this from my vantage point as an AI infrastructure provider, there are 2 slip streams. One, you have to be in the token flow or you have to be in the agent flow. And we are lucky in the sense that we are in both token flow and agent flow from a value creation perspective. Surge pricing or scarcity-based GPU pricing is not a slip stream. It is a temporary spike. I mean we are playing in that arena as well. But I think durable slip streams are token flow and agent flow, and we are in the middle of both of them.

Maybe just explain -- that's a really interesting concept. What is the difference between token flow and agent flow? What does each flow look like?

Yes. So token flow, for example, is when companies start consuming -- when they go into full-fledged inferencing, right? When they go into full-fledged inferencing, what do they need? They need to be able to take an open-weight model, for example, and they need to do post training. Post training has a number of different techniques. They have supervised fine-tuning. You also have reinforcement learning, which is basically giving it the ability to learn from actual user interaction, and there are companies that are actually doing it in a continuous loop every night.

They look at how their users interacted with their application and the model during the daytime, and then you have some ground truthing that happens and then you feed it back into the model. So you improve the model overnight and you redeploy in the morning based on some agent evaluations, right? So that's one example of post training. So your tokens on Tuesday morning are of higher quality than the tokens you've got on Monday morning, right?

So I'm just vastly simplifying this, but that is why not all tokens are generated equally. There is a quality aspect of tokens that -- it is very easy to do small-scale inferencing with flash models at a very low scale, it's very simple. But the complexity is exponential when you start talking about 2.8 trillion parameter models like K3, just standing it up is a beast. And then you need to think about the token generation from a cash management perspective.

Then you have to think about many other things like quantization and things like that to make sure that you are providing the best cost performance with acceptable quality from a token perspective, right? So these are all elements of what a true token flow business looks like. And also from an economic value capture point of view for a provider like us, not all tokens monetize the same way. There are -- the advanced reasoning models like K3 or GLM-5.3 monetized at a completely different rate versus a DeepSeek Flash.

DeepSeek Flash is great for hobby projects or projects to just find your product market fit or just get going. But then if you want to actually productionize and get to the other side with complex reasoning multi-turn tasks, you most certainly want to look into some near frontier models in the open-weight category.

Agent flow, on the other hand, is how do you build and scale agentic workflows, right? So right now, this is another case for the Cloud 1.0 being completely, I don't want to say useless, but it needs to be completely reimagined for agentic workflows because agents are very ephemeral, but they need persistent memory, right? Agents are very short-lived. So last week, we announced a new product called agent harness, Open Harness Runtime.

And this Open Harness Runtime enables customers to bring any harness, whether it is Hermes or Codex or OpenClaw, any kind of harness into our platform and we will take care of all the infrastructure behind it, like to run the actual agent in a secure sandbox, to do observability, to do the life cycle management of this agent. It's all done by us seamlessly. The reason why that is important is the agents can run in virtual machines, but it is very, very inefficient.

Virtual machines take typically multiple minutes to hydrate and dehydrate, while sandboxes can hydrate in hundreds of milliseconds and rehydrate in less than 100 milliseconds. So it's instantaneous. And you typically only pay for what you're consuming from a CPU cycle perspective. And when an agent sleeps and awakes, it has persistent memory, right?

So when you look at many of the -- like, for example, let's say, you're automating and agentifying an SDR outreach. Some of these agents take multiple days for it to complete a task. And you have multiple cycles of hydration, dehydration happening and the agent persists memory across these things. And then, of course, you need to have the ability to add security, unique identity management, right? You need to bring in providers like Okta or someone to make sure that your agents have persistent identity. So this is what I mean by agent flow, right?

And the interesting thing is, from our perspective, again, I'll bring it back. That's why most of you are here is to understand it from our perspective, the higher up in the stack you go, the more elevation you gain from like raw bare metal kind of infrastructure, the more you go from GPU economics to software economics. So the more our customers consume our agent runtimes, flash storage or databases to manage persistent state and things like that, the more it starts looking like software economics and not GPU economics.

Perfect opportunity to bring Maura and Matt into the conversation to put some numbers around that.

Yes. Let's talk about this progression from bare metal GPU to more of the managed services and tokenomics. You're at around 15% of bare metal AI revenue and the rest of the 85% is these higher-value services. Can you talk to us about the unit economics of these higher-value services relative to the bare metal 15%?

Yes. The balance, the 85% is made up of inference services, which is basically all of the token economics that Paddy was describing. It's kind of the reserved instances, spot instances where we layer on the orchestration and the Kubernetes and all things that you layer on from a software perspective, but it is also the pull-through of the core cloud.

So if you start from the highest margin, core cloud has been around for a long time. You knew what our margins were before we launched this AI kind of endeavor. You were talking about 70% gross margins, right, and very, very valuable and sticky relative to the services just being bare metal. In between, you have token economics, which if you think of that, it changes the game entirely. If somebody is going to contract a bare metal GPU contract for 5 years. They know what the price is. They know what the terms are. They know what they paid for it. They know what that yield is. There's really no upside to that, right?

And the upside is -- upside or downside, depending on how you look at it, is on renewal. What happens at renewal is a big deal. For us, every day is an opportunity for us to drive higher pricing and higher margins on those services because we've abstracted the delivery of the value, which is tokens from the underlying infrastructure. So the more efficient we can be, the more tokens we can produce out of the same infrastructure, the more we can dial in and optimize the model as Paddy was talking about.

And every single model that comes out requires different optimizations, and it's a daily game. You're constantly trying to increase your utilization. The more you can think about, well, what do I do in the off hours if I'm primarily serving tokens and most of that demand is North American business hours, what am I doing to monetize the value of that GPU in the off hours by looking at Asia, looking at other traffic sources, looking for other ways to either batch the inferencing.

So it becomes a price optimization and utilization game. So it's a very, very different model with very, very different muscles that are required. And we've been in the consumption-based -- 100% consumption-based business for close to 15 years. Like that's what we do is optimize platforms to drive the most utilization as we can. That 100% translates into the economics because we can drive materially higher ARR per megawatt out of that infrastructure by selling software margin services, but also by increasing the throughput on the platform that we're delivering beyond just, hey, it's X hours -- X dollars per hour over a fixed period of time.

And you recently increased certain GPU list prices by around 30%. How should investors think about the difference between the kind of supply-demand imbalance that we're in versus the actual software differentiation driving those pricing increases?

Yes. I would say that -- and I know that a lot of investors picked up that 30%. The nature of our business model, we have -- very little of our revenue is under long-term contract, which means we can adjust pricing on a daily or a monthly basis. We have short-term contracts. Upon renewal, we increase prices. We're constantly managing and optimizing the price optimization. And that's not something that we have to wait 3, 5 years to do because we're under long-term contracts. So it's inherent in what we do and how we price our services.

So we've been able to increase our pricing on customers. Even some of the tiny bit of bare metal because they were on short-term contracts. We're just increasing their prices or we're migrating them off of that onto our inference services. So we have the ability to turn more dials than I think a lot of folks do in the industry. And as Paddy said, once you start getting into the token economics, it's not a supply-demand, how many GPUs do you have? It's how many tokens can you give me at what level of quality. And we control the economics on the back end. So it's a very, very different model that we're playing.

And as we think about the supply-constrained environment that we're in, how do you go about securing incremental megawatts of capacity?

Yes, we've been really effective. People say this in sometimes maybe a pejorative way, but we don't play the gigawatt game. We play the megawatt game, right, which is we're out there looking for 10s and 20s and smaller amounts of megawatts. And the value for that is, we're dealing with Tier 1 data center operators who have been proven and are delivering these services, and this is what they do all day long. It derisks our execution, and it gives us the ability to be really confident in our ability to turn up data center capacity and focus on building software.

And you've seen our -- that in evidence, the 3 data centers we turned on this year, we turned all 3 of them on ahead of time and are doing quite well in terms of ramping those up. So as we're out looking for incremental capacity, we'll take bigger locations than we've looked at today. And when you look at the competitive environment there, we fare pretty favorably when you look at the alternatives that the data center providers have.

We have a very different credit profile. We're profitable. We generate cash. A lot of what they're looking at right now is, well, they could take more frontier model uptake, which is getting very concentrated, or they could take kind of other neocloud capacity. They are a very different credit profile than what we would have. So we've been doing quite well, I'd say, in terms of securing incremental capacity.

And in any given quarter, there's this dynamic where the capacity coming online is sort of baked into the model already. So how should we think about upside in any given quarter to kind of the metrics you laid out?

You want to take that or...

Yes, I can take it. So I think that is not really true in our case. So for most cloud providers, that may be true because the quantity is kind of limited, right? So for us, when we look at the value equation, we look at it as P times Q. I mean Q is finite, we understand going into a quarter, but we still have some levers in Q in terms of faster ramp into the machines, driving higher utilization with all the techniques that Matt talked about in terms of bin packing, follow the sun and things like that.

That's why we launched spot instances a few weeks ago. And again, there were a lot of questions. And my response is try, to get a spot instance from our farm, and I lease it back from you because it's just -- in the minute we put something on demand or spot, it's gone. And on demand, we never get it back. That's the thing. We have to call up customers and say, "Are you really using it or we could -- we would love to have it back."

Anyway, so there are -- and then storage pricing, right? I mean, everyone is doing storage pricing, that's fine. We are also doing storage pricing, that's fine. So these are all the levers we have in the Q. But our focus is almost entirely in the P area, right? I mean, P is price, right? Price of the yield that we can get from GPUs. As we get closer to the end of the year, we are doing fewer and fewer GPU-as-a-service deals, right?

We have come up with this concept internally of elastic compute. What is elastic compute? Elastic compute is a combination of different flavors of serverless inferencing, on-demand and spot. Because these 3 go hand-in-hand because of all the reasons Matt talked about, which is, we have certain token throughput that we expect during business hours North American time, during the day. And then we have the ability to have spot instances to soak up the cycles that we have from the same fleet during the off-peak hours or we can get traffic from other parts of the world to offset the token production from this fleet.

So the whole name of the game is to maximize the fleet utilization, and that drives up the yield per megawatt that we get, right? On top of this, we also have various other things that we monetize. We have our core cloud monetization. So the other thing that people miss is for every one of our new data centers, it is -- we are not just deploying GPUs, right? We are deploying a full stack AI-native cloud.

And increasingly, we talked about the fact that our AI customers are not just GPU customers, right? They are token customers. They are database customers. They are compute CPU customers because they are becoming more and more agentic. So we have the ability, the more AI traffic we drive, the more core cloud consumption we drag through and attach. So that's another lever we have. And we -- as Matt mentioned, we also have the ability to drive more short-term contract repricing to keep up with the market. Even over the weekend, there was an article that was published that H100 prices are going up again.

We saw it like 3 weeks ago, because we have a fungible on-demand fleet of H100s, and they are -- they go even before we have them available. It is almost like a real-time auctioning system. So we see all these signals, and we have so many different levers, and that's what you saw in Q2, why we beat our estimates super handily is because we are exercising all of these things. And this is part of our daily executive standup. We look at the price side of the equation really, really closely to maximize the fleet utilization. And at the end of the day, for us, it goes back to -- we fundamentally believe that software makes megawatts more valuable.

And as you're increasing the CapEx and equipment financing to fund this growth opportunity, how are you thinking about the payback period with all of these different offerings that you have among the inference and on-demand?

Yes. We've not changed our views of what an acceptable return on investment is, nor our payback. And we're very excited by the opportunities that we have in front of us to invest. You've seen this. We're, I'd say, appropriately conservative. When we underwrite a new data center, new GPU investment, we underwrite it with very conservative revenue assumptions. We assume price compression, which is -- in fact, we've been wrong, right?

Prices have been going up, but we're underwriting it with price compression. And we target paybacks that you would expect. We talked about in the 3-ish year kind of a range. What you're seeing now in the market is, the CapEx per megawatt is increasing, and it's increasing for a number of reasons. Part of it is just component costs are going up. But more importantly, the newer generations of NVIDIA and AMD gear, they have a tremendous amount of incremental token capacity, so they're more efficient per megawatt. They cost you more in CapEx per megawatt.

But the ARR per megawatt that you can generate from that is continuing to go up as well. Plus you have all of these new capabilities that Paddy described, which detach the pricing. Like, if you thought -- pick your GPU model and pick your dollar per hour, if you thought it was $2 or $3 or $4, it's a lot more than that if you can optimize and sell it as tokens, there's a lot more upside. That gives you the ability to have that upside lever on the returns to pull those returns in and those paybacks in.

So we're very encouraged and very bullish on our ability to continue to deliver really strong returns on the investments that we're making. And Paddy and I spend every day just trying to figure out how do we go faster? How do we get more capacity? How do we turn the token lever as quickly as we possibly can.

I have a couple of follow-ups here. So, talk to us about -- you've given us some indication of what the next 18 months of capacity adds look like. You've given us some commentary on megawatts and how you can secure capacity. As you think about the next 18 months, how much is based in terms of the Q part of the equation? And how much license do you have to pull in more megawatts over the next 18 months?

I'd say we've -- it was consistent with our conservative approach. We'll tell you when we've got things that are committed and we know and we've got certainty on those. But we've been super active in the marketplace. And I'd say we have -- the challenge for us is we're a profitable company. We generate cash. We've got great margins, and we've got a lot of upside, and we want to make sure that we're investing in capacity that has the similar return characteristics, right?

Right.

And so if you said, "Hey, how big could you get, how quickly?" Well, if we were going to pursue training workloads or bare metal contracts, we could get really big really quickly, but we would sacrifice some of the things that make us different. So I'd say our aspirations are to get materially larger than we are. We've got license from a -- as long as we're delivering on, I'd say, the things that make us special, and it's a software-oriented, inference oriented, I think we have the ability to drive to a materially higher capacity than we have today. And I can tell you, we've been working on capacity for the last 18 months, and we feel good about our ability to really flex that, the Q as well.

So we're all on the edge of our seats waiting for whatever you will eventually tell us about the 2027 guidance?

Yes.

You've given us some nuggets here on megawatts. You've given us some nuggets here on pricing. As we start to fine-tune our models, is there anything else that we should be thinking about as we think about the shape of 2027? And any other pieces that we should be thinking about that you're thinking about when you eventually give us the update on 2027?

Yes. And we'll provide more information on our outlook in November when we have earnings. We're not going to do that on a kind of mid-quarter. But think of all things that have changed since we gave the 50% plus guidance for 2027. One, we have some 9-figure deals to give us visibility that we didn't have when we had that. Two, we've added some incremental capacity. We added -- we announced 20 megawatts of incremental capacity that we had secured since we made that statement.

Three, we've launched the token business, and we're seeing a whole new way of monetizing the infrastructure that we do have. So the P times has now got a big lever. We've also increased the guidance and outlook for exiting this year to 35% plus. So we're already going to start a decent amount higher. So -- like a lot has gone really well relative to -- and we're turning on data center capacity on time and even ahead of schedule.

So -- and prices are going up, not down. So there's a lot, I'd say, that's embedded in that. But as you think about next year, it's one, we said 50% plus. That's a full year number. So if you start at 35% and you average 50%, what do you end at? You end at something north of 50% by the end of next year. So our goal is to take advantage of this massive opportunity that's in front of us. It's a generational opportunity. We're earning -- we've demonstrated we can earn really good and attractive returns with sticky customers that have real business models, and we're pretty bullish about our prospects for '27 and beyond.

I want to end here on a comment where Paddy, you talked about reassessing the go-to-market and using some more AI tools. At the same time, Matt, you've commented on the deals getting bigger. We remember 2 years ago, actually, it was probably 3 years ago now where DigitalOcean said, "Look, we're going to do direct sales reps. We're going to land large customers." And it was really hard to get off the ground back then. You've actually gotten it off the ground. So how does that go-to-market motion evolve from here? And where do you start running into more of the neocloud and hyperscalers?

Yes, it's a great question. So everything is moving at the speed of light, right? Our product velocity is off the charts. I was just talking to someone in the hallway. We are moving so fast and pumping out so many products, it's hard for go-to-market, honestly, to keep up, right, which is a great problem to have. So we have a new CRO now, Kevin, who came from Vercel. So he definitely speaks the AI native language. And we are we are surely reimagining our go-to-market. Our product-led growth machine is absolutely amazing. It is humming, right?

We launched our token business. Now it's probably like 120 days or something. We had 6,000, 7,000 customers already on it. And it is just a luxury that most companies don't have. So building on top of that, we are increasing our direct hand-to-hand customer acquisition strategy with the top, I don't know, 300, 500 AI-native companies. And it's really interesting that most of the companies that come to us, come to us for our software.

Yes, having capacity is an important lever, but the companies that come to us are coming to us because they can build on our software, right, not just coming to us because they can get access to GPUs and they are great at managing the infrastructure. Most of the companies that have come to us, or the ones that we are acquiring now don't want to manage infrastructure. If they want to manage infrastructure, they would go to a neocloud and get infrastructure. They're coming to us because they are on GLM-5.3 today. Tomorrow, they may be on K3.

They don't want to think about all of these things. They want to build an inference -- agent-native application for which they need a plethora of infrastructure management capabilities that will be super distracting and heavy lift for them if they were to build it from ground up. So we are fortifying our ability to go have these conversations.

We have 2 different FDE orgs now, one inside the engineering organization, sitting right next to the product development team. We have a field FDE team, which goes with our CRO and demonstrates to our customers that, "Hey, here's how you build an agent-native application." So we are trying to throw out any existing playbook. We must invent a new playbook because there aren't too many companies that have figured this out yet.

Fantastic. Please join me in thanking Paddy and Matt for their time.

Hello, everyone. Thank you for joining us, and welcome to the DigitalOcean Second Quarter 2026 Earnings Conference Call. [Operator Instructions] I will now hand the conference over to Radu Patrichi, Head of Investor Relations. Radu, please go ahead.

Thank you, and good morning. Thank you all for joining us today to review DigitalOcean's Second Quarter 2026 Results. Joining me on the call today are Paddy Srinivasan, our Chief Executive Officer; and Matt Steinfort, our Chief Financial Officer.

For those of you following along, an accompanying slide presentation is available on the webcast. Before we begin, let me remind you that certain statements made on today's call may be considered forward-looking, which reflect management's best judgment based on currently available information. Our actual results may differ materially from those projected in these forward-looking statements, including our financial outlook. I direct your attention to the risk factors contained in our SEC filings as well as those referenced in today's press release that is posted on our website.

DigitalOcean expressly disclaims any obligation or undertaking to release publicly any updates or revisions to any forward-looking statements made today. Additionally, non-GAAP financial measures will be discussed on this conference call. Reconciliations to the most comparable GAAP financial measures can be found in today's earnings press release as well as in our investor presentation that outlines the discussion on today's call. A webcast of today's call is available in the IR section of our website. And with that, I turn the call over to Patti.

Thank you, Radu. Good morning, everyone, and thank you for joining us today. We had an exceptional Q2 as we continue to accelerate growth in a disciplined way, and I'm excited to share the highlights with all of you. Let me start with 4 key takeaways from the quarter. First, our growth rate continues to accelerate. As we previewed several weeks ago, Q2 was another strong quarter for DigitalOcean. We were above guidance on every key metric. We delivered 29% year-over-year revenue growth while continuing to have strong profitability.

Second, our inference services, the collection of all non-bare metal inferencing capabilities on our AI native cloud is getting tremendous traction and grew almost 800% year-over-year. Launched in late April this year, our inference engine, which is a managed offering that includes server-less inference and related technologies is off to a flying start with over 6,000 customers, including material inference workloads from some of the most sophisticated AI native companies. Third, an AI-native flywheel is emerging, driving adoption across our full AI native cloud with a new entry point through our inference engine. We are already seeing early signs of this flywheel.

More than half of new AI customers added year-to-date have core cloud attached. We believe this flywheel will drive higher margin and stickier services, further increasing our ARR per megawatt and differentiating us from bare metal Neoclouds. And finally, we continue to focus on disciplined execution and durable growth. While we continue to manage the same supply chain challenges that face the entire industry, we are delivering our new 2026 capacity on time and in some cases, ahead of schedule. We secured an incremental 20 megawatts. We strengthened our balance sheet. We landed our first 9-figure annual commitment -- revenue commitments, and we remain focused on responsible investment and generating attractive returns.

With our meaningful progress and momentum, we are again raising our full year 2026 outlook. We now expect revenue growth of approximately 30% for the full year 2026 and to reach at least 35% growth by Q4 of 2026. While it is premature to give formal guidance for 2027, we are even more confident in our prior 2027 estimate of 50% plus revenue growth for the full year 2027.

I'll now spend a few minutes drilling into each of these 4 key takeaways. First, we delivered record Q2 revenue performance and the top line continues to accelerate with demand well in excess of capacity. Q2 revenue was $281 million, up approximately 29% year-over-year, which is more than double our growth rate in the same period last year. We delivered a record $93 million in incremental ARR in Q2, the most incremental ARR in a quarter in the company's history and nearly triple what we added in the same quarter last year. And we are doing all of this with strong profitability.

We delivered 40% adjusted EBITDA margin, 24% adjusted operating income margin and 17% trailing 12-month adjusted free cash flow margin in the quarter. We are driving this growth by continuing to deliver for our highest spending customers. ARR from $100,000-plus customers grew 98% year-over-year and our $500,000-plus customer ARR grew 160% and our $1 million-plus customer ARR rose 214%.

The higher spend the cohort has, the faster that cohort is growing, and this has been the case for 8 quarters in a row. Our highest spending cohort is also becoming a much bigger portion of our business, and a critical part of our growth engine, growing from 9% of total ARR a year ago to 23% in Q2. AI customer ARR reached $234 million, growing over 200% year-over-year.

AI customers come to DigitalOcean for more than just capacity. They come to us for software and the capabilities that help them accelerate their business. 85% of AI customer ARR in the quarter came from inference services and core cloud, not from bare metal. Inference services are the fastest-growing component of our AI customer ARR, growing close to 800% year-over-year and now represent over 70% of our total AI customer ARR. We are a full stack cloud platform with software that AI native companies depend on to build, run and scale production AI.

The second key takeaway is the growing traction of our inference engine. We launched our inference engine, which provides the right model at the right performance and price for every task as a part of our AI native cloud in late April. Since then, over 6,000 customers have leveraged the inference engine, while customer count grew an average of close to 60% month-over-month, and the token volume increased 30x over the last 60 days. We have seen open weight models climb up from around 15% of total token volume following our April launch to close to 75% today, highlighting the importance of open weight models in the AI native ecosystem.

This token growth is driven by strong demand from AI natives, not from individual users looking for a batch for the most token consumption. Tokenmaxxing was the industry's first instinct, maximize usage, throw the largest frontier model at everything and let the bill compound. As workloads shifted from human prompted to agent-driven, token consumption and cost exploded. For an AI-native company, tokens are both a source of value and COGS. So runaway costs are an existential threat to their unit economics. We believe that the market is shifting towards valuemaxxing, the right model at the right cost for every task, measured in business outcomes per dollar. This shift is a tailwind for us as we believe that value creation opportunities will expand from just whoever built the model to include whoever serves it the best.

Open weight models make valumaxxing possible. Open weights let customers post train on their own data and control their cost curve. Frontier quality open weight models at compelling cost performance characteristics have been a key adoption driver. For analyst firm artificial analysis, today's best open models trail the frontier models by only a few percentage points and are over 70% of token volume per OpenRouter, the largest and most popular AI gateway. An open weight file is necessary but not sufficient for companies to own their intelligence. Turning open weights into fast, reliable, economical production tokens is a systems problem our inference engine solves.

Continuous batching, quantization, KV cache optimization, speculative decoding, prompt caching, intelligent routing and workload-aware scheduling, all engineered as one system on infrastructure we own. Like traditional open source software, the model may be free, but making it useful and serving it well is the product. Our inference engine is much more than an API endpoint to an open weight model. It has become a full production run time solving today's most pressing needs. Our inference router optimizes requests in real time for quality, latency and cost across our full open and frontier catalog behind one unified API. Close to 1,400 inference customers actively use this feature to optimize dollars per unit of intelligence.

Model Synthesis, a new feature we just released, orchestrates a panel of models in parallel with the synthesizer merging their outputs, delivering frontier grade quality at a fraction of frontier cost. Model evaluations let customers test any model against their own business data. Batch inference handles high-volume asynchronous workloads. Prompt caching cuts cost and latency with 0 application changes. Server-side tools give agents web search, retrieval and function calling natively inside inference requests with built-in access to knowledge bases and MCP servers.

Together, these features turn model choice from a onetime decision into a dynamic ongoing engineering and business decision. On our platform, open weight models grew from roughly 15% of tokens following our initial launch to close to 75% today. And when Kimi K3, the largest open weight model ever released, went live on July 27, we were the only full stack cloud provider to be a launch partner, delivering day 0 access. Adoption has been incredible with over 400 net new customers just in the first week. Our model catalog now offers 75-plus open and closed source models through a single endpoint, including GLM-5.2, DeepSeek V4, GPT-5.6, OPUS 5, et cetera, with 14 day 0 launches since April of this year.

Our third key takeaway is that our AI native cloud is becoming a flywheel. Every layer a customer adopts pulls them into the next. In late April, we launched the DigitalOcean AI native cloud, 5 fully integrated layers from silicon to inference to agents with open source support at every layer. Since then, we shipped more than 80 releases across all layers, demonstrating innovation across the platform.

These releases included managed agent products like server-side tools, data and learning products like knowledge bases, the inference engine I just discussed and cloud primitives like our new insights observability service. An integrated full stack platform is foundational to AI builders because AI native applications require far more than raw GPUs or just tokens. They need a production cloud designed around inference and agentic execution. Building and operating that cloud is hard. It requires deep engineering across data centers, silicon, networking, storage, Kubernetes, databases, model serving, routing, evaluations, agent run times and much, much more.

Our integrated platform eliminates this complexity for customers and a flywheel is emerging as these AI builders adopt it. Customers enter the platform through one of the 3 front doors, inference, agents or core compute. Most AI native customers first need inference with the right model at the right performance and the right price for every task. From there, inference graduates into agentic workflows, which use and generate data that requires databases, storage, knowledge bases and observability. That generated data becomes raw material for learning, improving and customizing the models. Agent run times and learning drive demand for compute. And because that compute runs on infrastructure we own and operate every turn of the wheel improves our unit economics, better price performance for customers, spur even more tokens and the cycle accelerates.

Adoption in each layer drives the next and the effects compound. Inference is one entry point into a self-reinforcing cycle that pulls customers deeper into the platform and has been a leading indicator for full platform adoption. And this flywheel is already working. Let me give you some examples. OpenCode, a leading open source AI coding agent with over 7.5 million monthly active developers started by integrating with DigitalOcean Droplets to simplify agent development.

Now OpenCode is also using DigitalOcean's inference engine and AI native cloud for its inference needs, including access to leading open weight models. In addition to OpenCode, we have also integrated DigitalOcean AI native cloud into other leading coding and agent building environments like OpenClaw, Codex, Hermes and Grok Build. When developers build there, our inference engine is already in their workflows just one API call away. That opens the inference front door at ecosystem scale. Daytona, an advanced AI sandbox company, builds secure elastic sandboxes for AI-generated code and autonomous agents on DigitalOcean. This is a textbook full stack agentic workload running on our platform.

Its workloads require GPU acceleration, isolated compute environments, fast deployment, storage, networking and orchestration all working together. Vercel, a scaled agentic infrastructure platform is integrating DigitalOcean inference engine into their AI gateway to provide their customers with dedicated AI platform capabilities. Another great example of this is OpenRouter, which is both an efficient customer acquisition channel and a platform through which we can dial up or down on-demand traffic to test, learn and scale as we launch new models. We now serve more than 20 billion tokens per day on OpenRouter, up more than 330% over the last 60 days with much of that traffic being generated from agents.

These customers are examples of AI builders spinning our flywheel, and the flywheel does not stop at the first entry point. Every turn adds products to the stack we own, an integrated platform running on our own infrastructure spanning 20 global data centers. Owning the stack lowers our cost to serve, and that lower cost structure, combined with the emergence of high-quality, low-cost open weight models gives us better unit economics to serve our customers, which in turn enables us to win more customers.

For AI native, that advantage enables precisely what they value, better cost and performance on every workload, faster time to market, tight integration across inference, agents, data and compute and freedom from having to stitch together a myriad of services across vendors. This is clearly resonating with our customers as roughly 70% of AI customers having $100,000 or more ARR in Q2 have attached a core cloud product to their AI workloads, showing early evidence of this flywheel in action.

This value proposition is very differentiated in the market. Hyperscalers optimize for frontier labs and large enterprises. Neoclouds have built strong GPU rental businesses for model training and are adding software mostly through acquisitions. Assembling capabilities is not the same as building an integrated platform and customers often bear that complexity. Inference providers serve tokens well, but rent their GPUs with margins stacked on margins and leaving customers to stitch together inference, agents, data and compute.

Our approach is different. One, purpose-built AI native cloud tightly integrated from the ground up, enabling AI native to start and scale their agentic applications on our cloud. We will dive deeper into our AI native cloud at our AI Builder Summit on October 13 in San Francisco and we hope to see you all there, which brings me to our fourth and final takeaway that we remain disciplined in our execution and continue to focus on durable growth.

This discipline is evident not only in our financial performance, but also in our operational execution and in our responsible and profitable approach to growth. Driving growth approaching 30% in Q2 on a path to 50% plus next year requires focused execution. We remain on time and even a little bit ahead of our previously communicated schedule on all 3 of our new 2026 data centers. We launched our Richmond data center in Q1, our Kansas City data center in Q2, both ahead of target, and we remain on track for the second half launch of our Memphis data center.

Beyond just hitting our launch date, we've been able to allocate the majority of the capacity to specific customers or to our highly in-demand token fleet before we launch these data centers. We also secured approximately 20 megawatts of additional capacity this quarter, which is targeted to come online over the last part of 2027 and into 2028. This brings total committed capacity to approximately 155 megawatts, the majority of which will be online by the end of 2027. We continue to actively pursue additional capacity to drive further growth and meet customer demand.

Our discipline is also evident in the steps we took to strengthen our balance sheet. In July, we reduced our leverage with minimal dilution and use of cash by retiring approximately $472 million of our 2030 convertible notes, creating additional capacity to cost effectively finance our future investments. It is worth pausing on how different our profile is from many others in the AI infrastructure market.

Number one, our growth is driven by a broad set of AI native companies rather than by a handful of large bare metal offtake contracts with our top 25 customers representing only 20% of ARR in Q2. Next, our largely consumption-based model gives us the flexibility to adapt to market conditions and shift capacity to where it is most valuable. This flexibility enabled us to increase list prices on numerous GPU fleets recently by approximately 30%. Third, we are profitable with 40% adjusted EBITDA margins, 24% operating income margin and 17% last 12 months adjusted free cash flow margin.

And finally, we closely match our cash outflow with our revenue by financing equipment, efficiently funding our growth. There are very few companies with our combination of positive adjusted operating margins and projected growth of 50% plus. This is a generational opportunity, and we will go after it responsibly, building a durable business on the foundation of our differentiated software and full stack AI native cloud platform. With this momentum continuing to build, we are again raising our 2026 outlook. For the full year 2026, we now expect revenue growth of approximately 30% with an exit growth rate of 35% or more by Q4. That trajectory and the incremental committed capacity we've added both clearly strengthen our conviction in 50% or more revenue growth in 2027.

With that, I will turn it over to Matt.

Thanks, Paddy. Good morning, everyone, and thanks for joining. As Paddy shared, Q2 was an outstanding quarter. I'm excited to take you through the results, provide further context on some of the actions we have taken and provide some additional color on our updated outlook. Q2 revenue was $281 million, up 29% year-over-year, above the high end of guidance. The outperformance was broad-based, led by growth from our highest spending customers and our expanding AI customer base.

Our highest spending customers didn't just keep growing, they accelerated. ARR from our 100,000-plus customers grew 98%, up from 37% in the second quarter of last year. Our 500,000-plus customer ARR grew 160%, up from 64%. And our $1 million-plus customer ARR grew 214%, up from 92%. Each of these highest spending customer cohorts is now growing more than twice as fast as it was a year ago. We continue to gain meaningful traction with some of the most sophisticated AI natives. AI customer ARR reached $234 million, growing 212%.

And critically, 85% of that ARR is non-bare metal. This traction is evident in the material commitments we secured during the quarter, which collectively increased remaining performance obligations to $894 million, up more than 12x year-over-year with a 3.7-year average life. While changes to RPO will be lumpy, these commitments add visibility, and we expect to secure more of them in the future. They have not, however, come at the expense of our broad customer diversification as our top 25 customers represented only 20% of ARR in Q2, and this will only modestly increase as these deals ramp up.

One quick note on key financial metrics. Our business has changed dramatically over the last 2 years, with growth increasingly driven by our top customers and by emerging AI customers. Against that backdrop, net dollar retention, a strong indicator in the slow and steady growth SaaS world, has become a less useful measure of our performance. While our 102% NDR in Q2 is a 3-year high, we'll no longer highlight it as a key financial metric.

Growth today is shaped far more by our highest spending and AI customers than by the NDR trend across our 680,000-plus customer base. Profitability remained strong in Q2. Adjusted EBITDA was $114 million, and adjusted EBITDA margin of 40%. GAAP operating income was $29 million, a 10% margin, and adjusted operating income was $67 million, a 24% margin. Non-GAAP diluted net income per share was $0.45. Adjusted free cash flow in the quarter was $61 million.

Trailing 12-month adjusted free cash flow was $175 million or 17% of revenue. As Paddy highlighted, we proactively strengthened our balance sheet, reducing our leverage with effectively no dilution and minimal use of cash. In July, we equitized $472 million of our 0% 2030 convertible senior notes. The underlying shares were both already reflected in our diluted share count and were highly likely to be converted given where our stock is trading. And yet the full principal value was also reflected in our net debt, reducing our leverage capacity.

Through this proactive transaction, we retired more than half of our convertible debt 4 years ahead of maturity, reduced net leverage and did so with effectively no dilution and minimal use of cash, freeing up capacity to invest in further growth. Turning to guidance. We are raising our 2026 revenue outlook. For the third quarter of 2026, we expect revenue of $304 million to $307 million, representing 32% to 34% year-over-year growth. We project adjusted EBITDA margins of 38% to 39% and non-GAAP diluted net income per share of $0.28 to $0.30 on approximately 126.5 million weighted average fully diluted shares.

For the full year 2026, we expect revenue of $1.17 billion to $1.18 billion, representing approximately 30.5% year-over-year growth with an exit growth rate of 35% or more in Q4. We expect adjusted EBITDA margins of approximately 39%, non-GAAP diluted EPS of $1.35 to $1.40 and adjusted free cash flow margin of 11% to 13%, an increase to our prior guide. While it's premature to speak to 2027 guidance, the positive momentum we're generating and the higher projected exit growth rate give us even more confidence in our estimated 50% plus growth for the full year 2027.

Before I turn it back to Paddy, let me put our progress in perspective. Revenue grew 14% year-over-year in the second quarter of last year. In a single year, we have doubled our growth rate to 29%. We are now projecting to nearly double it again on an annual basis next year. And we are delivering this growth with attractive margins, appropriate leverage, a strong and flexible balance sheet and disciplined execution.

With that, I'll hand it back to Paddy.

Thank you, Matt. Before we move to Q&A, let me recap what we shared today. First, growth continues to accelerate, approximately 29% revenue growth, more than double the growth from a year ago, record $93 million in incremental ARR, AI customers and $1 million-plus customers each growing ARR more than 200%. We delivered this growth with strong profitability and free cash flow.

Second, our inference services are getting tremendous traction. Inference services grew nearly 800% year-over-year. Token usage on our inference engine is compounding monthly and open weight models have climbed from 15% of token traffic to close to 75%. Open weight model adoption leverages our strength, turning open models into fast, reliable, economical production tokens. Third, adoption of our inference engine is creating a growth flywheel. Inference is the entry point and every layer a customer adopts improves their token price performance and pulls them deeper into the platform.

Leading AI builders like OpenCode, Vercel and Daytona began spinning that flywheel. And because the entire cycle runs on infrastructure we own, it drives customers to higher margin and stickier products, increasing our potential ARR per megawatt. Finally, we remain disciplined in our execution, deploying planned capacity on or ahead of schedule, securing 20 megawatts of incremental capacity, delivering strong margins and strengthening the balance sheet. Our momentum and solid execution enables us to raise our 2026 outlook and positions us for strong performance in 2027.

Before I end my comments, let me connect these 4 key takeaways because the connection is the real story. Software makes megawatts more valuable. Our software attracts high-quality AI native customers with insatiable demand. Those customers adopt more of the platform than just capacity and that broader adoption increases what each megawatt earns, driving durable growth, higher margins and cash flow in future years. Strategy is becoming results and results are building momentum. Platform shifts like this come along once in a generation. Quarters like this one show that we are becoming both an enabler and a beneficiary of that shift. With that, let's open it up for questions.

[Operator Instructions] Your first question comes from the line of Gabriela Borges with Goldman Sachs.

I wanted to ask a little bit about DigitalOcean's ability to scale. Paddy, to your point, the hyperscalers are optimized for large enterprises. DigitalOcean has historically been optimized for smaller customers, but you're actually landing these larger flagship customers that have larger commitments, have larger backlog deals and require perhaps a different type of sales process, a different type of operational process.

So twofold question for you. How are you meeting those demands of the larger scaled customers? And then I think just maybe partly for Matt, how are you thinking as you scale these larger chunks of megawatts, talk to us about some of the operational puts and takes to being able to get those megawatts online at the right time and up and running.

Thank you, Gabriela. It's a great question. We feel very confident in our ability to scale, given our track record, like we have been doing this at a global scale, running a cloud business, managing a global network of data centers for the last dozen-plus years with hyperscaler SLAs and serving over 0.5 million paying customers along the way. So we feel very confident in our ability, and we are demonstrating that by bringing capacity on time and also before schedule. I always work backwards from the customers we are targeting and what they are coming to us for.

Like right now, they are coming to us not just for capacity, as I mentioned. So they are not expecting some exotic bespoke hardware or network configuration. They're predominantly coming to us because of the richness of our AI native cloud. So from a platform innovation perspective, our pace of innovation, as I described, is just staggering with over a major release every business day and sometimes multiple. And our engineering talent is absolutely world-class, and we aggressively keep adding to it.

To augment that engineering talent, we have also stood up a forward deployed engineering organization to work with some of our larger, more sophisticated customers with demanding workloads to ensure that they're getting the right price performance, throughput accuracy combination. But most of our core software doesn't have to be really customized to meet their needs. From a go-to-market point of view, we just added Kevin Van Gundy as our CRO, who comes with tremendous experience working in the digital and now AI native ecosystem.

We added Leo as our CMO, who brings a wealth of marketing experience from Google Cloud and Oracle Cloud. And they're in the process of scaling up our go-to-market muscle to help us address the next phase of our hyper growth. But this is something we feel very confident. We've been doing this for a number of years. And I'll let Matt answer the infrastructure question. But I think from a talent density perspective, both on core engineering and go-to-market, I feel really good. And we have demonstrated in the recent past, and that's why we keep talking about our $500,000 and $1 million customers and how that flywheel is spinning and has been doing it for about 8 quarters in a row now.

Yes. And I would just add to that, Gabriela, that the customers that we're dealing with, while they're bigger, these aren't your traditional kind of brick-and-mortar enterprise companies. These are very, very sophisticated technical customers where their founders and leaders are often deeply, deeply technical. And they very much appreciate the depth and the breadth of the engineering talent that we have and our ability to work with them, as Paddy said, which I think uniquely and very well positions us to be able to meet their needs.

From an infrastructure standpoint, as you've seen, we're working with some of the top data center operators in the industry that are very, very familiar with and experienced bringing up capacity. We've got a deep and talented team that works alongside of them. We have great partnerships with the leading chip manufacturers. We've got a great supply chain with a diversified set of OEMs that are all global. And we've been able to manage the implementation schedules and turn up capacity despite some of the challenges that everyone faces in the industry.

We've been able to do that on time and meet the requirements that these large customers have put in front of us. And we're very encouraged by the partnership we have with those customers. We're doing a lot of joint development already. So I think it's more than just turning up infrastructure. It's having engineers working side by side with these very sophisticated and talented customers, and we're bringing really strong talent to bear, and we're very encouraged by the progress we're making.

Your next question comes from the line of Jason Ader with William Blair.

Two questions. First, just if you could provide any specifics on the impact of pricing on the revenue growth in Q2 and then for the updated outlook? That's the first question. The second question on equipment financing, Matt, for 2026, where do you expect net leverage to be at year-end? And could you provide any specific guidance on the free cash flow for the year, including all the leases?

Yes, Jason, good questions. On the pricing, we've -- as you saw, we increased our list price on a number of GPU generations by about 30% a while ago. A lot of that pricing, we had already been, I'd say, upgrading as we -- given the short kind of contract duration for some of our customers, we had already been upgrading their prices and increasing their prices upon renewal or in some cases, pulling capacity back from a customer that we thought we have a better use for it, either the capacity in our token factory or in -- with a different customer that was willing to pay a higher price.

So all of that pricing is included in the '26 guide, and it's part of how we went from saying we're going to exit the year around 30% to now exiting it at around 35%. And it's a good setup for us in 2027 as well. On the equipment financing side, the -- we continue to get access to very attractive rates and have ample capacity to fund the growth over the committed capacity that we've taken down.

If you look at the pro forma net leverage, just take the Q2 balance sheet and just simply -- and the LTM EBITDA and simply subtract the amount of debt we retired in the equitization, puts us at 0.7x net leverage. We're in a very, very good position to stay well below that 4x net leverage that we had articulated. And in fact, it should be well below that. And that's part of why we did that. We're now sitting with an incredibly strong and flexible balance sheet. We have the ability to take on incremental equipment financing and equipment-related borrowing capacity and fuel our growth. So it was a great step for us, and our leverage is going to be very comfortably below that guideline that we had provided.

And just on the free cash flow.

Free cash flow -- sorry, Jason. Yes. No, it's great -- that's a great question. Yes, so free cash flow, as we said, would be 11% to 13% for the year. That's on an adjusted free cash flow basis, which is higher than what we had guided previously. And if you take all of the principal payments and everything, we'll still generate cash in 2026. So we expect to be free cash flow positive on any metric that you use, whether it's adjusted free cash flow or take complete cash generation and take out the principal payment. We will continue to generate cash in '26.

Your next question comes from the line of Mark Zhang with Citi.

So I wanted to actually dig in a little bit more into the 9-figure deals that you guys were able to sign this quarter. Number one, sort of wanted to get a sense of, I guess, like the inferencing and the core cloud -- these logos are committed for. And sort of what's the adoption of the other aspects of the 5-layer stack and monetization road map looks like going forward? Because I think like the go-to-market philosophy here is really looking for large deals that can make good sense that can continue to expand going forward. So I just want to get a sense of the opportunities from here as we go forward with the AI stack.

Yes. Thank you, Mark. So in terms of the larger deals and pretty much any deal that we are talking about these days, I think we had a couple of different stats that I used in the prepared remarks. Over 70% of AI customers that we added this year at any significant scale are already using some aspects of the core cloud. See some of our AI native cloud layers are still new, and that's why we have another version of our AI Builders Conference scheduled on October 13 to talk a little bit more about more specifically the managed agents layer of our platform.

So if you take a step back and think about these types of workloads landing in our platform, they typically land on one of the 3 front doors that I talked about, right? And the front door is really, really important because that's the dominant use case for which any of these sophisticated workloads are coming to us for. And immediately, they are attaching some part of our other layers of the cloud, whether it is databases or storage or orchestration.

In many cases, it's a combination of all of the above and gives us more confidence that they're coming to us not just for tokens, not just for capacity, but they're coming to us appreciating the value of the full platform because these workloads, they're not proof of concept. They are building agentic applications from the ground up. So by the nature of these agentic applications, they need far more than just GPUs or tokens. They need a place where they can do some post training. They need a place where they can store memory and context. They need a place where they can run agents in secure sandboxes.

They need a way to orchestrate these agents. So we feel increasingly confident, and that's why I spent so much time talking about the flywheel of the more we can get these AI-native workloads to consume more aspects of our platform, we feel really good about the durability of the revenue, durability of these workloads scaling up on our platform. And the early results are really, really encouraging given the attach that we are seeing on the platform.

Got it. That's very helpful. And then maybe just a quick follow-up. You also mentioned that with the new CRO, Kevin coming in, you guys are certainly in the process of scaling up the go-to-market muscles. Can you just maybe give a sense of what the early -- I guess, the early preview of what Kevin's plans are for the go-to-market organization? Should we expect more investments into sales and marketing and go-to-market for the enterprise -- at the enterprise level going forward from here?

Yes. Thanks, Mark. So the primary focus right now is to land very high-quality AI-native workloads, right, like the ones that we discussed on the call. And these are top-tier AI native companies. And as I described, just in the last 90 days for our inference engine, we've added over 6,000 customers. That is just incredible. I mean, think about it, right, 6,000 customers in 60 to 90 days. And a lot of that is still standing on the shoulders of our incredible world-class product-led growth motion.

And we are tapping into the ecosystem at scale, whether it is OpenRouter or OpenClaw or Hermes Agent, we are getting customers from all kinds of ecosystem hooks, and that will continue. In terms of very specifically the human-based sales, yes, we will fortify our enterprise AI native enterprise go-to-market motion. But again, here, it is about nailing that motion with forward deployed engineering. It is nailing that motion with enterprise sales reps that know how to go and qualify these opportunities and hold their own with very technical founding teams rather than scaling it. We will scale it eventually. But right now, it is all about quality of engagements and nailing that motion before we scale it. So in terms of investments, I don't see the investment scaling anytime soon. It is all about getting the right quality of engineering-oriented technical sales to enable us to attract and expand these AI native workloads.

[Operator Instructions] Your next question comes from the line of Wamsi Mohan with Bank of America.

I appreciate the comment that it's still a bit premature for 2027. But if we look at your performance here, which has been really strong, RPO, the timing and on time or even earlier ramp of your data centers, your comments on token usage, higher exit rate for 2026 and put all of these together, should we not assume directionally that there is further upside to 2027 than what you thought 90 days ago? Any color there would be helpful. And I have a follow-up.

Yes, Wamsi, I think that's the appropriate conclusion. The challenge for us is the revenue growth is so predicated on the specific timing of data center implementations and the turn on of capacity. And we're sitting here in August, and there's still a fair bit of moving parts in terms of the dates and times for next year. So we felt it's premature to give a specific number.

But clearly, the message is we're exiting the year a lot faster growth than what we had said we were. We've got tons of RPO and we're landing bigger customers. So we're very bullish, and we expect there to be additional upside. We're just -- it's too early to put a number on it. And so we'll wait until later this year before we provide any more specifics around that. But all the indications are we're -- as you saw by the virtue of the fact that we increased our guidance for '26 and the exit rate, we're better positioned than we were just 90 days ago. It's a good conclusion.

Okay. And then maybe, Paddy, just on the open weight models, you, I think, quoted that it's risen from roughly 15% to now nearly 75% of token volume since launch. How much of that usage is recurring production traffic versus maybe some batch inference where you have some discounts? I think you mentioned it was not batch, but I just want to make sure of that. And is the cloud core services attach any different between customers using closed versus open models?

Yes. Thanks for the question, Wamsi. It's a great question. So I'll go from the reverse order. So there isn't any major difference in what these workloads are attaching based on whether they are open weight or closed source models. They are attaching the same kind of core cloud parameters. And one pattern that we are observing is most sophisticated production workloads are now becoming a combination of open weight and closed models.

It is almost always a fusion or that's why we released this new feature called Model Synthesis, where we can actually do the heavy lifting on behalf of the customer where we run the same query in parallel across to multiple models and synthesize the results using a synthesizer rather than the customer having to stitch together these kinds of infrastructure plumbing technologies. So if you -- going back to the first part of your question, are these production workloads, absolutely yes.

I can't put an exact number on this, but you can see from the combination of the throughput, latency, accuracy that these companies are demanding, it's very easy to find out whether they are running a production workload or some internal proof of concept. And I feel a lot of the traffic we are seeing is production traffic. And as the open weight models pick up in traffic, cost is an important factor, but it is not the only factor because as you see some of the sophisticated mixture of experts models like a K3 or the about to be released Qwen 3.8, for example, these are 2.8 trillion, 2.4 trillion parameter models.

These are very big bulky models with active parameter count like 140 billion, I believe, was the K3 model. So these are not cheap models to serve. And when you look at the cost per intelligence task, yes, it is cheaper than the frontier closed source models, but they're not cheaper by an order of magnitude. But it is creating surely a Jevons paradox of the more open weight models at a reasonable cost performance that we are starting to see the adoption is just going through the roof. And as I mentioned, there's just a tremendous amount of demand that far exceeds our supply. So I feel very good about these production workloads, whether it is in coding or generative media or business workflows. These are production workloads that are scaling and they have insatiable demand on our systems.

[Operator Instructions] Your next question comes from the line of Sanjit Singh with Morgan Stanley.

I wanted to revisit the revenue per megawatt story at DigitalOcean. You guys have obviously been at a huge premium to the Neoclouds. The bare mix is obviously coming down. You guys have previously said that as the AI mix starts to increase, the revenue per megawatt will come down a bit from its current levels. Is that still the right thinking given we have the inference engine, given the success with attaching to the cloud portfolio? Where do you think -- or what are some of the levers to drive support for revenue per megawatt over time?

That's a great question. So we expect the incremental ARR that we get per megawatt to increase over time. The decline that you described is from when we were a general purpose cloud generating north of $22 million in ARR per megawatt without much AI. As we add incremental megawatts, we're adding more ARR per megawatt than our Neocloud peers because we offer higher layer services beyond just bare metal, because we sell to a broader customer base that isn't a single customer with a multiyear commitment that's going to drive pricing and margins down.

And because we offer a core cloud and CPU services that we attach to those AI workloads. So we expect that to increase as the mix of core cloud to an the attach rate increases. But we're also installing higher capacity equipment in the same megawatts going forward. So as you see the generations of NVIDIA and AMD increasing their token throughput capabilities, it gives us more revenue potential. The costs are certainly higher per megawatt as well, but the revenue potential is also higher. So we expect it to be a combination of more attach, higher and more mix of inference services beyond just the GPU as a service and the higher token capacity of the equipment we're putting in. All of those will contribute to increasing our ARR per megawatt on an incremental basis.

Your next question comes from the line of Tom Blakey with Cantor.

I think it's maybe a dovetail off of Sanjit's question. Could you just talk about maybe the pricing impact to this very strong ARR number, the net new ARR number that you reported this quarter? And then maybe give an update to the megawatt cadence that you're looking at here in calendar '26. As you mentioned you're a little bit ahead of plan, I'm just wondering if there was any details you can give us about 2Q '26 and if we're still looking for 25 megawatts in the second half.

Just to answer the latter part, we've got 15 megawatts that are left. We announced that the 10-megawatt Kansas City facility was launched already. So we have 15 left in 1 facility, and it's -- we had said it would come on in the second half, and it's on track, and we expect that to come online as we had expected over the balance of the year. And...

Yes. The first question was the pricing impact on the net new ARR.

Modest.

Yes, it's very modest in Q2. And as Matt already answered, it is baked into our guidance for the rest of the year. It's not what you might imagine right off the bat because we raised the list prices across the board for on-demand and spot instances. But as Matt mentioned previously, as some of these contracts roll out, we have been adjusting the prices to market levels for our existing contracts. But in terms of its impact in Q2 and the $93 million in net new ARR, it had very little impact on it. So I don't want the takeaway to be that that's how we had a blowout quarter. That's not the case at all.

Your next question comes from the line of Jackson Ader with KeyBanc.

I was just curious about the -- what exactly is baked into the out-year outlook? Like if I think about all the activity that you guys signed or contracted in the second quarter and the impact either here on '26 or '27, if I just think about forward guidance, is it right to think that, okay, we're at 155 megawatts, the majority online by the end of '27. And any incremental activity that happens in the next few months, like that is all incremental to the expectations for 2027? Or do you guys have certainly line of sight into a bunch of activity that's coming down the line. And so that is also factored into what you're expecting for that -- for the 2027 numbers?

That's a great question. And what you'll observe about us is we're very good, I think, at having measured and appropriately conservative outlook based on what we've already communicated in terms of capacity. And so it's a good observation that were we to add incremental capacity and were we to add incremental deals beyond what we've articulated that there would be upside.

From a '27 impact standpoint, you're getting pretty late in the year this year to have a huge impact on -- from a capacity standpoint on the calendar year '27 just because data centers typically have kind of a year-ish from lease signature to when you're generating revenue. So you're getting to the point where you might impact the exit growth rate, but the full year calendar year revenue might not be as impacted.

And that's part of why we're saying we're not going to provide formal guidance right now. There's just a lot of moving parts. But what you and what Wamsi also highlighted is, clearly, we've got a ton of momentum. We've made a great amount of progress in just a quarter. And all of that upside is not reflected in the prior estimate of 50% or more growth for next year. But it's too early for us to put a precise number on it other than, hey, we're exiting the year at a much higher growth rate. We've got a very strong RPO backlog. We're very active in the market looking for incremental capacity. So we certainly believe there's upside.

And your last question comes from the line of Radi Sultan with UBS.

Just one quick one. your customers using your AMD deployment, speak to how you see the mix between NVIDIA GPUs and inked in. And then maybe Matt, the unit economics on a per megawatt basis for AMDs compared to NVIDIA GPUs?

Yes, we have a good healthy mix of different types of accelerators in our farm. For obvious and competitive reasons, we don't get into the details of what we use to host what type of models and things like that. But it is a mix of both, and we continue to keep pace with the innovation in this market. And we certainly don't want to discuss the unit economics of different hardware throughputs. And I would just stop at that because it's a good mix of different types of accelerators. And as you can imagine, we are really good at taking whatever hardware is available based on the capacity we have and running the state-of-the-art models, right? Most of the state-of-the-art GLM-5.2 or K3, we run on all kinds of hardware. And that is the beauty of the software optimization layer that we continue to build and refine where we are almost becoming hardware agnostic.

We have reached the end of our Q&A session. I will now hand the call back to Rahu.

Great. Thank you, Paige. Thank you, everyone, for joining, and this concludes our second quarter earnings presentation and conference call. Apologies, we couldn't get to all your questions, but look forward to speaking to everyone later in the day on our follow-up calls.

Thank you.

This concludes today's call. Thank you for attending. You may now disconnect.

Welcome to BofA's Global Tech Conference Day 2. I know it's a session after lunch, but that's why we got Matt over here to keep you all on your toes and stay awake. Welcome to DigitalOcean. We have the CFO, Matt Steinfort.

Matt, Thank you for joining us. You guys have done an incredible job here over the past few years in changing the entire sort of model of the company going from a developer cloud to what you are today. So do you want to take a minute to maybe talk about what has changed over the last couple of years, and we'll get into Q&A.

Yes. Thanks, Wamsi, and thanks for having me, and thanks for the nice walk-up music with a little Colorado theme, I appreciate that. You and I were talking about before the session, how much the company has changed in the last couple of years and how the narrative has changed in terms of the questions that we get.

A couple of years ago, we were still the third largest cloud by customer count, but we view it as kind of the smaller kind of toy cloud that was largely targeting developers and small kind of businesses. And I think that narrative there was, okay, don't your biggest customers tend to just go away from you because they get -- they outgrow you and they graduate. And we spent a lot of time. Literally, the #1 question, like the hard question when we would prepare for these sessions was -- the number 1 question was, why do you guys exist? Why can't the hyperscalers do what you guys do, and we had to spend a lot of time explaining that we have carved out a niche, and we serve an underserved portion of the market in the core cloud by those customers who aren't well served by the hyperscalers because they need more attention, they're digital native and they need more care and feeding, they need simpler, they need less requirements for long-term contracts and those kind of things. And -- but we couldn't explain away the fact that, yes, we did have a challenge, a leaky bucket in our top customers.

And so you roll that forward to today, and not only have we fixed that particular issue in the core cloud, we're now our $1 million-plus customers are growing. I think it's like 180%, and we hadn't had any churn in the last 4 quarters in that bucket. So we've fixed that challenge. But we've also launched and taken the same kind of fundamental value proposition into the world of AI. And we've developed a very compelling full-stack AI native cloud that is resonating incredibly well with our AI customers and the AI customer revenues growing 220%. And we built a very strong business there with a number of marquee inference oriented customers. So very, very different kind of business in just 18 to 24 months and required a lot of development, required retooling the leadership team at all levels of the company and kind of rediscovering our roots around focusing on the technology customers and the end experience but not on the individual single developer, but more on the larger digital and AI native companies. So it has been a dramatic shift.

Yes. No, sure has. Maybe it's getting reflected in the guidance that you provided, which you have opt a few times now. And I guess as you look over the last, call it, I don't know, 6 months or 90 days, like what has really changed in what you're seeing in discussions with customers that's giving you the confidence to keep adding this power and expansion plan that you'll have that you have communicated to investors and also taking the guidance up now to like 4% to 50% for next year?

Yes. It's the traction that we're getting with the customers, this is kind of -- it's a virtuous cycle, right? So when we won character a while back, that was the first kind of marquee recognizable name that we had won. And we had a lot of smaller customers who are interesting. But winning character gave us credibility in the market to be able to secure the cursors of the world and start to win some of their business, like hippocratic and idiogram and others. And the more we learn from them the more we can develop our software in a differentiated way and build on the platform advantage. And the more we do that, we're getting more attractive customers are coming on. So it's just -- it's snowballing. And so that gives us the ability and the confidence to take down incremental capacity. And as we're getting more experience deploying that capacity, we're having more confidence in our ability to execute on the time lines and turn up some of the new technology like liquid cooling is very new for the industry. And the first ones that we turned up are -- in fact, the first of the data centers we turned out this year was air cooled. So it was like that. We're just now getting into the liquid cooled, but we're delivering even our second data center that came on in April, which was fully liquid cooled, came on ahead of schedule. So we're just getting more confidence in our ability to execute. So that's kind of 1 part of the guide.

But the other part is just the customer traction that we're getting and the reaction we're getting and the demand signals that we're getting from those customers and the customers in the pipeline is very compelling and enabling us to really lean in.

Yes, you've gone from like 40-plus megawatts and added like another 30-plus now, like 60-plus more. How should we think about this incremental 60, when does that come online?

Yes, it comes online over the course of '27. So it's not entirely at the beginning, and it's not back-end loaded. So think about it, it's going to come on. It's 4 different data centers. So there's diversity there of providers and locations. So we feel good about our ability to manage those, and they'll come on over the course of the year. We'll provide more clarity when we get to the '27, the kind of more formal guide. We feel really good about the providers and the partners that we have. They're all very experienced data center operators. These are generally existing facilities where they're building out a data hall for us. Where there's existing campus, where there's a building that's like a shell that's ready to go and they just need to kit out for us. We feel good about the operational execution risk associated with that.

Yes. It's not like you're targeting a gig of capacity. These are much more measured sort of incrementals. But even despite that, it looks as though the market is relatively tight in terms of being able to procure whether it's servers, whether it's everything in our power, it's getting like all the liquid cooling in place. So are there any particular constraints that you see or what are maybe the tightest constraints that are going to determine how that 60 megawatts gets rolled out?

Yes. And I mean it's -- what is it, it's June now, which is crazy, but it's June, and we just announced 60 megawatts will come online in '27. So we're 12 to 18 months in front of that. So there's clearly -- the data center providers need to deliver on schedule. Like I said, this is not data center capacity being built from dirt, right? This is existing facilities. Power is already there. So there's not as much risk there, but they still need to kit it out.

The equipment side is less a scarcity thing than it is a timing and cost, meaning you can get the capacity [ pick that are quantum ]. So like you want the amount of GPUs that we're buying, you can find them. Multiple OEMs have them like we work with Dell and Super Micro and HP and Lenovo. We work with a lot of them. The question is, can I get it to you when you want it. And can they give you some kind of certainty on the price, because the cost -- the component costs clearly are going up. And so that's something that we have to deal with. But it's less of can you get it than it is can you get it when you want. And we've been pretty good about our relationships -- working our relationships with the OEMs and the evidence of that is we turned on the 6 megawatts of the 31 we're deploying this year. We said second quarter, when we provided guidance, we delivered in March. The second of the 3 is a 10 megawatts, we said second half, and we delivered it in April. So -- and when we say we delivered it, when we deliver it, it comes with the GPUs are right behind and the CPUs are right behind it.

Yes. Okay. That's helpful. How much visibility do you need to sort of go ahead and put this incremental capacity and sign these data center leases, like what sort of visibility do you have from your customers, both in terms of duration and as well as sort of magnitude?

Yes. Well, we've been, I think, very clear multiple times certainly, the demand that we have right now from our existing customers and could fill up the capacity that we have. And that we've got 3 to 4x the capacity or the demand and the capacity that we're bringing online. And that's what gives us the confidence and the signals that we're getting from our customers, the feedback that they're giving us on, the impact we're having on their total cost of ownership and some of the new capabilities that we've announced that are very compelling to our customers, many of them developed jointly with our customers, like the infras router capability. All of that gives us the confidence. Plus, the market opportunity is just -- I mean you're seeing -- you had a number of other companies here at your conference that are doing very well, very good companies that are also seeing similar demand. Like the demand in the market is, I won't say limitless, but it's -- there's certainly not enough capacity to meet the demand that's out there right now. And I think we're benefiting in that customers are beginning to understand the difference between coming to someone who can provide them a full set of inference capabilities with a full cloud stack next to it versus someone who can provide them with access to bare metal, which is valuable, but requires you to do a lot more and invest a lot more in. Our customers tend to not want to spend their effort and their resources worrying about like the infrastructure level and they want to focus on building their own software.

Yes. That makes a lot of sense. Matt, so you just said demand is like 3 to 4x. I mean literally where -- what you're able to potentially scale to. Why not -- why is 60 megawatts the right number in terms of getting when demand is like so strong?

Yes, I think that's probably the single question that everybody in the audience has got on their mind as well. 60 is what we had signed as of the last earnings. So that -- there was no magic number there. It was the number that happened to be signed at the time that we did earnings. And what we had said is we're continuing to actively evaluate incremental capacity for '27. We're looking at '28 capacity already. Clearly, we're looking at how do we grow even faster. So part of the executive team and the Board, 1 of our primary focuses is exactly that, like how fast can we go and that has -- there are a lot of dimensions to that. Like how fast can you go without being too highly levered. Like, okay, well, there's a throttle we have to have. And fortunately, we've demonstrated the ability to tap into the equity markets and do things to give ourselves an incredibly strong and flexible balance sheet. It has implications on, okay, well, what about the capacity? Like do you have -- how much -- how many megawatts do you have that you could sign right now? And then it also has the balancing that you just described of, okay, how much of are we willing to do with our business model of not having a 5-year bare metal offtake with an investment-grade kind of counterpart.

And so we're -- we look at our customers and their credit profiles and their growth trajectories and their funding and we allocate capacity to try to optimize in that. And so I'd say we're in a more nuanced game because we're investing ahead of the security of the committed revenue where a lot of our peers in the -- that are approaching it more from a neo cloud, bare metal side, they tend to have a customers and then secure the -- like that -- then gives them -- the finance and gives them the data center kind of financing, et cetera.

Yes. No, that's a great point. You mentioned on the cost side, we like that it's not about just maybe just availability, it's the cost of sort of bringing up this capacity and in the past, you have spoken about $20 million to $25 million per megawatt from a cost standpoint. The market has been very inflationary in terms of all these components, including now CPUs and memory obviously has been ongoing for a while. So as you think about putting that all together, what do you think about the incremental cost. Is that still the right range and it's the high end of that range? Or are we sort of talking about a big step up in sort of the cost to bring this up?

Yes. It's -- the $20 million to $25 million per megawatt in CapEx that we talked about was for the 31 megawatts that we were turning on this year. And again, you can see that. And if you look at our first quarter results, we added about $144 million of equipment finance obligations, and that was associated with the first 6 megawatts. So it puts you right in that range that we talked about.

As you pointed out, the costs are higher now. And they're higher on 2 dimensions. One dimension is just component cost. So the cost of storage or memory is the primary driver, but even storage, like NVMe drives or -- I'm probably missing a letter in there, are a lot more expensive than they were. And so you're starting to see component costs across GPU and CPU go up.

But the -- I'd say the bigger thing is actually the type of gear, the versions of the -- like some of the NVIDIA's latest technology that we're putting in, the B-300, you get more tokens per megawatt, it cost you more, but you get more tokens per megawatt. So the revenue potential is higher, which is good because that drives our $13 million in ARR per megawatt that will drive that up. But there is extra cost in there too, and so I'd say the good news with that is, well, it's everybody's got that cost and that's the cost in the industry. So at this point, we've been able to pass that on as just part of the pricing. And that's why I think you're seeing some of the prices of older generation also increasing just because it's -- you've got scarcity and you've got rise in cost structures, that's -- it's very difficult for industry investing this much to absorb that. It's got to be kind of passed on to the end customers.

What have you seen lately from your perspective on pricing in that regard?

Yes. I think you're seeing -- and this has been definitely a surprise for me, which is when we underwrote our investment in the 31 megawatts and even when we underwrote the investment in the 60 megawatts, we're projecting costs come down. I mean not that cost, prices come down, right? The older generation, you've got new generation technology, prices will come down. And so you got to be comfortable that you're underwriting a good business case even if that happens. Well, that hasn't been happening, right? The prices for H-100s and H-200s are increasing. And you've had a lot of people that -- Jensen and other folks, again, some of the folks at this conference talking about explicit increases in the and the pricing, we've seen that. And we're actually in a really interesting position in that because we don't have long-term contracts, we can rotate through a contract from 12 months ago comes up for renewal. We're not renewing that, could leave it close to what the price was before. We'll either increase the price, we'll move them off if they were a bare metal customer. We got a little bit of that left. We'll move them off. We won't even offer that to them and we'll move them to a higher grade level service or we can literally repurpose it and say, we're not going to sell it as GPU per hour. We're going to sell it as tokens and we can increase the monetization even then.

So in general, the pricing for older generation technology is not only stabilized, it's increasing, which is -- it's a really good thing from a near-term perspective.

So purposely, you actually want a higher churn in some ways of where you price...

We're [ churning ]. We were talking about this in 1 of the sessions earlier today. Traditional SaaS metrics are really not super relevant right now. Like people ask you about NDR and when are you going to include AI and DR. I mean I don't know if we ever will. One, you got to be like 12, 13 months in before that actually makes sense. But if I rotate out a customer at a certain price, because maybe they were a marketplace. And at the time, we didn't have a go-to-market and so we're selling through a marketplace. And I said, "hey, we're not going to sell to you anymore. We're going to sell the cursor." And then if you looked at NDR, you'd say, "Oh, you had churn." Did I? Like I'm using the exact same equipment, I mean no gap in revenue with a better customer at a higher price. So it's just -- it's -- like you said, it's a different world and different metrics matter.

Yes. Interesting monetization opportunity that maybe doesn't get reflected in some of the metrics. As you think about these new capacity ramps, how should we be thinking about sort of the margin trajectory as well as you're bringing on this capacity? Clearly you had some pressure at gross levels, but can you just talk about sort of how we should be thinking around some of the puts and takes on margins?

Yes. No, it's a great point. The -- when you add data set capacity, it does 2 -- drives margins down, like point bars, it can get it around that. Why? Well, because as soon as you turn the data center lease on, you take lease expense. And because we pay for our equipment over time instead of largely. As soon as you take the delivery of that, the equipment finance depreciation-related expense hits right away. So your gross margin takes a pop, like right out of the gate, and then you grow into it with a -- as you fill up the capacity and you generate revenue. So it has near-term temporal margin kind of impacts. You also have -- as AI is growing, like it was, what was it, it was mid-teens percentage of our overall ARR, right, in the last quarter. As that becomes a bigger mix, well, the margins on the core cloud business are -- if you didn't have AI, they'd be in the 65-ish range, right. And you're weaving in margins that are not that high. They're not like 25% margins like you would get if you're just doing bare metal, but they're not 65. So you've got 2 things going on. You've got the merging of the mix shift of more AI that's got lower inherent margins, but then you got some temporal stuff every time you turn on data centers.

But to me, the more important thing is to look at the aggregate margins when you include operating expense and you include everything else because the actual like resulting margins are pretty strong.

Yes. Maybe to touch on sort of what you announced at deploy, right? Like you spoke about the agentic stack that customers can now use. What's the adoption? I know it's very early days, right? But what are you looking at? What should investors be focused on in terms of across the several layers of the stack that you described, obviously, at the infrastructure layer like you have significant penetration. But as you go up the stack, how should we be thinking of the progression of some of those in terms of adoption rates? What are you anecdotally seeing, maybe if it's too early to sort of make a final judgment on where thing are?

Yes. I think like the simplest way to assess whether the strategy is working is to look at ARR per megawatt, right? Because it's like are we getting more value, higher layer services for the investments that we're making in capacity. So that's a simple metric. We also break down the -- within the AI customer revenue, how much of it is coming from each layer of the stack, right? So how much is being sold at bare metal, it's 19% now it's going to go down fast, and it -- like it's not only going down on a percentage, but on an absolute basis. We don't need to sell bare metal anymore.

The inference layer will be where you'll see the most expansion like the fast because that's growing like crazy. And that's the products that you talked about that we launched that deploy the inference router and we're now selling for token, like we're selling serverless, where it's like people come in and like, hey, I want -- gibe me tokens," and "here's my latency in my throughput. Here's the price." And they don't even know what the infrastructure is. So that's like starting to really take off. And then a lot of the agent layer capabilities that we're selling, that's been really, really interesting. We're the -- I think the top deployment of Open Claw. We have like 70,000 active Open Claw like instances where there's -- things are happening so fast. There's a new 1 Hermes something that before we're 1 of the top in that. And so we're starting to see a lot of that. Those things are going to be hard for the market to see outside other than the anecdotes. But if you look at ARR per megawatt, and you see that continue to progress. And then as we -- we'll continue to disclose the mix of bare metal versus inference versus core cloud pull-through. And you'll see more core cloud pull-through, I think as well. I think those are the -- hey, is there really a differentiation here, and is this 5-layer software stack, is this real? Or is it marketing speak, and they're just talking about software. Everybody talks about software. If you talk about software and you get $9 million to $10 million in ARR per megawatt, you might have some really interesting software. It might be a differentiator, you're winning more bare metal than your fair share, but you're not getting higher layer services on top of it. And that's where I think the proof is in, okay, well, how are you monetizing that investment.

As you think about that getting customers into incremental layers, like where does that go from a $13 million per megawatt kind of range that you articulated currently to like where could that go when you think about using more and more pieces of the agentic stacks?

Yes. It's probably too early to say how far it could go. But what I could say is it's already going higher than 13. We underwrote the 60 megawatts at a higher ARR per megawatt than that. Part of it was because of the incremental token capacity as we talked about that we get from the higher CapEx that we're investing. But also we just launched a lot of those inference capabilities. We're getting now -- like I'll give you an example, when we won Character a while ago, we had -- we're unproven. We had to prove to them technologically that we could provide the service, and we had to demonstrate that it would save them 30% to 40% or improve their throughput by 30% to 40%. And so we got their inference workloads, but we didn't get any core cloud. And now when you're talking to cursor, they come with core cloud. They're, okay, we need this and we need NFS, and we need other stuff. And so our confidence in the core cloud pull-through is increasing. So we're definitely confident in our ability to drive that above 13 on an incremental basis, and we're seeing that in our -- in the customer wins that we're getting.

Okay. Well, here maybe -- look, the stacks had a phenomenal move, right? And I think it's reflecting sort of the enthusiasm around, a, the product offering, b, your execution and c, frankly, like your guidance has gone a lot like you've shown revenue acceleration, which is very, very meaningful going from mid-teens to now a pathway to 50%-plus. What do you think that are maybe some of the key risks that investors should watch out for, especially on the execution side? And anything else that you would point out?

Yes. I think if you think of the competitive dynamic, you've got kind of 3 different -- 4 different players. You've got the hyperscalers that are very, very focused on their own -- supplying their own capacity for their own internal use and on some of the frontier models. You've got the neo clouds that started with scale and capacity and are trying to layer on software because they see the same thing we do, which is inferencing is going to be a much bigger opportunity than training, and there's a lot of better characteristics of it because it's a production workload that runs, it's not episodic and it's harder to move, and it requires all these other things. So they're trying to add software capabilities. You've got inference wrappers that are taking advantage of, hey, it's really expensive. And so if I can optimize this, I can sell to people because I can save them money, but they don't own any infrastructure. They don't own GPUs. They don't own CPUs. They don't have any core cloud capabilities. So you think about all of those 4 players, and then you take us, where we started with software and we've got a very differentiated cloud where other than hyperscalers, the only 1 who has a CPU full cloud plus all the AI capabilities, but we don't have as much scale. And so it's a race, right? It's like everybody knows what they're missing. And the question is how quickly and how hard is it to fill that. So for us, it's like, okay, how do we add scale in a way that's still profitable and not overlevering ourselves, et cetera. For the other folks, they've got to worry about, okay, how do you actually build software if you've never run a cloud and your primary customers, 95% of them you actually don't even see their software. They just run it on your infrastructure. So that's a different challenge. And if you're a wrapper, you're like, I don't have any scale at all. I've got features and functions. And so what do I do? And so I think if you fast forward this in 2 or 3 years, the question will be, well, who won in that regard. And we like our chances of adding scale and growing from a capacity standpoint because we believe that running a production cloud both AI and core compute is pretty difficult. And there's only a handful, 3, maybe 4 other players, I'll call hyperscalers that do that.

Yes. No, I'd bet on your chances, too, because it just seems as though filling the gaps of the other places is a lot harder, and you already had customer evidence that has shown that if they were customers of some of these competitors, it was probably more economical and easier to move on to your platform and use the stack as a whole. So definitely excited about the opportunity here.

Maybe to wrap over here since we only got a couple of minutes here, Matt. What should investors be most excited about as we look to the future over here? I mean, it feels like you've got this tremendous open-ended opportunity. So I would love to hear your parting thoughts on what investors should be most excited about?

I think that the thing that I'm most excited about, and I would hope that the investors would be excited about is, this is still such an early part of the growth trajectory of this market, right? There's clear evidence that there's going to be a massive amount of compute required, right? And I don't think anybody questions that. I think there's some questions around, okay, well, once there's a little bit more of an equilibrium and you don't have the scarcity, will there be commoditization and will there be people who got too far over their skis and are going to struggle. And when I look at what we're doing, it's like, well, what would you want to have? And when that happens, you would want to have highly differentiated services with more than just kind of a rental kind of a business. You would want to have customers that are embedded in your full stack of not just compute but storage and database where there's gravity, right, where there's really difficult to move that kind of -- those kind of workloads. And we're just kind of -- I think we're -- well, like you said, we've had a good run and the people are not paying attention to is, we're just getting started. There's like so much that we can do. And I think that we add some scale, we get some half, we'll be able to do some really exciting things.

Yes. Well, that's fantastic. In 23 years of doing this, I would say this is probably the BofA tech conference where the infrastructure bullishness that we've heard from across the board has been so consistent. And everyone thinks it's early days just because the demand is strong but the pipelines are exploding higher. So tremendous opportunity ahead of you. Great execution, and thanks for being here.

Great. Thanks a lot.

Thank you, Matt.

Thank you.

Thanks, everyone.

All right. Good afternoon, everyone. Thanks for braving another beautiful, wonderful 80-degree day here in Boston at the TMC conference. Look, I think this is going to be a really exciting panel this afternoon because I can think of very few public companies better positioned for the broad trends that we're seeing around inference, developer adoption and the future of the AI landscape than the DigitalOcean team, who we're privileged to have here with us today.

Just quickly, my name is Kevin Curtin. I look after the AI infrastructure investment banking business at JPMorgan, and it's been our privilege to have been aligned with DigitalOcean since their IPO in 2021.

But it's very clear, given the recent share price performance and investor reaction that the market is just now discovering a lot of the capabilities that we've seen for a long time wherein this platform i only accelerated by AI.

So it's a privilege to have here today both Paddy and Matt, CEO and CFO, respectively for, I think, a very great discussion. We've got about 35 minutes. I've got a list of questions prepared. But to the extent you all have questions, hopefully, you've heard from the conference organizers how to get those up digitally.

I'll be continuously taking a look at this thing throughout, but we'll also, of course, leave some time for audience questions as well. So with all that said, thanks again for joining us for the session this afternoon.

The first question, Paddy, Matt, you guys have noted that global inference traffic is expected to grow 10x by 2030 in that DigitalOcean's AI ARR is now 80% plus non-Bare Metal. So clearly differentiated from some of the other public Neoclouds. How does your software-centric approach to inference differentiate your margins and stickiness compared to other GPU rental businesses?

Great. First of all, Kevin, thank you so much for hosting us. It's an annual ritual and a very short commute for me personally. So I always enjoy this conference. So a great question to lead us off with. We embarked on this strategy for 2 specific reasons. One is, obviously, the financial profile. And the second one is just a self-reflection of what we are good at and what we want to focus on.

So what we have been really good at over the last dozen plus years is taking complex infrastructure concepts and making it super simple and accessible to developers at different stages of their respective life cycles. So that's what we were really good at in the Cloud 1.0 era. And that is the reason why we said we have a phenomenal opportunity to replicate that playbook and do it for the AI native era as well.

So when the initial stages of the AI wave started about 3 years ago, there was a flurry of activity in getting massive GPU farms stood up and cater to the needs of various frontier model companies that were looking to procure large clusters of capacity to train their models. And we were very passive participants in that market because we knew that our strengths lie in building great software and not necessarily running data center operations or building hardware systems.

But then fast forward to about T minus 18 months or so, we started seeing the inflection of -- inferencing was just on the horizon. And we started learning from our AI native customers in terms of what makes a great inference stack. And fast forward to 2 weeks ago, we announced the industry's first AI-native cloud which is an integrated architecture of 5 different layers, all the way from silicon to agents, all integrated into a single stack, which makes it very, very powerful.

And the results are there for everyone to see, as you mentioned, over 80% of our AI revenue is non Bare Metal. And the Bare Metal component is decreasing every quarter as we publish our results. And the primary reason for that is all of our AI customers are AI natives who are building and monetizing software, and they predominantly use this for inferencing and they use not just the inferencing services, but also they drag through a lot of our core cloud computing stack as many of these AI-native applications become more and more agentic, they dragged through a lot of our other parts of our software stack as well.

It's very clear the, I think, revenue growth trajectory is strong as a result of those trends you just described, Paddy. And so you guys recently raised your '27 growth outlook to 50% or more on the revenue line, which is a significant jump from just last quarter. What have you seen in the early utilization of your new committed capacity that gives you such high confidence in the acceleration in your business?

There are a couple of things. One, we added 60 megawatts on top of the 75 megawatts we already have online or bringing online this year. So the reason why it gives us a lot of confidence to do this is we are seeing a lot of strong indicators. Our growth has been accelerating every quarter for the last several quarters that I've been here.

And on top of that, almost every quarter over the last recent past, we have been setting new records from a net new organic ARR added perspective. So that's a great leading indicator, and then pretty much every leading indicator metric has been improving over the years, including the traction we have with our top customers, $1 million customers, the 500,000 customers and so forth.

So we are seeing a lot of leading indicators that gives us the confidence to go take this capacity down because we feel with the AI-native cloud stack that I just talked about, we have a very strong competitive moat and our customers are appreciating it, and the quality of the customers that you're seeing from us has also significantly improved over the last several quarters, and that gives us a lot of confidence that what we are taking on from a capacity point of view is something that we are very comfortable with, given that demand is still far exceeding the supply from a capacity perspective.

And so I think in terms of those leading indicators, it's both revenue and new customer wins, right? And so as we heard a really strong message to deploy your launch of the AI-native cloud is enabling high-growth AI natives like Cursor and Ideogram leaving hyperscalers for DigitalOcean. How does the 0 lock-in open-source stack that you guys highlighted at deploy, help you to drive better unit economics, better outcomes for these really desirable customers innovating at the frontier?

Yes. It's a great question. So -- and this journey started several quarters ago when we announced Character.ai. They also moved from a hyperscaler to us. And the -- there are multiple reasons for that.

Number 1 is, just purely from a cost performance perspective, we give them the kinds of throughput at a very low latency and accuracy that is not easy to get from other cloud providers, right? The work that we do at a kernel level, optimizing the software to ensure that we get them the type of throughput that reduces their total cost of ownership by 30%, 40% on many instances on the same class of hardware is a very compelling value proposition for these AI natives.

And we also have to understand that these AI natives build and monetize software. So for them, any spend on infrastructure hits their margin profile, right? And the more successful they are, if they're not careful, if they make the wrong choices in infrastructure and wrong choices in the models that they are supporting, it is just detrimental to their ability to keep scaling their business.

So they're very aware of the choices that they are making. So that's a very important part. The second thing is what we are about to witness in the general market, which is something that we've been seeing over the last 2 quarters, which is the step one for AI natives was to introduce intelligence into their workflow. That was number one.

Number two is the transition that most companies are going through is make their workflows agentic, right? It sounds easy, but it's really, really hard to do. And having the ability to do both the thinking part and the doing part in the same stack is very unique to our AI native cloud. There aren't too many cloud stacks where you can get intelligence to power the thinking of your application and get the right modern computing primitives to be able to perform the actions part that is required for the agentic, the new modern agentic applications. So having the ability to do both in a single unified stack is very, very powerful. And that's one of the things that AI native companies really appreciate from us.

Yes. You actually anticipated my next question. Agentic workloads, as you guys see them consume 15x the tokens and use 4x the CPU power. So if we're talking about building sustainable businesses, the unit economics of compute really matter. So at your deploy day, we saw a lot of really innovative features and full platform elements that you guys rolled out, like the model router, for example, that enable really preferential unit economics for your customers. Can you talk about the adoption of features like that and how much traction you're seeing with these AI natives?

Yes. It is still early days. Deploy was only 2 weeks ago, but we are seeing a tremendous amount of reception for the features we just launched. And for those of you who've not seen it, we announced a feature called intelligent router. So essentially, it is a layer that sits on top of the various open source and closed source models. In fact, Marc Benioff was on the podcast, All-in pod over the weekend. And he talked about the fact that Salesforce is using $300 million worth of Anthropic tokens. And he talked about the fact that, hey, there will be a company that will come and invent a layer that will do smart routing of tokens because not every task requires, Opus 4.7.

And I'm actually going to send him a note saying, "Hey, that company is already here, it is public, and we have actually shipped this 2 weeks ago." And the intelligent router does exactly that. And we showed a demo where we took a workload and routed all of the traffic to 4.7 on one side, and we used our intelligent router on the other side.

And as we went through the 5-minute demo, not only were we able to see that the cost differential was at least an order of magnitude, the second thing which is shocking to a lot of people was for many of the tasks like writing a unit test code or doing a very simple translation of actions to outcomes, these kinds of things were significantly faster and more complete using an OSS model like Kimi 2.6 or Qwen3.2 or something like that, for a number of reasons.

So we are getting into a paradigm where you need some of these sophisticated tools to manage your footprint for 2 reasons. One, obviously, the total cost of ownership or the ROI that you get, especially when you're trying to monetize software, you're very margin-aware and margin mindful and the second reason is you also don't want to be boxed into a single model provider. So we see a lot of our AI natives becoming multi-model and embrace open source in a big, big way. So to accomplish all of these things, you need a modern AI-native stack.

Yes. If you guys haven't seen videos, I checked them out over the weekend on YouTube, it's actually pretty cool. So if you want to see more about how that works, the videos are there. Really quickly, I think the IT security here is so good. It's actually kicking the iPad off, so to the extent you all have questions, I'd like to pause right now, in case you've submitted them, they haven't come through and give you guys a chance to ask. If not, I'll just keep rolling through our questions.

So sorry, again, for those of you that might have submitted them electronically.

All right, cool. We'll keep rolling. The key metric that really matters for you guys among many, is $1 million plus customer ARR and that metric grew 179% year-over-year this quarter, which was significantly faster than the overall business. What's driving success in retaining and growing these top customers when the market had previously worried about a graduation effect these days?

Yes. And so this is an overnight success 2 years in the making. It has taken a lot of heavy lifting from our side to ensure that we identify the reasons why some of our top customers with sophisticated workloads were forced to take some of those workloads to other hyperscalers, and we methodically started addressing those things.

So it took us about 4 quarters to put a dent on that. And there are a lot of things we did from a performance enhancement, advanced networking features, fix some of the security requirements for more modern distributed global developer organizations and things like that. So there were maybe half a dozen to a dozen capabilities that we're missing from the platform that we've addressed and fixed. And over the last 4 quarters, not only has the $1 million and $500,000 cohort started growing significantly, we have also seen a lack of churn there, which is really remarkable. And we are starting to see the same thing happen with our top AI workloads as well.

So it is a very deliberate strategy from our side because we felt like this is something that is so foundational for the success of the company that needs to be addressed. And this time, we want to be more proactive and address this type of graduation effect. And that's why we are really focused on addressing the needs of these AI natives proactively because the decision that they are making today when they are between $25 million to $100 million in ARR is very likely the same platform decisions they're going to stick with when they are at a $5 billion run rate. So we want to make sure that we catch them and catch them young, but have the ability to have them grow with us with our platform.

So Paddy I'll give you a break for one and call on Matt for a question on capitalization and how you guys are thinking about your balance sheet. You ended the quarter, Matt, with $1.1 billion in total liquidity and repaid the Term Loan A in full. How do you think about using your balance sheet flexibility to secure the data center capacity that you need, the long lead equipment, making investments on behalf of customers more effectively than peers who might not have access to the same pools of capital you guys do as a now $18 billion public company?

Well, I think we've demonstrated a couple of things. One is we're going to make very thoughtful kind of economic decisions about the pace at which we add capacity and the returns that we generate. And one of the tenets we have is we're not going to run the company as a public LBO. We're not going to be highly levered and burn a ton of cash. And so we have a handful of guardrails. Our guardrails are, we're not going to run the company at above 4x leverage for any kind of meaningful period of time. And if we do it, it'd just be because we turned on a data center and it was ramping. And we're not going to burn a ton of cash.

And so it was important for us then to position our balance sheet, which is now incredibly strong, very flexible balance sheet to be able to orient our leverage towards growth. And so to do that, we looked at our Term Loan A, and that was a $500 million facility that was costing us about $50 million a year in terms of mandatory prepayments and interest. And we said we could use that capital much more efficiently by using it to lease gear to pay for gear over time, which has been our primary financing vehicle for the new data centers when we bring them on.

Instead of paying $100 million upfront for gear, we'll pay it over 4 years or 5 years, and we'll still own it at the end. So by paying down the Term Loan A, we'll pay off the stub of the '26 convert at the end of the year. All of our leverage capacity, we can then deploy towards fueling the growth, which enabled us to add the 60 megawatts that we announced last quarter. And that's not the end. We still have a lot of room, and we've said publicly that we're continuing to evaluate adding incremental capacity that could hit '27.

Certainly, we're in the process of looking at '28 and '29 capacity as well. So we feel really good about where we sit. We think that we've demonstrated that we can tap into a lot of different parts of the market very effectively, the almost $1 billion of equity that we raised earlier this year and the stock was actually up that day was, I think, a good signal of the market appreciating us and the differentiated approach that we're taking to chasing the AI opportunity.

So maybe I'll just pause there because I know your capital structure as a public company with low leverage, that's not trying to lever Bare Metal contracts, so to speak, is a little bit different and pause for questions from the audience on capital structure specifically.

Over there. There is one there. We might have a mic for you in the back.

Just maybe talk about going forward, given the recent equity raise, how you plan to finance these new data center capacity additions? Mostly how much is from equipment financing, how much maybe from cash on hand versus other sources?

Yes. The primary funding vehicle that we're pursuing is equipment finance. So we will continue to lease equipment. And I hate to use the word lease because it confuses a lot of people. We will pay for gear over time. We'll pay for it over 4 years or 5 years or if we can get it 6 years, and we'll own it at the end. And it's the same as buying it, you just pay over time.

That's a highly effective tool for us. And we've been able to tap the OEM markets, the bank markets, and we still have some runway, a decent amount of runway to continue to tap those markets. And then beyond that, as you get into bigger quantums, there are other sources of capital that you can tap into in that same equipment financing structure.

Having said that, we're not wedded to that as the only vehicle that we use to fund growth. As we demonstrated, we could use equity. We've done converts in the past. We will optimize the cost of capital, and we will optimize the kind of cash burn. And sitting here with incredibly low leverage right now with a lot of growth in front of us, we have a lot of degrees of freedom. So I'd say we're open-minded and we'll be economic about what's the best cost of capital. But right now, the equipment financing market has been very attractive for us.

Any other question before we go on, on that topic?

Okay. Great. NDR has recently stabilized at about 101%, and you guys this quarter delivered a record $62 million in incremental organic ARR. Should we view Q1 as the definitive inflection point where your net expansion will now outweigh churn in the "legacy business? "

I think we passed that inflection point a while ago. And I think the better of those 2 metrics is the incremental ARR. We're adding incremental ARR at a record clip. Every quarter is an elevated number. NDR for us, again, you got to remember, we have 650,000 customers, only 20,000 of which are what we consider to be digital-native enterprises.

And so we think about that 630,000 customers is effectively a paid freemium group. The NDR of that cohort, just by the virtue of indie developers, they're tiny, in some cases, individuals, the NDR is always going to be something less than 100. It's always going to be in the mid- to high 90s. That waters down the overall NDR. And then you look at NDR, that doesn't even include any of our AI revenue.

So for us, the better metric is to just look at the growth rate of the key customer cohorts. And we disclose what we disclosed on purpose. We show you the $100,000, $500,000, $1 million customers and the NDR in each of those cohorts, which we disclosed last -- for the fourth quarter, I think it was 115% NDR for the $1 million-plus customers. We're not going to disclose that every quarter, but it's higher this quarter. All of them are higher this quarter. They're just going up.

And that's what you want to see is are your big customers spending more with you? Are you growing your big customers, and we are. You also want to see that your AI customer revenue is growing, and it's growing 220-something percent. I mean it's all of those leading indicators that people look at like NDR, which is a lagging leading indicator, are trying to get towards, are you going to grow your big customers? Are you going to be able to have them accelerate, and we're doing that. We're demonstrating that.

So we look at the absolute growth rates of the target segments much more than we look at the NDR metric. That's kind of a SaaS metric, and it's useful in certain contextes, but it's not a perfect metric for our business at this juncture.

One metric that stands out about you guys is the 220-something percent growth that you just quoted and you guys still have a consolidated 40-ish percent EBITDA margin. So how have you been able to figure out how to grow and invest in capacity profitably?

Well, one is we've stuck to our strengths. We don't chase Bare Metal opportunities where there's not a lot of margin, and it's a scale play. So we look for customers that are going to take multiple layers of the 5-layer stack that Paddy articulated so that they're buying not just GPU and to get GPU access, they want to buy inference services. They'll attach core cloud, and they'll be able to drive higher margins for us and higher ARR per megawatt.

And then, of course, we have the core cloud business, the CPU side, which I think people are starting in the industry to become aware is going to be an even more critical differentiator in the market for people who are offering AI services is you need a CPU-based cloud as well for all those primitives. And the margins there are materially higher than they are in the GPU world today.

And so we've been able to marry those 2 things together. But the other thing that I don't think people appreciate is the margin that matters for us is operating income, right? And that's -- if you're going to be apples-to-apples and compare us to software companies or compare us to AI infrastructure companies, our operating margin, GAAP operating margin is in the top kind of quartile of companies out there and none of -- very few of them are growing at the rate that we're growing.

So we're marrying very good kind of underlying economic margins. We're getting a ton of operating expense leverage as we grow. We're going to grow 50% plus next year, certainly not going to add 50% to our cost structure. There's other economics that are different than our core. These big AI natives, they don't pay with credit cards, so you're not paying a couple of points to Stripe. The bad debt profiles are different. There's just a lot of leverage that we can drive in terms of improved margins that offsets the fact that AI in general across the industry is a lower-margin business than historical kind of cloud computing. And by doing all of that, we've been able to maintain very, very strong operating margins and EBITDA margins, and we expect to continue to be able to do that.

Before we go back to Paddy to talk a little bit more about the business and performance. Any questions for Matt on capital structure or P&L?

All right. So Paddy, I think we've talked a lot about unit economics, about TCO advantages delivering rationally for your AI-native customers. All that would be one part of the story, but you guys have really proven that there's high performance, predictability and value to the platform. Independent Benchmarks recently ranked DigitalOcean #1 in output speed for models like Deepseek V3, outperforming hyperscalers by nearly 4x.

How much of this is driven by your vLLM optimization, your software stack, the decisions you're making that are value additive versus the underlying hardware in the architecture of your system?

Yes. It's a good question. So the underlying hardware in this case, it was B300 Blackwell is very important. But every cloud provider has access to the same hardware. So the difference is in the software optimizations we make at a kernel level. And it is the combination of the kernel optimizations we make for a specific family of models for a given hardware.

That's where the secret sauce is or the tuning that happens at a kernel level. And that is what is enabling us to get -- in this case, 230 tokens per second is significantly higher than what you can get from like most other cloud providers, including hyperscalers.

And we take a lot of pride in that. And that -- the reason why that is so important is right off the bat, if we are 50% better, that is 50% fewer tokens that our customers have to spend on. So it makes a meaningful difference for them in terms of how much the total cost of ownership, that results or delivers for our customers and we have a couple of very detailed technology articles, which we have authored along with our customers to showcase exactly how that happen.

So for example, there is this concept of disaggregated inferencing, where we split the inferencing step into multiple steps and optimize every step along the way using GPUs and the software that we are optimizing. So that's how we are able to get this type of throughput with extraordinarily low latency.

So if you look at the artificial analysis, you'll see the throughput and latency as the 2 axis, and we are #1 in both. So that really matters when you have inferencing happening at scale. And the other thing I want to also mention is, it's not just a result in a lab, right? Yes, that is important, but it is also equally, if not more important, that inferencing is a real-world workload. So one customer that we had up on stage for deploy is a company called Hippocratic AI, and Hippocratic AI provides voice agents for hospitals.

So this is as mission-critical as it can get because it is providing primary patient care in acute postoperative type of scenarios in hospital settings. So you cannot have -- you have to be super mindful of the latency. You have to be super mindful of the uptime for these kinds of workloads. And given that we have gone through the school of hard knocks for the last 15 years in terms of knowing how to build, operate and manage global cloud infrastructure really helps us make the transition from GPUs for training to actually deploying cloud infrastructure, AI infrastructure for inferencing at scale, which is a real-world workload.

With an eye on the clock, we've got about 5 minutes left. Any questions from the audience for Paddy or Matt?

So my last one, you guys have had a pretty blowout quarter in Q1. I think the stock reaction solidifies that. As you guys have done investor callbacks and obviously, the story is getting out much more broadly into the broader research community, what haven't we talked about today or gone into in as much depth that you think investors, analysts should know about DigitalOcean that maybe isn't quite out there yet?

So I'll start from a more strategic perspective. So first thing is this is once-in-a-generation opportunity that we are looking at. And what we haven't really talked about a lot is the competitive moat that we already have and we are building actively. So it is one thing to say, you have access to the same GPUs, I agree. But in inferencing and agentic applications, most of the magic happens in software. And that requires a very sophisticated software stack because the next generation of AI native applications are just starting to formulate, right?

You have coding as a micro vertical is in full bloom, but pretty much every other SaaS category, vertical software categories, physical AI are all in a very nascent stage. So there's a huge wave of new agentic applications that are going to be coming.

What do agentic applications need? They need a lot of intelligence and they need a lot of agentic computing. And what we have even today is a 5-layer integrated stack from silicon to agents, all working in a single stack. And I think that competitive moat is very, very important because this is what AI natives need. And this gives us a structural advantage, both from a technology perspective, but also from a unit economics point of view. And that's why we feel like we want to step into this generational opportunity, and we have the ability to do so.

Matt, anything you'd add to that?

I think that to me, the conversation has shifted notably with investors from a year ago, probably not even a year ago, we were still getting questions on why do you guys exist and why can't the hyperscalers do what you do. And I think that the market is starting to understand the difference, and I think that's reflected. I don't think they fully get the amount of addressable market opportunity we have.

There's a lot of start-ups that have a lot of buzz, inference wrappers that are providing layers of value added in the ecosystem. And as Paddy just said, we're launching capabilities that can consume some of those layers. And I just -- I don't know that the market fully gathers that yet. And I think that as we continue to win customers like Cursor and others that are huge customers of those kind of wrapper companies, but they're coming to us because they see the value that we can provide and the value that there is when you disintermediate that extra layer, I don't think that's fully contemplated by the market at this point.

I agree. And I think that's a great place to leave it. Thanks, everyone, for your time this afternoon and join me in thanking Matt and Paddy for theirs.

Thank you for standing by. My name is Jill, and I will be your conference operator today. At this time, I would like to welcome everyone to the DigitalOcean's First Quarter 2026 Earnings Conference Call. [Operator Instructions] I would now like to turn the conference over to [ Raju Patrike ], Head of Investor Relations. You may begin.

Great. Thank you, Jill, and good morning, everyone. Thank you all for joining us today to review DigitalOcean's First Quarter 2026 results. Joining me on the call today are Paddy Srinivasan, our Chief Executive Officer; and Matt Steinfort, our Chief Financial Officer. For those of you following along, an accompanying slide presentation is available on the webcast.

Before we begin, let me remind you that certain statements made on the call today may be considered forward-looking statements, which reflects management's best judgment based on currently available information. Our actual results may differ materially from those projected in these forward-looking statements, including our financial outlook. I direct your attention to the risk factors contained in our earnings -- in our filings with the SEC as well as those referenced in today's press release that is posted on our website. DigitalOcean expressly disclaims any obligation or undertaking to release publicly any updates or revisions to any forward-looking statements made today.

Additionally, non-GAAP financial measures will be discussed on this conference call and reconciliations to the most directly comparable GAAP financial measures can be found in today's earnings press release as well as our as well as in our earnings presentation that outlines the discussion on today's call. The webcast of today's call is available on the IR section of our website.

And with that, I'll turn it over to Paddy.

Thank you, Raju. Good morning, everyone, and thank you for joining us today. We had an outstanding Q1 2026, and I'll start with four headlines.

First, our momentum is accelerating. Q1 revenue was $258 million up 22% year-over-year, with million dollar plus customers growing 179% year-over-year to $183 million in ARR. AI customer ARR grew 221% to $170 million, and we beat every financial target we shared in our last call.

Number two, we launched the DigitalOcean AI native cloud last week, the most significant product launch in our history. With more than 15 new product launches across five fully integrated layers built into a modern, open unified stack, purpose built for the [ inferencing ] and Agentic Era.

Third, we are investing to meet our growing customer demand and to seize the material opportunity in front of us. We raised $888 million in equity during Q1 to strengthen our balance sheet and quickly utilize that flexibility to secure 60 megawatts of incremental capacity that is slated to ramp throughout 2027, bringing our total committed capacity to 135 megawatts.

And finally, we are again raising our near- and medium-term guidance on the strength of customer demand and the incrementally committed capacity.

For 2026, we are increasing our full year revenue growth projection from 21% and to approximately 26% year-over-year and expect to exit Q4 approaching 30%. And this revised 2026 growth is entirely driven by our previously committed capacity, without any top line benefit in 2026 from the new 60 megawatts. With the projected ramp of the incremental 60 megawatts in 2027, we are now projecting revenue growth of 50% or more in 2027, meaningfully higher than the 30% growth we communicated just last quarter.

I'll now spend a few minutes drilling down on each of these four headlines. The momentum we are generating is clear evidence of both our differentiated position and our strong execution across the board. It starts with the accelerating top line growth. Q1 revenue was $258 million, up 22% year-over-year and up over 400 basis points over Q4 2025 already strong 18% exit growth rate. We are delivering this growth by continuing to delight our top cloud and AI native customers.

Our AI customer ARR reached $170 million, growing 221% year-over-year. Our $1 million customer ARR rates $183 million, growing 179% year-over-year. These are not just customers experimenting on our platform. These are cloud and AI native companies scaling their businesses on DigitalOcean.

Our rate of acceleration is also increasing. We delivered a record $62 million in incremental organic ARR, the highest in the company's history. Customers see our differentiated value and are leaning into our platform. [ RPO ] reached $243 million, up an extraordinary 1,700% year-over-year. And we are doing all of this with strong profitability. We delivered 41% adjusted EBITDA margin and 18% trailing 12-month adjusted free cash flow margins.

Drilling into our growth. Our largest customers continue to be our fastest growing and their growth continues to accelerate. ARR from our $100,000 customers grew 73%, while our $500,000 customer ARR grew 132%. ARR from our $1 million-plus customers reached $183 million, growing at 179% year-over-year versus 123% last quarter.

Our AI customers are the other key driver of accelerating growth. AI customer ARR reached $170 million, growing 221% year-over-year. And most critically, inference and core cloud pull-through increased to more than 80% of total AI customer ARR, up from 70% in Q4. That number tells you something important. We are not a GPU rental business. We are a full stack cloud platform that AI native companies depend on to build, run and scale their production AI software.

Last week, at our Deploy conference in San Francisco, we launched the DigitalOcean AI native cloud. And let me explain why this is a very significant step. Four forces are fundamentally reshaping AI right now. [ Inferencing ] has overtaken training as the dominant AI computing workload. Open source AI is now in production at over half of AI native companies. Reasoning models are driving the majority of token consumption. And Agentic systems are rapidly moving from experimentation to production. Together, these forces represents AI evolution from "thinking" in which AI plays an advisory role to both thinking and doing in which AI delivers outcomes by executing autonomous tasks. The thinking part is powered by AI bottles in inferencing mode and the doing part is delivered by a variety of modern cloud computing modules, all working together to take intelligent, autonomous real-world action.

DigitalOcean's AI native cloud is purpose built for AI natives building exactly these types of workloads. It starts at the bottom with foundational layers. We operate a global scale infrastructure with 20 data centers purpose built for AI workloads running a full stack core computing platform with a complete set of computing primitive that Agentic workloads demand. Kubernetes, CPU and GPU droplet, advanced networking stack, including virtual private cloud, object block and file storage and high-performance NFS. This is part of the doing layer, the foundation that vast majority of GPU-centric cloud simply don't have.

Last week, we launched a new inference engine, which we co-invented with our customers to address their most critical inferencing needs, and it delivers a lot more than just serving tokens. It provides serverless and dedicated end points for serving up AI models batch processing for asynchronous token generation, an intelligent policy of our inference router that automatically selects the best model for cost and performance a catalog of over 70 open source and close source frontier models with day 0 access, multimodal capabilities and guardrails.

For customers who want to run their own models, we support BYOM, or Bring Your Own Model. This is the "thinking" layer, and it is far more than just serving tokens. It is about serving tokens efficiently with best-in-class performance, tightly integrated with other parts of the cloud.

Augmenting this new inference engine is our data and learning layer for which we announced an enterprise version of our managed MySQL and [ PaaS CRIs ] databases for advanced workloads. We also announced new vector database support for building Agentic workloads. We also launched a brand-new managed agents platform to give AI native everything they need to build, execute and operate autonomous agents at scale with open harnesses, sandbox, state management, agent observability, toolbox for external integration and [ Plano ] based orchestration on an open platform without getting boxed into a single LLM or platform provider.

This is the DigitalOcean AI native cloud, five fully integrated layers from silicon to agents with 0 lock-in because we offer open source options at every single layer. This is absolutely essential as our target customers are AI native companies who are creating and monetizing software. AI infrastructure is a material cost of revenue line item for these AI natives, especially when they scale, maintaining flexibility across models and platforms and leveraging the most efficient model capabilities for every specific task is an existential requirement for them. AI natives are increasingly adopting open source at every level, including multiple open source models to open agent [ harnesses ], open source vector databases and so on, to a wide lock in and deliver compelling unit economics for their customers as they go into hyper growth mode themselves.

Building a truly open, fully integrated platform is hard, and that difficulty is precisely what makes our platform durable. The market is validating what we have long believed that infrastructure without intelligence, without orchestration and a full cloud platform is insufficient for what AI native workloads actually demand. Agentic applications require intelligence CPU-based execution, stateful memory, manage high-performance storage and databases and orchestration, all working together natively not assembled after the fact.

Our integrated stack is built for exactly this architecture, and that's what enables us to deliver differentiated performance with compelling unit economics that matter to our AI native customers.

Leading independent benchmarking company, artificial analysis recently reported that DigitalOcean delivers the #1 output speed for leading open source model like DeepSeek version 3.2, Qwen version 3.5, the $397 billion parameter model across all cloud providers. Our 230 output tokens per second on DeepSeek V3.2 is 3.9x faster than one of the leading hyperscalers. This wasn't just a hardware story. It required co-designing every layer of the stack from NVIDIA's Blackwell ultra GPUs to custom VLLM optimizations, including speculative decoding and kernel fusion, which is exactly the kind of deep engineering that differentiates the modern AI native platform from GPU farms and inference wrapper providers.

The clearest validation of our strategy is the caliber of customers choosing to build and scale on us. We recently onboarded Cursor one of the fastest-growing AI applications ever built, for production inference, model fine-tuning and core cloud services. Ideogram, a leading text-to-image foundation model company migrated production inference from a hyperscaler to our AI infrastructure running their own model [ weight ] at scale. And Higgsfield AI, serving over 20 million creators with cinematic video generation run its full multi-model workflow on our integrated stack. Three different AI native companies in hyper-growth mode, running their production AI on our AI native cloud. And our pipeline continues to grow in both volume and strategic scale.

Let me spend a couple of minutes on our competitive positioning with our new platform announcement. At a high level, unlike the hyperscalers, we are more open, purpose built for modern software without the legacy complexity of enterprise workloads designed for the previous era. Compared to the GPU Neoclouds, which are optimized for large training clusters, we are a full stack inferencing and Agentic platform. And finally, while the inference wrapper providers offer tokens, we offer the breadth AI-native builders need to build complete modern software without forcing them to stitch a platform together themselves.

What makes our position genuinely durable is three compounding layers. Number one, our AI middleware. The [ Plano ] data plane and inference router built on technology from our recent Cataneo acquisition completed last quarter, sits between the agents and the underlying infrastructure, intelligently steering workloads across models, regions and accelerator types based on cost, latency and availability trade-offs at real time.

Second, our managed agents platform extends computing primitives up the stack with secure run times, execution sandboxes, background workers, observability, orchestration and much more. All purpose-built for Agentic applications to be built and scaled on this platform. And the third is data gravity through managed databases, vector stores, cashing and object storage, production data lives inside our DigitalOcean AI native platform. Models and GPUs are not sticky, data is.

For AI native, the decision of where to build is rarely about a single feature. It is about platform breadth quality of abstractions, openness of the platform and the absence of friction. Delivering that requires deliberate integrated engineering across every layer from silicon to agents. It needs an AI native cloud, which is what digital ocean has been building towards with millions of R&D hours over the last dozen-plus years.

The market opportunity is generational and we are poised to earn more than our fair share. Global inference traffic will grow 10x by 2030, and Agentic workloads consumed 15x more tokens than human users, a multiplier that compounds as AI matures.

And we're already seeing it in our numbers. Our AI customer ARR is growing 221%, and over 80% of that is coming from infant services and core cloud, not Bare Metal, these are companies running full stack production AI on digitation and they're accelerating. We are investing to meet this growing customer demand and to seize the opportunity in the massive inferencing and Agentic markets. In Q1, we raised $888 million in equity proceeds that enable us to expand our data center and GPU capacity to meet our growing customer demand while strengthening our balance sheet.

Matt will provide more details on the equity raise and our capital strategy later in our comments. But let me give you a brief highlight on our expansion plans. Starting with our existing committed capacity. We remain on track to deliver our previously communicated 31 megawatts as planned in 2026. With our Richmond facility beginning to ramp revenue in March.

On top of this, we have now secured approximately 60 megawatts of incremental data center capacity across four locations. Capacity that will ramp revenue throughout 2027. This brings our total committed data center capacity to approximately 135 megawatts. And given growing customer demand, we continue to actively pursue additional capacity beyond this new 60 megawatts capacity that will be targeted to come online in 2027 and 2028. The opportunity in front of us is enormous genuinely once in a generation. Every data point we see from our growing customer pipeline to the demand signals we are seeing and hearing from our largest customers to the reactions and interest in our AI native cloud reinforces that conviction. As we scale our business to meet this opportunity, we will continue to make the right long-term business decisions to seize this moment while building a durable and profitable growth engine.

With momentum continuing to grow, we are further raising our near- and medium-term outlook for the full year 2026. We now expect revenue growth of approximately 25% to 27% year-over-year with an exit growth rate approaching 30%, a full year ahead of the guidance we provided just last quarter. This accelerated 2026 growth is based solely on the performance of our previously committed capacity and doesn't include any projected revenue uplift from the newly committed 60 megawatts. We expect to deliver this 2026 growth with high 30s adjusted EBITDA margins and 9% to 12% adjusted free cash flow margins, which does include some start-up costs for the new 60 megawatts.

Looking further out, we now expect 2027 revenue growth of 50% or more, up from our 30% guidance last quarter with approximately 40% adjusted EBITDA margins and high teens adjusted free cash flow margins. This combination of rapid revenue growth and true durable profitability puts us in a ratified company. DigitalOcean is one of just a handful of names across a broad set of software and AI infrastructure players, delivering both attractive GAAP operating margins and material revenue growth.

As I shared on our last call, growth and discipline are not trade-offs for us. They're both operating principles. And our execution of these principles is clear in our results.

With that, I will turn it over to Matt to walk through our Q1 results and our updated guidance in more detail. Matt, over to you.

Thanks, Paddy. Good morning, everyone, and thanks for joining us. As Paddy just shared, we had a very good quarter. In my comments, I will review the financial results in detail, walk through our recent balance sheet and capital allocation actions and then provide an update to our near-term and medium-term outlooks.

Starting with Q1, our results were very strong, and we exceeded the guidance we last provided on all key metrics. Q1 revenue was $258 million, up 22% year-over-year. above the top end of our recent guide. The vast majority of this Q1 revenue beat came from strong retention in our top [ DNE ] cohorts and from expansion in our top cloud and AI native customers.

The Richmond data center, which began ramping revenue in March, contributed less than $500,000 of revenue and less than 20 basis points of year-over-year growth in Q1. Our top customers continue to drive our growth. Our $1 million customer ARR reached $183 million, growing 179% year-over-year. AI customer ARR reached $170 million, growing 221% year-over-year.

And we continue to deliver both durable and profitable growth. First quarter adjusted EBITDA was $105 million, up 21% year-over-year with an adjusted EBITDA margin of 41%. GAAP operating income was $37 million, with an operating income margin of 14%. Adjusted operating income was $64 million, with an adjusted operating income margin of 25%.

Trailing 12-month adjusted free cash flow was $171 million or 18% of revenue. Trailing 12-month adjusted free cash flow less lease principal payments was $154 million or 16% of revenue after including $17 million in financed equipment principal payments over the last 12 months.

Next, I'll spend a few minutes on the recent equity raise and what it means for our financial profile and for our capacity plans. In Q1, we raised $888 million in equity, and we have already put the proceeds to work across two important priorities.

The first priority was strengthening the balance sheet. We repaid our full $500 million Term Loan A, saving roughly $50 million per year in cash interest and mandatory prepayments. We intend to use a portion of the remaining cash to retire the outstanding $312 million 2026 convertible notes when they mature. Collectively, these actions result in a flexible balance sheet with no material maturities until 2030.

The second priority was expanding capacity to meet demand. As Paddy shared, we have secured approximately 60 megawatts across four new locations, an 80% increase in our committed capacity. This capacity is projected to begin ramping revenue over the course of 2027. While there won't be any 2026 revenue impact, the build-out of some of this capacity is likely to start in late 2026, which will impact 2026 cash flow and margins.

We expect the CapEx per megawatt in this new capacity to be higher than for the equipment ordered last year, for the 31 megawatts. The increase is driven both by the rising component cuts the entire market is seeing and higher cost and higher token capacity equipment that we plan to install. We expect the incremental ARR per megawatt to be higher as well. And importantly, we expect to generate the same or higher return on investment in these new data centers. We are likely to continue to align the timing of our investments with revenue by financing a material portion of the equipment for these facilities.

With all of this, we expect to exit 2026 at approximately 3x net leverage with no material debt maturities until 2030. Looking forward, we are again raising our near-term and medium-term outlook. The strong Q1 retention and growth in our top cloud and AI native cohorts has continued in Q2.

For the second quarter of 2026, we expect revenue of $272 million to $274 million, representing 24% to 25% year-over-year growth. We expect second quarter adjusted EBITDA margins in the range of 37% to 38%. And which is $102 million at the midpoint, up 14% year-over-year.

We expect non-GAAP diluted net income per share of $0.20 to $0.23. And based on approximately 121 million to 122 million weighted average fully diluted shares outstanding. Note that our shares outstanding projection includes a benefit from the projected anti-dilutive impact of the cap call that we purchased along with the issuance of our 2030 notes.

For the full year 2026, we are again meaningfully raising our outlook. We now expect full year 2026 revenue of $1.13 billion to $1.145 billion, representing 25% to 27% year-over-year growth, with a negative growth rate approaching 30% in Q4. Again, this does not include any projected revenue from the newly committed 60 megawatts.

We expect strong full year adjusted EBITDA margins of 37% to 39%, which is $432 million at the midpoint. Projected adjusted free cash flow margin will be in the range of 9% to 12%. And which includes roughly $100 million cash flow impact in 2026, a projected nonrecurring start-up costs for some of our newly committed capacity.

Without these costs, adjusted free cash flow margin would be roughly 18% to 21% for the year, above prior guidance. We expect adjusted free cash flow margin less equipment finance principal payments to be slightly positive for 2026, including the impact of the $100 million in cost for 2027 capacity.

We expect full year non-GAAP diluted net income per share of $1.10 to $1.20 on $118 million to 119 million weighted average fully diluted shares outstanding. This is an increase to our prior guidance despite the equity raise as the interest savings from retiring our Term Loan A more than offset the impact of the higher share count.

We are also increasing our medium- to long-term outlook, the 30% 2027 revenue growth outlook we provided last call was based solely on the 75 megawatts of capacity that we had active or under contract at that time. With approximately 60 megawatts of additional committed capacity, projected to begin generating revenue over the course of 2027, we now expect 2027 revenue to exceed $1.7 billion, full year growth of 50% or more year-over-year. We will deliver this growth while working to make smart investments generate attractive returns and maintain a strong and flexible balance sheet.

Our margin outlook for 2027 is healthy. We project approximately 40% adjusted EBITDA margins and high teens adjusted free cash flow margins. While we are excited by our progress and the increased growth outlook, we're not stopping there. We continue to actively look for opportunities to further accelerate durable and profitable growth. With that, I'd like to turn it back over to Paddy.

Thank you, Matt. Before we move to Q&A, let me recap what we shared today. First, our momentum has never been stronger. Our $1 million customer ARR reached $183 million, growing 179% year-over-year. Our AI customer ARR reached $170 million, growing 221%, and over 80% of that is coming from infant services and core cloud, not Bare Metal. We are an AI-native inference cloud, not a GPU landlord.

Second, we launched the DigitalOcean AI native cloud. We unveiled our full platform last week at Deploy conference. We acquired Cataneo to accelerate our open source AI stack. We landed multiple marquee AI-native customers, including Cursor. Our differentiation is clear. The pipeline is deep and the wins are real. We are the AI native cloud.

Third, we are investing to meet our customer demand. $888 million raised 60 megawatts of incremental capacity committed. We are building for 2027 and beyond with disciplined capital allocation and a strengthened balance sheet.

Finally, we again raised our near- and medium-term outlook. Projected exit 2026 revenue growth approaching 30%, accelerating to 50% or more revenue growth in 2027, attractive margins and a flexible balance sheet. We continue to build a durable and profitable growth engine.

The inference and Agentic economy is real. The demand is real. And DigitalOcean with its AI native cloud is purpose-built for this opportunity.

With that, let's open it up for questions.

[Operator Instructions] Your first question comes from the line of Kingsley Crane of Canaccord Genuity.

Needless to say, congrats on the momentum you've earned it, you continue to earn it. It's great to see One of the ideas over the past couple of weeks is that the mix of CPU and GPU should be closer to 1:1 with the Agentic workloads compared to pure LLM calls. And you talk about that new arrow thinking and doing in your deck, which was really well prepared. Just curious how relevant is that CPU renaissance for your business given your large core cloud and CPU footprint? Just trying to think about the quantitative benefit that could create.

Yes. Thank you, Kingsley. Appreciate your question. Yes, I think it is unmistakable that we are moving more and more towards an agent ear where more software is going to be rearchitected and there will be a heavy dose of autonomous agents performing tasks that were previously handled by humans. So in that era, the doing part, as I mentioned, will also require intelligence, but it is going to require a tremendous amount of computing that until about 12 months ago or more precisely until open [ plot ] really showed us the blueprint. We were in really as an industry contemplating how compute intensive it is going to be.

When I say compute intensive, it is just not CPUs, right? It is high bandwidth memory. It is advanced databases like the ones that we just announced last week, it is safe agent execution, it is orchestration between these agents. There is a tremendous amount of modern computing primitives that are required to orchestrate all of this. So I don't know whether the ratios that have propped up with say, CPUs to GPUs will go from 1:12, as we were previously thinking to 1:1, I don't know exactly what that ratio will end up being. But what I can tell you is that we are going to need a hell a lot of more compute to do all of these things as more software gets rearchitected over the next handful of years to be more Agentic, which requires both inferencing for the thinking part and a lot of computing for the doing part. So we are preparing for that, the new capacity that we have just took on.

All of our new data centers are deploying our full stack AI native cloud. So it is just not inferencing services. It is the full stack AI native cloud that is getting deployed in these data centers. And we are getting ready for a compute-heavy future, and we are starting to see that in a very pronounced way from some of our advanced AI native customers as they themselves move into an Agentic Era.

It's really helpful. And then for either Paddy or Matt, we've been thinking about low to mid-teens revenue per megawatt for AI. You mentioned that the incremental capacity you're bringing on could be higher. And then just in addition to that, like to what extent can software capabilities like inference engine and [ French ] router, open source model adoption, agent framework, push that revenue per megawatt higher. I think we're all doing that megawatt math, but just curious to what extent that figure can become untethered from the peers there? .

That's a great question. We definitely expect that we can increase that $13 million per ARR per megawatt over time. I mean you're already seeing that non Bare Metal over 80% of our AI customer ARR, and that should increase the ARR by itself.

We're also expecting, as you just pointed out, there's going to be a lot of core cloud and a lot of compute that gets pulled through with that. Right now, it's still -- it's a modest amount of core cloud pull-through, and we think there's upside there. And then to your point, all of the capabilities that we announced to deploy, the serverless inferencing and a lot of these other capabilities, they detach the pricing and the value creation from a dollars per GPU hour and enable us to capture both higher revenue and higher margins with stickier services. So we're very optimistic about our ability to drive the ARR per revenue up over time. And certainly, that's part of our investment thesis as we've taken on this incremental capacity.

Your next question comes from the line of Gabriela Borges of Goldman Sachs.

Paddy, you start up this conversation talking about how the beat in the quarter was not driven by new capacity coming online, but rather previously committed capacity. So I defer your thoughts on that. Talk to us a little bit about how we should think about the beat on rate [ cans ]. You're already giving us visibility into 2027 based on capacity coming online. But in any given quarter, what levers do you have to be in raise? And maybe if you could comment on the pricing dynamics and believe as you can pull on pricing within that.

That's a great question. I think when we guided to 2026, and we outlined the pace at which capacity was going to come online this year, there's a number of assumptions that we had to make in that, that gave us the ability to have very strong confidence in the guidance that we were providing.

One was the timing of the facilities coming online. The second was the -- our ability to sell into that capacity as it came on and the third, the pricing at which we're selling into that capacity.

And if you think about all of those dimensions, again, when we provided that guidance, which was late last year, early or early this year, we had to make sure that we had enough cushion. And what we're finding is we're doing pretty well on all three of those dimensions. The Richmond data center came online. We had said second quarter, it came online in March. It didn't contribute much the first quarter, but it's online and ready to go ahead of what we had said. We're able to sell into it. Much, I'd say, on a very appropriate and aggressive time line, which is really good. And then as you're seeing in the market, the pricing for GPO hour, even services right now is not seeing any kind of price compression. In fact, we're seeing increases in the prices for [ H100s ] and [ H200s ] and some of the legacy gear.

So I'd say we have sufficient ability to continue to beat and raise we just outlined the incremental 60 megawatts for next year. And we're taking a very similar approach, which is we'll be cautious about our expectations around timing of delivery, we'll be cautious about expectations of how long it takes to sell into it, and we'll be cautious about the pricing that we get and then we'll work to exceed that.

Matt, maybe I'll pick it up just some of those comments on being cautious. So I think we can all agree that we're pretty early in what is going to be an incredible product cycle. At some point, the product cycle will peak. So I guess the question is for the both of you. What are the demand signals that you're watching to be able to figure out whether it's 2027 growing north of 50%. Is that the peak growth rate? Does it accelerate from there, does it normalize and come down? What are some of the metrics that we could potentially be tracking from the outset? And what do you track internally?

Yes, I can start at a high level, and then I'll let Matt comment on your specific 2027 question. So we all agree, Gabriela that this is such a tectonic shift in how software is built and delivered. And one thing that I also want to highlight here is that inferencing and Agentic workloads will scale very differently compared to training. Training is a onetime, almost episodic turn on, the entire cluster comes online and just stays static from a workload perspective. While inferencing and Agentic workloads have more of a cloud kind of characteristics in terms of how the workload ramps, although the gradient of the ramp has been significantly steeper than we have ever seen with traditional cloud software.

So a lot of our confidence is coming from observing our big marquee AI native customers and seeing their workload growth and hence, the inferencing demand that they translate on to us and our platform. So in terms of the product cycle peaking, I think that is -- we are still a few revisions of our products, certainly and also as an industry to get to that peak cycle.

[ Openly ], I have to remind everyone is barely 100 days old. And since then, there have been a few other personal productivity agents like Hermes agent and a few others that have come and the whole industry is now figuring out what agent harnesses should look like. It is still a very, very early days of the Agentic architecture. So I expect the product cycle refresh to continue for quite a bit into the next several quarters before we can say, okay, we now have a blueprint for how these modern autonomous systems are going to be built and operated in scale. So I think we still have a lot of innovation ahead of us.

And what gives us a lot of confidence is having this front-row seat working with these marquee AI-native customers gives us a tremendous opportunity to learn about their application patterns. And this luxury is available to us because we are not just a Bare Metal provider. These customers want us to be in the room where they are solving these problems, and that's how we were able to build a lot of these things that we saw last week in terms of innovation, like the intelligent routing, the -- many of the cashing techniques that made us the #1 in DeepSeek and Qwen token throughput and time to first token and things like that, it gives us a front-row seat and a co-invention opportunity to do this alongside our customers. So I definitely feel like the product cycle is not going to peak anytime soon.

And I think the best metric to watch, which we're watching is ARR per megawatt. I mean if you think of token efficiency being one of the primary differentiators in terms of your ability to provide value to your customers is how much revenue can you get for those tokens and how efficiently can you provide them? And how sticky are those services that you're providing, that should all translate into higher ARR per megawatt, which is why we've introduced that metric, we track it internally, and it's all about optimization for us, and that's where we're focused that's what we would point the market to watch as well.

Your next question comes from the line of Mark Zhang of Citi.

So very nice to see the growingness of the non Bare Metal ARR this quarter. Just want to dig into some of the dynamics there in the input. So I wanted to get a sense of contributions from just new land versus existing conversions of the existing Bare Metal customers? And then how should we sort of like think of the pace of the mix shift going forward? And can you give us a sense of the ASP upfront, when you convert from Bare Metal?

Thank you, Mark. Your line was a little choppy, but I think I got the essence of your question. So in terms of the mix of the customers, it's a healthy mix of AI native customers that are new to our platform, that are not just consuming core AI services, but also by the nature of their inferencing workloads, they use storage systems and database systems and also increasingly core computing primitive, but we also have some of our existing digital native enterprise customers also starting to ramp up their AI innovation and AI workloads. So it goes both ways, and we are super happy to see that.

And in terms of the Bare Metal consumption, pretty much most of the customers that come to us now are coming to us because they see this rich set of inferencing entry points. So last week, we announced serverless inferencing, dedicated inferencing, batch inferencing and things like that. Increasingly, customers are realizing, especially the AI natives that they were forced to deal with all this complexity over the last couple of years, not because they wanted to, but they have to because there were very few vendors who were able to provide this kind of kernel optimization and performance enhancement using software and hardware codesign.

But now that these kinds of capabilities are available out of the box from our AI native cloud. We are seeing a lot more appetite from our customers to come in at a higher altitude in our platform and we are not having to sell Bare Metal at all. In fact, we don't even have that as part of our standard pitch.

And from a timing standpoint, this is one of the benefits of our consumption-based model with but where we're not locking in bare metal prices for 4 and 5 years. As these Bare Metal customers, if you notice in the materials we provided the Bare Metal not only decreased as a percentage, but it actually decreased in absolute dollars of the AI customer ARR. That's because as these customers come up for contract renewal, we have the opportunity to resize and reconfigure that capacity. If we want to make that available to serverless inferencing, where we know we'll earn a higher return than Bare Metal, that's what we do. And so we have the ability to steer that percentage down by not consuming our scarce capacity for Bare Metal services.

So not only are new customers not asking for it, but the customers that are on it right now, we can rotate them off into the new services or we can repurpose the capacity for higher-margin services, and we control that.

No, that's terrific. And then just maybe a follow on. It's terrific to see the new five layers also referencing a new platform that you guys had provided last week at the pot. How should we sort of think of the maybe like changes to the gold market from here? Obviously, there's a lot to sell. There's much more products to for customers to consume. How do you -- how are you thinking about just in terms of the go-to-market partnerships and how you really like officially land new customers won this new module?

So our go-to-market over the last several quarters has been aimed at getting marquee AI-native logos. And that's how we have landed some of the customers that I was so proud to announce today. And we just have to scale up in doing what we are already doing.

So just as a reminder, we have a very small but mighty team of AI native focused sellers that are quite capable of selling our AI native cloud stack. On top of it, we also have a very focused start-up ecosystem team that nurtures high-quality AI native companies in Silicon Valley and nurture them through their growth phases. We also have a tremendous luxury of having perhaps the best product-led growth machine, which keeps growing in strength. So we get a tremendous amount of traffic and volume through our product-led growth flywheel, which includes a heavy dose of AI native customers that absolutely just love the simplicity and the absence of friction in our platform that enables them to just come and try our platform and do it without any human intervention.

So we have multiple front doors as a way to solicit customer entry into our platform. So we'll be fortifying some of those things, and we have a very strong partnership team that enables us to build relationships with various frontier model and open source model companies in the rest of the ecosystem.

Your next question comes from the line of Jason Ader of William Blair.

Paddy, you guys are exploiting a gap in the market right now, especially with the Neoclouds, but the Neoclouds are all messaging shifting to a full stack approach and a focus on inferencing. So I guess my question is, how sustainable is your differentiation relative to the Neoclouds and what drives that?

Yes. Great. Thank you, Jason. I think the market opportunity is just huge and tremendous, right? We feel that the Neoclouds adding software capabilities is a great validation of our strategy and we've been saying that for a long time.

But we are in fundamentally different businesses than the Neocloud. They're training first, and that's a great model. And they have a small number of highly concentrated customers with take-or-pay agreements and their needs, that type of contract needs a tremendous amount of infrastructure and discipline and execution to pull that off. So it is a significant heavy lift to deliver on these massive hyperscaler offtake contracts.

So I like our chances of continuing to innovate on the software stack, as I said, it takes a lot of hard work to build a well-integrated stack like the one that we announced last week. It is just not a stack that lives on a PowerPoint slide. You can log into cloud.digitalocean.com and see how these layers work together. We are also incredibly proud of the fact that we have made the stack completely open with open source options at every single layer.

That is a pretty big deal that I want everyone to appreciate because our target customers are AI-native customers. and they feel very uncomfortable boxing themselves into a single LLM provider. That is just not how their businesses will scale. And for them, having open source work as well as close source as part of the native stack is very, very important.

So driving this kind of integrated open source enabled stack is really hard. And I like our focus. I like our discipline in terms of doing this. And the market opportunity is going to be so big that I feel very, very convinced that if we focus on learning and understanding our customers better than anyone else and translate that to product innovation, everything else is going to take care of itself.

I keep telling my teams be extraordinarily customer-obsessed and competitive aware, not the other way around. We should obsess over our customers first so that we can build the best product for them while being aware of competition, and not the other way around. So I feel we have a lot of room to run with this strategy.

Okay. Great. And then one for Matt. Matt, for 2027, you talked about adjusted free cash flow margin in the mid- to high teens, I believe. Could you give us a sense of what it would be, including lease payments?

That's a great question, Jason. It's hard to answer, though, because it will depend entirely on the lease terms that we have. So whether we lease over 4 years or 5 years or a longer period, and it will also depend on the mix of what we lease versus what we pay for upfront. That's why we're not guiding to that at this point.

What I can tell you is that we continue to make very disciplined investments, we've created a lot of balance sheet flexibility for ourselves with the equity raise. We've got a lot of options at our disposal. And we're very excited by the return on investment that we're underwriting for these new facilities.

So we'll continue to operate with discipline, but we can't provide specificity on the -- what the lease payments are going to look like in 2027 because we don't know yet.

Your next question comes from the line of Wamsi Mohan of Bank of America.

Paddy, for -- when you look across your customer cohorts, how much penetration are you seeing of AI-driven workloads, as you look at sort of $1 million plus in the $500,000 plus customer cohort? And are you actually seeing because of AI, do you expect over the next 2 years to have an even higher chunk of customers graduating from this $500,000 to $1 million-plus cohort as you look through the next few years? And I have a follow-up for Matt.

Yes. sees, you're absolutely right. I think the short answer is yes to both. We have a good mix of AI as well as cloud native customers in the $500,000 and $1 million customers. And yes, it is a very important motion that we drive internally to look at every 100,000 customer and drive our teams to find out what is blocking our customers from being a $500,000 customer. And similarly, we look at every 500,000 customer and find out how we can make them $1 million customer and so forth.

So with the increased adoption of AI in these customer cohorts, we fully expect those numbers to keep going up to the right for sure.

Your next question comes from the line of Tom Blakey of Cantor.

Congratulations on the great results here. Maybe a couple of questions on my side. Paddy, we've talked prior about 3 to 4x demand in terms of your 75-megawatt capacity was really impressive to see you announce Cursor here, a great win. Congratulations.

Just wondering if you could just maybe update us on the framework of what you're seeing there in terms of your customer selectivity and maybe even turning some customers away in this type of market?

And then secondly, for Matt and maybe the team just CapEx per megawatt, I think investors would love a little bit more color in terms of how much higher this can go for the 60 megawatts. And would it be difficult to just upgrade the prior capacity from a software upgrade perspective to the AI native cloud capacity to maybe kind of pull some of that in, that would be helpful.

Yes. I think on the last thing, we -- it is hard to have a non-AI data center deployed with AI hardware because of the limitations, especially all of the new ones that we're deploying are all direct liquid cooled and the hardware specs are just different, Thomas. So that's that.

And going back to your first question around the pipeline coverage and how we allocate capacity. I mean, that is some -- a new muscle that everyone in the industry is learning, right? Our pipeline, as I mentioned several times, is 3 to 4x, if not more, in terms of the actual capacity that we have. Which is a great problem to have, but it is a problem that we are very and very thoughtful about resolving because we have to make some bets just like our customers are making bets on us. We have to make bets on how we want to allocate the capacity. Because, as I said in the last call, if we decide to just sell the capacity to the first or the biggest or the loudest customer we'll be all done. We can go home and the capacity will all be taken. But we have an intention to run this like a cloud, right, where we want as many customers as possible so that we can learn, we can build a better product and build a bigger competitive moat that customers that only have -- or platforms that only have a few concentrated customers simply don't have the luxury to learn and innovate as fast as we are.

So it's a balancing act that we are trying to figure out, but so far, so good with the types of customers we're bringing on board.

In terms of the cost of the CapEx, it's certainly going to be higher than what we experienced for the 31 megawatts equipment was ordered in 2025. And you're seeing broadly across the industry, component costs are going up. But more importantly for us, we're putting in gear that has higher token kind of capacity and capabilities. And we expect to get the same or higher ROI on the investments that we're making.

So we'll invest a bit more. We see a phenomenal opportunity in front of us. We got a very differentiated position. We're going to get more capacity out of the investments we make, and we're going to earn similar or better returns on the investments.

Your next question comes from the line of Josh Baer of Morgan Stanley.

Congrats on a wonderful quarter. I was hoping you could double-click a little bit on GPU and other pricing trends that you're seeing in the spot market. And wondering if you can quantify the portion of your business that's on demand and exposed to spot versus what portion is contracted and has fixed pricing? And any way that you can characterize the benefit in the quarter or the impact of the 2026 guide from spot market pricing?

It's interesting, Josh, that you point to the spot pricing. So we have a portion of -- a small portion right now of on-demand because most of our capacity is locked up with a customer. But if you think about the core of your question, which is how much exposure do we have to the ability to raise GPU prices along with the market. Because we don't have 4- or 5-year contracts with our customers, if we're locked into a customer, it may only be for 3 months or 6 months or a year. And as I said earlier on the call, as those contracts are coming up, we can rotate. One, we can just raise the price on that customer to whatever the current market prevailing prices. Two, we can rotate it completely out of if it's a GPU per hour price, we can say we're not going to sell that capacity in that model any longer. And if you're interested in that you've got to take our on-demand pricing or you're going to take serverless inferencing.

So we have the ability to adjust to the market, I'd say, probably more readily than maybe some of the other folks in the industry. So we feel very, very good about our ability to adapt to pricing.

And as I said, to Gabriela's question, that ability and our ability to execute that is part of the reason why we're able to raise the guidance for this year without getting any benefit from the incremental capacity that we just announced. So that's a great question.

Your next question comes from the line of Radi Sultan of UBS.

If you think about adding more capacity and as the existing AI customer cohort scale, like how should we be thinking about the gross margin profile, this incremental capacity you're looking to add once it's fully utilized. And you mentioned, Matt, the increased component costs. But yes what are the key puts and takes there we should be keeping in mind just on the margin side of things.

I think you'll note in our materials that we highlighted, non-GAAP operating margin. And the reason that we did that is because, again, if you think of where the industry is going and how different this business is than the business that we had several years ago, gross margin is one input, but operating margin is a better, more holistic view of what's going on in terms of the overall profitability because the revenue growth is so rapid and it's certainly at a lower gross margin, but it comes with tremendous operating expense leverage. And so the operating margins are very strong and very compelling, and we expect those to continue to be very attractive.

Will we see a small decrease in operating margin as we invest to accelerate our growth, given some of the same timing-related issues with bringing on new capacity, we certainly will. But if you look at the rate of revenue growth, if you look at the strong operating margins, if you look at the fact that we've been very, very disciplined with cash flow, and that we're earning very good returns. I think you'd agree that we're positioned very, very well for very durable and profitable growth.

Your next question comes from the line of Patrick Walravens of Citizens.

It's amazing results you guys, congratulations. So Paddy, when I was at your Deploy conference, the speaker got interrupted by applause like five or six times. But two of the times were when you talked about the inference router and then also when you guys talked about support for the latest DeepSeek model. So can you just talk a little bit about why your customers are so enthusiastic about that?

Yes. Thank you, Patrick. And first of all, thank you for coming to deploy last week. So you bring up a really, really important point. And for those of you who have not seen the keynote video recording from last week, I encourage you to please do that.

The two points that Patrick just mentioned are really important because AI Natives are doing something which is incredibly interesting. Number one is they are all running multiple models, right? Because as I mentioned, this is a cost of revenue line item for them, and it will be crippling if they are just beholden to one closed source model. Last week, there were two different models that were announced. One is DeepSeek version 4 and the other one was the latest version from OpenAI.

And the difference in price was 10x. In terms of the output tokens, it was literally $3 versus $30. So AI natives are doing three things: One, they are all becoming multiple models. Number two is they're running a lot of open source. And number three is, many of these AI natives are also running their own version of a model, which is distilled from an open source model or something like that. So there's intelligent router becomes extraordinarily important so that the router can find the right model for the task you're assigning.

So we showed a demo, which was super compelling where it showed better performance at lower [ TCO ] per token by routing the incoming prompt to the right model.

And the second thing is Patrick mentioned that there was a lot of supplies for our DeepSeek support, which is fairly obvious because AI natives are embracing open source up and down the stack in a very pronounced manner. So that's why it is really important to understand, our target market is very different. These are AI natives that are building and monetizing software and for them, multiple models, open source and having destiny over their intelligence is an existential thing.

Great. And Matt, if I could ask you a follow-up. Cursor is an amazing win, congratulations. We've all seen the news about SpaceX having an option to buy it. So just how did that fit into your guidance? How did you think about that?

Cursor is a fantastic customer. And as you said, it's a great indication of the quality of the platform. And we're really excited by it based on the fact that they're using -- this is not a Bare Metal contract. They're using our inference services. They've made commitments around the NFS and some of the core cloud capabilities, so we're very encouraged by that, and we have a fantastic relationship with them.

We haven't predicated any of our long-term guidance on any single customer. We have, as Paddy said, to the demand for the capacity that we have available and we were very confident that there'll be a good part of that, but we're not basing any of our forecasts on specific customer.

And your last question comes from the line of Raimo Lenschow of Barclays.

Two quick questions. Going back to Gabriela's point in terms of like how big the market is. At the moment, it looks like most of the work is getting done on training models and inference is only starting. Like Paddy from your perspective, which innings are we on inference actually because it seems very, very early still to get an idea about like how long this can go on for. And then, Matt, for you, the one thing that comes up in the market is a lot of like capacity of new data centers, et cetera. You're not building 100,000 GPU to have data centers who are much smaller, but like what's the constraint of finding sites to kind of go beyond the capacity you announced today?

Thank you, Raimo. So to answer your question succinctly, since baseball season is just starting. I would say from an inferencing point of view, we are probably in the top of the second inning. And Agentic, we are just in the national anthem. It's just getting started. So I think there's a lot of room for a lot of innovation. And I am the one thing that I'm super proud of with all the announcements we made last week is 15 new product launches, not just features, 15 new product launches and the velocity and the intensity from our engineering team is just -- it's going to make a difference in terms of our ability to establish a leadership position.

And then Raimo, your -- what was the second question? Second part of the question?

[indiscernible] how did is it like, yes?

Sorry. The -- we've been able to secure the data center capacity that we've been targeting. We're still in active conversations on additional capacity beyond the both for '27 and '28. And we've not had an issue getting capacity that we've been trying to track down.

That concludes our Q&A session. And this also concludes today's conference call. Thank you for your participation. You may now disconnect.

Perfect. Thank you. So thank you, guys. Thanks, everyone, for attending. Really excited. My name is Mark Zhang. I help cover cybersecurity and infrastructure for Citi Equity Research. Today, we have the pleasure of having Paddy Srinivasan, CEO of DigitalOcean. Thank you so much for attending, Paddy.

Thank you, Mark.

Really appreciate the time and insights.

So maybe just to kick it off, Paddy, you've been with the company for, call it, 2 years and change. Been really just very focused on turning the company around in terms of profit portfolio go-to-market. Just want to give us a sense of what's changed over the past 2.5 years? What stayed the same? And how you're thinking about like going after this AI [Technical Difficulty]

Yes. Thank you, Mark. First of all, it's wonderful to be here. This is turning out to be quite the conference, a lot of energy, a lot of familiar names. It's just amazing, truly amazing the times we are living in, and it's an honor to be here. Yes, it's a little over 2 years that I've been at DigitalOcean. It's been a lot of different things, some planned, some unplanned. But when I joined the company, I've been in the developer ecosystem for 30 years. I started my career at Microsoft and did a few different things, including a start-up also in developer ecosystem. So when I got the opportunity to engage with DigitalOcean, it was -- I was super excited because we are a relatively young company. We're only like 12, 13 years old, 5 of which has been in the public market, relatively young.

And over the first 10 years of the company's existence, we built an iconic developer cloud. Like developers absolutely love us even today [Technical Difficulty] with developers. We have 650,000 paying customers, and it just keeps going, and we get tens of thousands of customers signing up on a monthly basis. So that machine and that goodwill and the credibility we have with developers is a constant.

What has really changed? Number one is when I joined the company, we were approximately $600 million of revenue. And there was one big problem that the company was facing, which is our top customers were unable to scale with us to go take their workloads to other hyperscalers to be able to keep growing and scaling their more sophisticated workloads. And that was my #1 focus to fix because at the scale that we were in at $600 million, it was really impossible for us to keep our growth rates accelerating [indiscernible] losing customer acquisition energy to [indiscernible]

So that was the first thing that I wanted to fix, and we fixed it by really focusing on the fundamentals of our product. And for the first 4 quarters, I was repeating the same theme. This quarter, we shipped 40 features, 50 features, 60 features. And here's the impact, here's the adoption. And then slowly that turned into business outcomes. Over the last -- in the last earnings call, I talked about how we've taken what was a questionable even a weakness of the company, and we have turned it into an absolute strength of ours. Our top customers are our growth engine now. Our million customers -- $1 million-plus customers, $500,000 plus customers, $100,000 plus customers, they're all growing significantly faster than the market growth rate. Our $1 million customers are growing in triple digits. Our $500,000 and $1 million customers had 0 churn over the last 4 quarters. And by any measure, they have become a big reason why our growth is reaccelerating. And that was purely a function of [Technical Difficulty] mention, Mark. We made a bet on AI. And when I was getting started 2 -- generation ago in terms of AI, it was just getting started large training clusters. We made a conscious decision to not go after that primarily because it was really not aligned with who we are as a company. What we know how to do is build great software, take very complex concepts, make it simple, accessible, affordable, predictable for our customers. And inferencing -- when inferencing started emerging, we felt like it was really in our wheelhouse. It played to our strength, and that's why we decided to double down on inferencing over the last 6 months or so. And the results are here to see.

In Q4, I announced $20 million of AI customer revenue, growing at 150% plus year-over-year and just getting [indiscernible] And 70% of our AI revenue is non-bare metal, which is the exact opposite of all the other Neo clouds. You see it's mostly bare metal services, whether it is mostly training, but also inferencing is all bare metal. For us, it's the exact opposite. And it really enables us to build a very different financial -- as we scale our AI track.

Great that you brought in the AI interest and opportunity. I think the AI interest term is around the industry quite I think if we really parse through the various levels of inferencing, right? There's various opportunities. How do you see sort of DigitalOcean's opportunity in inferencing? How do you differentiate yourself from, let's say, the other providers? Even outside the Neo cloud. The ones that are trying to put in more managed services and now AI inferencing. How does DigitalOcean play? And how do you sort of see the longevity of your position within just inferencing?

Yes. Great question, Mark. So I will answer this question purely perspective as a platform provider. So I mean, obviously, there are different types of customers, starting with individual users, we all use -- for OpenAI and things like that in our daily lives or open clock. So that's one set of users. The second set of users are enterprise companies like a Citibank or Walmart or Eli Lilly, those kinds of -- or a restaurant chain. Those are end user enterprise companies. The third category, which is our ideal customer profile, our target customer profile are cloud-native companies or AI native companies. In the old world, they used to be called independent software vendors or ISVs [Technical Difficulty] building and monetizing technology. That is their business model. So we are [indiscernible] only on the last segment, which is our [ 600 ] majority of them digital native enterprises or cloud-native companies. And now [Technical Difficulty] revenue from AI native companies that are using our technologies to [Technical Difficulty] It is really important to understand that.

So in this category, there are 3 types of [Technical Difficulty] business, and they're trying to introduce or they're trying to have a heavy dose of intelligence in their existing application, whether it is a SaaS application or something else. That's category 1. Category 2 are AI-native companies that are reinventing a category with an AI-centric approach. So this could be a new travel management application, not Concur, but the next generation of Concur that is built using an AI-native approach, right? So that's the second type of company -- company that says. And the second type is still having humans as the primary persona with a heavy use of intelligence to support the human workflows.

The third type of companies is what I call as agent-native companies. These agent-native companies are just emerging, and they're taking a very different approach to say, why do humans need to be involved for a travel management application as an example. Agents can do 90% of the work and 10% of the work will require a human in the loop. So for all those 3 types of customers, they're all intelligence hungry. They need a lot of inferencing, and that's who we are. We are an inference factory, and I'll get into the details and you ask me how we differentiate and things like that. I'm happy to get into the details, but we are an inference factory.

Can you get inference tokens by renting out a GPU and building everything? Sure, you can. You can also visit a lumber yard, buy timber and build your own cabinet. You can do that, too. But you come to somebody like us to get smart intelligent inference tokens so that you can focus on building your travel management application rather than having to go manage the life cycle of a GPU and things like that. So we are a platform provider that produces intelligent inference tokens compared to a Neo cloud that is still in the business of renting hardware.

That's a great overview. I think like certainly, the space is evolving and even the Neo cloud are providing or creating their own managed services and layering on top of the GPUs. So how do you sort of like think of the competitive landscape in terms of -- how do I protect my mode? How do I sort of like continue on with providing this platform as of [Technical Difficulty]

So before I answer that question, Mark, let me explain to you with these kinds of new generation of applications, how is that translating to a new stack, right? Because when I look at the stack that we have built over the last dozen years, a lot of it is applicable in the new world, but the new type of applications are also demanding a complete relook at the stack that we are providing, right? As these companies scale. One very, very important thing we have to internalize is for an AI native or a cloud-native company that is building an intelligent application, they are carrying the cost of inference in their cost of goods sold. It is a COGS line item for them.

And if they are not mindful, they can very quickly get into a negative gross margin territory. Like there's no business to be scaled, right? I think that's a really important distinction compared to a personal use case where you can put some guardrails on how much you want to spend and things like that or even an enterprise that is spending, like, let's say, I have a $100,000 employee, you can say, I want to spend $20,000 to augment that human employee with tokens and things like that. But for an AI native company that is building and scaling an intelligent application, this actually sits and drains your gross margins. And if they're not careful, it just -- there's -- you cannot build a business, right?

And we have a lot of companies coming to us saying, "Hey, we started, we wanted to nail our product market fit. We started with a closed model, and we have figured that out. Now I'm scaling. And if I don't do this, I'm going to be -- I don't have a business to scale. And we have this combination of open source models and open source harnesses, and we're trying to stitch all of these things by hand. Can you help us automate some of these things?" This is how -- 6 to 9 months ago, this was the starting point of our new inference cloud is. We started observing and working with companies that are trying to do this at scale with -- in a post-product market fit stage. And that's when we said we have to really rethink how we are doing this. And we started investing very heavily in a couple of different layers. So at the bottom of layer, we have infrastructure just like any other Neo cloud, right?

On top of it, we have a fairly sophisticated set of compute primitives that sit on top of GPUs, CPUs and storage and things like that. So we have what we call as the GPU droplet, which is a virtualized instance of a GPU. We have core compute. We have many other artifacts that you would expect from a real cloud. But then we have built 3 other layers, which are very new in this new world. First of all, intelligent applications start and end with feeding real-time data. So this whole concept of, oh, I'll generate a lot of exhaust and capture data, and then I'll pass it into a different system where I'm going to do an ETL and get the data prepped and I'll do all the processing and then feed it back to the application, that is dead.

Like we now are having customers that want real-time access to analytics and data fed into a reinforcement loop so that RL as a service is improving their application and the model that they're working on in real time. So there is a reinvention of the data layer that is happening right as we speak. The layer above that is what we call as the inference engine.

Inference Engine has two parts. One is we have a very rich list of models that we host. And the models are both open source models that we do pass-throughs and enrichment of, and we also have dozens of open source models that we offer day 0 support for.

And what does the day 0 support mean? We have the ability to take these open source models and optimize them at a kernel level to make sure that the efficiency of tokens is top notch. So when you are paying for tokens, you have to get not only the best quality tokens, you also need to get the best efficiency of tokens. And that is an opportunity for us, right? So that's on one side.

The other side is there are dozen or so services for an inference engine, including things like quantization because you're training your model using FP16, but when you're running it, you're doing FP4. So you have to quantize the model. You have to do -- you have to evaluate different types of models at real time. You need to do guardrailing of your models. You need to have many other primitives to make sure that your models and your -- the quality of your inference output is really top-notch.

And then the final layer that we are working on is the agent life cycle layer. Agent run times, agents are very idiosyncratic in the sense that you cannot -- you just simply don't have the time to hydrate and dehydrate virtual machines. You need something a little bit more lightweight. So you have this emergence of this new thing called agent sandboxes and so on and so forth. And agents also need persistent memory. They also need memory and things like that. So there's a whole suite of capabilities to run agents. So these are the 5 layers that we are inventing as we speak, working with our customers to serve the needs of this new AI native application stack, right?

So I'll come -- I'll now answer your question in terms of how are we different? We are different because a Neo Cloud has the bottom most layer. They have phenomenal infrastructure or data center operations, and they have GPUs and CPUs available for rental. But then if you are a company that wants to build let's say, a new health care application with AI native or AI-centric health care application, you have to go reinvent all of the other 4 layers.

Even if you are one of the inference providers, you get one of the layers, but you have to go build and scale your data layer on AWS. You have to go to a provider like DigitalOcean to get access to raw GPUs or CPUs because most of these agentic applications need GPUs, but they also need a lot of CPU firepower to be able to preprocess and post-process workflows.

So there's a lot of incompleteness with the Neo clouds and the inference providers in the market that they have very small pieces of the puzzle, but not [indiscernible] The hyperscalers surely have it, but the hyperscalers are busy building it for the city banks of the world where it is really hard to do that and also make it so simple that an AI native company can come and start consuming it in a matter of is what we try to do. So a long-winded answer to your question, but the [indiscernible] inferencing and agentic applications, we are reinventing the stack because the old stack as important as it is, it is simply not enough to be able to build and scale these new types of [Technical Difficulty]

That's very [Technical Difficulty] So across 5 layers opportunities, maybe we could like think about it from a product road map standpoint. Is it -- where do you see sort of the low-hanging fruit of media like demand and opportunities for you to expand your TAM and expand the opportunities with your current customers? Is it -- should we think of it from just a bottom-up approach? Or do you think from a product road map standpoint, where the key opportunities are if [indiscernible]

Yes, great question. So we see the opportunity as we are in a very secular tailwind situation here because it doesn't matter what type of application is being built. Future of all technology is going to include a heavy dose of intelligence. And we are in the business of providing tokens, intelligent tokens and efficient tokens. So if you think about the currency of new types of applications is token-based. It's no longer seat-based. And a lot has been written about that. So I'm not breaking any new news here. But what is important is that the business model is also fundamentally changing from -- for us, it is changing from charging for inputs, which are GPU dollars per hour to changing for output, which is tokens and the quality of tokens.

So as a platform -- inference factory platform provider, for me, my margins are going to improve if I can figure out how to produce more intelligent tokens because I can monetize them significantly better. And I can get more token efficiency and token throughput for the same GPU I'm deploying compared to my competition. So those are the 2 things we are really focused on.

Now coming to your question, Mark, where do we see the most opportunity? We see the most opportunity as the inference provider for tech companies that are modernizing their application, injecting intelligence. That's number one. Number two, we see an entire ecosystem that has already emerged in Silicon Valley that are very compute hungry and I should say, inference hungry and more and more demand that we are seeing today is for inference token consumption rather than GPU consumption.

So I'm beating up on my sales team to say, as soon as when a contract is up for a GPU, shut it down, move it to my on-demand and inference pool so that I can monetize it in a very different fashion. So the fact that we don't have a lot of super long-term contracts is an absolute blessing for us because I can take that pool and redeploy it and monetize it in a very different way and using a very different currency.

The third thing is we are seeing a lot of traction starting to appear on the agentic space, where agents are not only intelligence hungry, they're also compute hungry because the nature of agents is that they're autonomous and they're goal seeking in the sense that they will continue to go in infinite loops until they get the job done, which requires intelligence, yes, but they require an enormous amount of compute. And I mean, obviously, we've all seen the announcement from Claude on managed agents and things like that. And I have a point or two to make on that, Mark, if you will allow me.

That's the next question.

So because there's been a lot of questions around it, right? And I've been doing this for 30 years, so I've seen a pattern or two. And I want to start by saying closed source systems validate the category. Open source systems scale the category. So it is fantastic that Anthropic is validating that, hey, agentic apps are here, and this is how you deploy it. If you're an individual user, you're obviously -- that's a great place to do it rather than messing around with an open claw and things like that. Even if you're an enterprise and you are willing to make a bet on a closed ecosystem, that's also great. But our target customer profile is somebody that's building a business on intelligence. And it is -- as we have seen over the last 30 years, it is really hard for an ISV or an AI-native company today to bet their business with a closed ecosystem. Our top 10 customers, literally nobody is a single model company. Literally, everyone is using some closed source models, but a lot of open source models. That is how these AI native companies are scaling.

And that, to me, is a great opportunity, whether you're building an agentic system or just developing a new AI native application with a lot of inferencing or you are an old school company that is trying to inject intelligence into your application, all of them are inferencing hungry. All of them are compute hungry. And I feel like I'm an arms dealer in this massive replatforming that is happening, and we cannot provision compute and intelligence fast enough to serve the needs of all of these use cases.

Agentic workloads are just one of the use cases, and they need a lot of intelligence. They need a lot of compute and a lot of agents are not going to be built within the 4 walls of a closed ecosystem. There's going to be a tremendous amount of open source across the 5 layers that I was talking about, right? There's open source in every layer that we support, and it is amazing to see the emergence of open source in pretty much every aspect of this new computing stack.

Yes. No, that's terrific. Thank you for densifying some of the Frontier lab announcements. Maybe to that point, we're maybe, I guess, at the realm of the next frontier lab announcement. But from your purview and what you see from what they've done so far, what do you see sort of the risk to your, call it, the 5-layer stack of the inferencing opportunity? Yes, there's open source [Technical Difficulty] risk that you see Frontier Labs making in terms of like their splash within that inferencing opportunity?

Yes. So the Frontier Labs announcements or generally the model threat is at a layer above where we sit. We're still in the business of providing essential infrastructure. The application space is where I see most of the disruption happening. Will some of these model companies also announced different parts of the inferencing stack or some of the infrastructure? Sure. But to me, it is all great validation. So for example, my first job out of college was working at Microsoft writing kernel level code for Windows NT operating system. And if you remember, like early -- fast forward a few more years, early 2000.

[indiscernible] you're miked up and ready to go. [Technical Difficulty]

[Technical Difficulty] Servers had a dominant market share. But another 5 years or so, Linux started taking off. And today, I think it's -- 90% servers have Linux. [indiscernible] Ecosystem [indiscernible] how do you think of the build versus buy [Technical Difficulty]

Yes. So that's a great question, Mark. So we definitely look at [indiscernible] 3 dimensions. [Technical Difficulty] We are obviously building a lot of what we are shipping today. We partner [Technical Difficulty] And on April 28, we have our deploy conference here in San Francisco, and we'll be talking about a tremendous amount of [indiscernible] for the last 6 months that we'll be announcing. And like you said, Mark, we just did an acquisition of a technology company called -- And we will [Technical Difficulty] for ways by which we can accelerate our product road map. Like we cannot ship these features fast enough. Literally, our top customers are snatching these features from our hands and because they don't want to sit around messing with infrastructure.

So I'll give you one great example of the company that we just acquired, they have this technology called Plano AI. Essentially, for the last 6 months, we've been working with some of our customers. And this is why I keep going back to the fact that none of our top AI native companies are single model. In fact, we have many companies that have clusters of models that are looking at incoming prompts. We have a company that uses 31 different models to process a single prompt. And that is not an outlier, by the way. We -- most of our companies -- most of our customers have at least half a dozen to a dozen models that are processing incoming -- single prompt is processed by a cluster of models, right? And then we have some companies that are saying, hey, can you actually do this routing of traffic to the right model at the right time for a combination of accuracy, cost, latency and throughput? Can you come up with some heuristics either that we can -- actively or you come up with dynamically at run time in a matter of milliseconds and route it to the right model at the right time because inference quality matters, inference efficiency matters even more.

So these are kinds of features that we started building and then we came across this piece of technology. We quickly moved to acquire them. So we are always scanning the market to partner with great companies and eventually, if things work out and if it is the right thing, we'll acquire as well.

Terrific. That's great to hear. Maybe just to wrap it up and ask a final question, I'm sure all investors are interested in this question is how does the -- economy change monetization? And how do we think of the unit economics and the opportunity?

Yes. So from our customers' perspective, as I said, they're moving into a [indiscernible] based [Technical Difficulty] smarter, no credit to us, like we haven't done anything, but we are very comfortable with consumption-based pricing model. That's what we have done for the last year. As I was mentioning, the -- we are moving from an era [Technical Difficulty] per hour. The vast majority of products are moving more and more towards an outcomes-based like token-based pricing so that they can market up and companies are consuming tokens, enriching the tokens and providing outcomes to their customers. So as long as we are in the path of value creation and align our unit of consumption and our business model to the unit of consumption and the business model of our customers, we're going to be in a great shape. But as I said, we have had 12 years of experience being in this consumption-based pricing model. So I feel really good that finally [Technical Difficulty]

All right. Before we begin, for important disclosures, please see the Morgan Stanley research disclosure website at www.morganstanley.com/researchdisclosures. And if you have any questions, please reach out to your Morgan Stanley sales representative.

Hello. I'm Josh Baer, software analyst at Morgan Stanley. We are thrilled to have the DigitalOcean leadership team here. We have Paddy Srinivasan, CEO; and Matt Steinfort, CFO. Thank you so much for joining us.

Thank you. Thank you, Josh, for having us here.

Excellent. So Paddy, I was hoping you could start it off talking about your strategy. You came in a few years ago with a very clear product and go-to-market strategy. I want to check in on where we are as far as those initiatives, your focus with large customers, your AI strategy and what's evolved.

Yes. Thank you, Josh. Yes. It's been 2 years since I've been here, and we just crossed a big milestone, which is $1 billion ARR in December. So it was a natural time for us to take stock of how far the company has come. It's an incredible story. Two years ago, when I joined, the company already had a phenomenal foundation of having built a very iconic developer cloud. But what was missing was as these developers started scaling up their footprint, we were missing some very critical enterprise capabilities, which prevented them from -- prevented us from scaling with their needs. So that was priority #1 for me is to plug all these gaps and make the platform enterprise ready. So that was number one.

Number two is as these companies were scaling, we had to reinvent our go-to-market to be in service to them. And then the second big pillar was 2 years ago, AI was just emerging. And as an infrastructure provider, we needed to have a strong platform, a strong story in the world of AI. And we made a very conscious decision. We could have chased the training world of AI, but that would have meant that we would have to reinvent ourselves and make ourselves into a GPU form and become a landlord, which would have been a reinvention of the company or what we decided to do was to lean into our strength, which was we were really, really good at software. We are really good at capturing mind share of developers. So we said we are going to focus on inferencing. And that's the second pillar of our strategy. And 2 years later, I think we have done a really nice job of executing on both those things. And for me, strategy is all about not just being clear on what we are going to do, but also on what we are not going to do. And we have been very, very focused and disciplined on these 2 pillars.

Great overview, and we'll dig into a lot of that. I think it was last week, Time is an interesting concept. You announced Q4 and had an investor update with a path to 30% growth. Could you unpack some of the key takeaways from that combined results and investor update?

Yes. So Q4 was a phenomenal quarter for us. And -- it was a capstone quarter for what was a very defining turning point year for DigitalOcean. So one of the things that we announced last week was that we brought in $51 million of incremental ARR, the highest organic ARR in the company's history, right? So -- and we talked about 4 key takeaways. Number one, the top customers that I was just talking about was once a constraint. Now it is our growth engine. So we had phenomenal results from our $1 million customers, 500,000 customers and $100,000 customers. Our $1 million customers are growing at 123% year-over-year, and we have had 0 churn in that cohort for the last 4 quarters. So we took what was once a constraint and really made it into our growth engine and a strength of ours. So that's number one.

Number two is, we are hearing and seeing a lot about how software is eating software. AI is disrupting all types of software. And as an AI infrastructure provider, we are on the right side of this disruption. We are equipping both cloud natives that are defending their territory and AI-native companies that are the insurgence trying to capture markets in whether it is horizontal SaaS or vertical SaaS and things like that. And we showcased many customer examples that we have won over the last 90 days. So that's number two.

Number three is, how are we doing this? We're doing it with a very differentiated software stack, a stack that includes not just core cloud. It has a very robust and diverse lineup of GPUs. But the most important thing is we have built a full stack inference cloud capability on top of all of this. So this is what is helping us drive AI customer revenue of $120 million, growing at 150% quarter-over-quarter consistently for the last several quarters. So that's 3.

And we are doing all of this in a very responsible manner, right? And you mentioned, Josh, last year, we finished Q4 with 18% growth. We guided for 21% growth this year. We said we will exit at 25% in 2026 and guided to a 30% plus growth in 2027. And we are going to do all of this with a rule of 50 plus. And when we said 50-plus, people started crunching numbers and they took 50. So we are reiterating it's going to be 50-plus, and we will do it profitably. So that's the balance and the responsible investments that we continue to do.

And Paddy, you're talking about line of sight to 25% growth at the end of the year and then even growth into 2027. Like what gives you the visibility to look into '27 for the full year? How much revenue is coming from existing bookings, existing contracts? What do you need to secure to get to those?

Yes. So we are in a very unique position when it comes to inferencing, right? So we are -- RPO is never a thing for us. But in Q4, we announced very robust RPO, which grew 500% year-over-year. And it was double from the quarter previous. And -- but what is important to us is the demand that we are seeing from cloud-native companies, some of them are here in this conference. If we were to give all of our capacity to the first customer that asked us, we'll be -- we can take some time off for the rest of the year. But that is not our business model, right? It will look great on paper, but what we want to do is let a few dozen flowers bloom in our inference cloud because these customers are really taking market share and they are disrupting the software landscape, and we want to be part of as many of these stories as possible.

So we are not going to get carried away by what the other training clouds are doing in terms of announcing 1 customer or 2 customers and they're sold out of capacity for 4 years. versus we have a very different business model where we want to have our platform be used by several dozen AI native companies that are experiencing hyper growth. And we work very diligently with our prospects and our customers to make sure that we can provide capacity on demand, help them scale and do some of them take big chunks of our capacity? Sure. But what gives us confidence to guide what we guided to, by the way, with only existing committed data center capacity alone, we can grow in excess of 30% next year. And it is all stemming from the demand and the pull we are seeing from the market for our inference cloud.

Excellent. And I thought it was really helpful to isolate, all right, bring on 31 megawatts and that gets you to the 2027 revenue. But like one of the topics of conversation I've been having with investors is, well, that's not -- it's not like you're going to just stop adding capacity and add nothing in '27. So what's the impact to financials? And so maybe to bring that into the conversation, I mean, what framework would you suggest like provide to investors just thinking about what's going to be the pace of incremental capacity looking forward beyond '26?

Yes. I would start with a couple of things. One, we very intentionally guided that way, right? The company just a year ago was growing 11%, 12%, 13%, pretty stable, not a lot of growth capital in the company. And as we're accelerating, that changes the dynamic quite a bit. And so the bringing on of 31 megawatts of incremental capacity this year is like a 70% increase in our total capacity. And it causes lumpiness, right? It causes lumpiness around the margins. It causes lumpiness around the equipment that you need. And the thought was, well, shoot, one, we haven't committed to any incremental capacity beyond 31 or we'd tell you. And two, we need to give the market a clear view of what does it look like once you reach some level of kind of steady state with that capacity where that capacity is, I'd say, healthily utilized. And that's the way we guided.

And we actually posted an investor supplement in the -- to our investor website this morning describing a little bit of those dynamics around the impacts on cash flows and the impact on margins, the impact on leverage and how equipment financing kind of impacts that. So I recommend everybody to take a look at that. But then I'll come back to the, okay, what should you expect going forward? As Paddy said, demand is already well in excess of supply. You take that and you couple it with the fact that to get data center capacity, you have to be planning 12, 18 months in advance. So we're already actively in conversations, and we have been in conversations with potential data center providers talking about '27 capacity, talking about '28 capacity. We'll certainly share more information when we've locked down some plans and we can share what the resulting increase in growth would be because that's the big question is, well, how much are you going to add and what is the economic of that once you add it? We tried to give people the blueprint with this 31 and been very transparent about how that's going to impact our financials so that hopefully, the market you can get your models geared up.

And when we come back and say, okay, well, now we've committed to some incremental capacity, you can just flow that through and see how that works. So very excited about the growth potential, expect to add incremental capacity, expect to communicate more on that in the coming months.

Makes a lot of sense. And how should investors think about financing all the equipment for that incremental capacity? Maybe we could talk about 2026, what is committed, but then on a go-forward basis as well?

Yes. No, that's a great question. And we've gotten a lot of questions about the dynamics of equipment financing, and we can talk about that in a second. But if you just think about the quantum, one, we don't guide to -- we never guided to CapEx or to the amount of equipment that we would need to put in. It's a super lumpy metric if we bring CapEx or equipment on and the last week of December shows up as the full year number, even though it's clearly for ' 27. So that's not a spectacular metric that we guide to. But if you just think about order of magnitude, we spend, call it, and this varies, and I'll explain why it varies in a second, call it between $20 million to $25 million on equipment per megawatt for a new data center.

And you're like, why did you get to that number? Well, it's not just GPUs. We don't just put GPUs in our data centers. We're a full stack cloud. We put full core cloud capabilities in there with our general purpose cloud with storage and database. There's networking. There's other gear that you need to put in there. And then the amount varies depending on what kind of gear you put in. So if you put in some of the NVIDIA latest gear, it's more expensive than some of the latest AMD gear. So it depends on what you put in there. But I think order of magnitude, $20 million, $25 million. And so while we don't guide, you can take the incremental megawatts that we've committed to and you can do some math. And don't forget to add a little bit of just general purpose cloud kind of growth. The rest of our network is still growing. Our top customers are growing. So you can kind of back yourself into what a gross number would be.

But to me, the gross number is not as important as well, what are the long-term margins you're generating? And what should you expect from cash flow generation over time. And that's why we've added the guidance that we have to give that clarity.

That's really helpful. Could you talk a little bit about your decision process or your framework for determining when to use cash to buy this equipment, when to issue debt, when to enter into equipment finance leases? How do you think about the...

Yes. That's a great question. I think people give us and maybe me too much credit. They're like, that seems like a really complicated like sophisticated financial structure. And I'm like it's either paying upfront for equipment or it's paying for it over time. It literally -- it's that simple. If you think about -- we have an option of -- and I'll just make a number up. If we're going to put in a bunch of equipment, it costs $100 million, we could pay for it right now and you say, well, how did you pay for it? Well, we use cash, but we had that cash because we had borrowed money and we have a TLA and it's got a little bit of interest on it. So arguably, you're paying upfront for that equipment and you're paying interest on that equipment because you've borrowed money.

The alternative is, well, I can pay for that over 4 or 5 years, and I'll pay a similar amount of interest because the interest rates aren't that different. And everything else is the same. So we still own the equipment at the end. We still operate the equipment. There's nobody operating the equipment. We're not outsourcing that to anybody. We're running the equipment that's in our facility. It shows up as debt for leverage purposes, exactly the same, either that TLA amount or the total obligation that you have on the -- for the liabilities and paying off the principal. It's pretty much the same. So you said, well, why would you do that? Well, it's way better for us as we're scaling to align the investments we make with revenue, right? We'd much rather pay 20% or 25% of that each year because we get revenue that covers that, and we're generating cash on the back of that from year 1. If you take it all upfront, it just -- it's a limiter. You can only take so much because you're going to burn a ton of cash and then you pay for -- you pay it back over time.

So again, it's not sophisticated. It's fully transparent. It's visible. It's on balance sheet. It's -- there's no real friction. It's just you're paying over time instead of paying upfront.

So we've been fielding a lot of questions around this topic and free cash flow targets. Could you unpack or maybe bridge between the unlevered free cash flow guidance and targets, levered free cash flow? And I think maybe you've covered CapEx, but anything else with that bridge?

Yes. So when we were, again, growing 11%, 12%, 13%, not a lot of growth capital, didn't really have any leverage to speak of with a 0% -- 0 coupon bond. Our reported adjusted free cash flow was a pretty simple metric, and it didn't have a lot of CapEx in it, not a ton, and it didn't have any interest in it. So people could use it as a proxy for our unlevered free cash flow. As we grow, though, that becomes a less useful metric for valuation purposes because you start to put in a bunch of capital. Clearly, you don't put in -- you don't include all growth capital in your multiple of free cash flow because you'd be digging us for all of the investment we're making now times the perpetuity value. And you also clearly don't include interest.

So we said, okay, well, we need to start breaking this out and providing more visibility. So we introduced unlevered free cash flow. Unlevered free cash flow captures the normal cash from operations. It captures the CapEx that we do spend if we pay for something upfront. But it's not a perfect metric because it doesn't capture the principal payments associated with leases. So that's fine. The guidance for the unlevered for the year for '26 is 18% to 20% growth. If you take out the lease payments, if you say, I'm going to burden you with lease payments and you take that out, that number drops to about 12%. And you say, well, is that a good metric to use? I'd say it's still -- you still have valuation challenges you have if you're using that metric. One is if you're putting a multiple on unlevered free cash flow minus principal payments, well, you're double counting the debt because we're already showing the debt as -- in that obligation to those future payments.

And the second thing is, again, you're still putting a multiple on growth capital, which is probably deserving to be treated differently. You can then add leverage on that, which is to get to kind of a fully levered with all cash payments metric. And we've disclosed that in the materials that we posted today. I think the challenge that we have and you'll have as investors valuing the company is you can't just take the old free cash flow method and just apply the same kind of multiple to it. You have to tease out. We've got growth capital that you have to be able to value for the potential revenue creation that's going to drive and profit creation, and you have to make sure you're taking out the leverage.

And so we're just trying to show all the different components so that people understand what's in there and they can value us however the market wants to value us.

Perfect. Let's shift the conversation back to Paddy and talk about the business. You framed sort of your AI strategy and focus in the opening remarks. Can you go a step further and really lay out your competitive differentiation?

Yes, sure. So when I look at our competitive differentiation, so first, let me start with what we currently have, right? So in the earnings deck from last week, we had 2 slides that I want to refer back to for those online. So there's a Slide 19 where we show our full stack and Slide 20, which is a Harvey Ball comparison chart with us and other providers. So our stack has 3 major components. One is the full stack cloud, which we have gone through this cool of hard knocks over the last dozen years, building a full stack cloud, running it in 20-plus data centers across the world. It is not -- you cannot wipe code a full stack cloud platform, right? There's a lot of sweat equity that goes into building and operating a full stack cloud. So that's number one. And the full stack cloud has the obvious stuff like compute network storage, Platform as a Service, database as a service, orchestration with Kubernetes and whatnot.

So it's a full stack cloud comparable to the hyperscalers. The second piece that we have is, of course, the GPU infrastructure. So -- our GPU infrastructure is slightly different from the ones that you may find from GPU farms or the neo GPU clouds in the sense that our GPUs are purpose-built for inferencing. Even this morning, we announced a very deep technical paper talking about how our inferencing scaled up for a public company called Workato, where they're achieving more than 2/3 cost optimization and almost 80% reduction in the time for first token and things like that. So the point is our GPU infrastructure is purpose-built for inferencing. And we do a lot of different things optimizing GPUs for that. The third thing that we have is our inference engine.

The inference engine is -- starts with a lineup of all kinds of leading models, open source and closed source. We have kernel optimizations and optimizations that enable inferencing customers to get the best bang for their buck, and they measure this in 4 ways. One is the throughput rate, low latency, high accuracy and the best TCO. These are the 4 things that matter when companies go into inferencing mode. And we have a bunch of artifacts and modules that deliver these 4 things, right? And then we allow our customers to come into our platform whichever way they feel comfortable. Some customers say that, "Hey, just give me raw GPUs and just have some orchestration on top of it, we got the rest."

Most customers are increasingly starting to prefer other ways of entering the platform. For example, serverless inferencing. They just want a bunch of models to be available to them using API endpoints so that they can focus on their business and not managing the infrastructure. We have dedicated inferencing clusters. We have run your Python code in a container, but consuming inferencing endpoints. So we have multiple ways of consuming our inferencing infrastructure. So those are the 3 big clusters, right? So you have the core cloud, we have GPU infrastructure and then we have the inference engine. And when I look at the competitive landscape, the only class of competitors that even have the breadth and the depth of what we offer are the hyperscalers. Of course, we've been competing with the hyperscalers for the last dozen-plus years, and we win our fair share. We build a $1 billion business by winning a fair share of those customers that prefer simplicity, lack of open standards and lack of vendor lock-in that is a big angle for us.

And number three is predictability and transparency of pricing. That is very unique to DigitalOcean. So that's how we win against hyperscalers. But when you look at how we differentiate ourselves with the neoclouds or the inference wrappers, they typically have 1 of the 3 things I talked about, right? neoclouds have GPU forms. They don't have a full stack cloud. They don't have an inference engine with a software differentiation. The inference wrappers have the inference engine, but they don't have a full stack cloud and they typically come to people like us for GPUs. So they have 1 of the 3 pillars that we have. So I feel what we have built already is very sophisticated and we have a lead in the market when it comes to inferencing. And that lead is only going to keep expanding.

On April 28, we have our deploy conference here in San Francisco. We are going to lift the covers on a lot of the things that we have been busy building, and it is going to increase the lead we have competitively with all of our competitors. So super excited about what we have already built, where we are, and that lead is only going to keep increasing.

Excellent. Kind of related to some of the competition, I mean, you provided some transparency into your ARR per megawatt, which is around $22 million currently, and you're expecting it to sustain around $20 million per megawatt even with the incremental 31 that comes on board. And this -- you can calculate some of the public neoclouds anywhere, $8 million, $9 million, $10 million, $11 million. So a big premium for you. How do you -- like where does that come from? Is that durable? How does that flip back to your comments on attractiveness from a pricing perspective?

Yes. I guess the way to think about it is we were at $22 million per megawatt in Q4 of last year. let's just literally take our ARR of Q4 and divide it by 43-ish megawatts that are active at the end of the year, you get 2. And you're like, okay, well, that's great, but that's a lot of core cloud and you're just building the AI business. So then you got to say, okay, well, what's the incremental. For every incremental megawatt that you add, what do you think you'll get? And how does that compare to the competitors?

And based on the guidance that we provided, if you fast forward to the end of '27 and make some assumptions around -- we gave you 30% growth for the year and you back into the AI growth rates and everything, you'll likely conclude, the overall answer will be around 20 is what we said. And you can back into the -- that we're implying there's around $13 million per megawatt of incremental capacity that you would consider AI customer revenue. And that compares to the $9 million to $12 million that you said for the neoclouds. And you say, well, why are you getting more revenue? Are you just charging higher prices for the same thing, and it's not that at all. If it's just Bare Metal, there's a lot of price transparency in the market and people generally know what those margins are, and they're not particularly strong. But if you layer on inference services, which is like even just layering on GPU droplets on top of the wrapper around the Bare Metal to abstract that kind of some of the administrative capabilities, offering serverless inferencing, offering some of the other higher layer AI services, but then also pulling through core cloud.

So you're getting database, you're getting storage, you're getting bandwidth, you're getting compute, CPU compute, all of that is incredibly higher margin. Like the margin on the core cloud is, I think, 70s, 80s percent where the margins on the Bare Metal, as everyone knows in the industry is like 25-ish plus or minus. It depends on how you're really thinking about it, but it's not spectacular. So we're able to get more of the wallet of the customers using that GPU infrastructure. It's higher layer, which means it's stickier, right? You start getting data and database that makes that workload stickier. So one can move a Bare Metal training workload. It's harder to move an inference application, and we get a higher margin for it. So we think that, that number only goes up over time and is the embodiment of the differentiation that Paddy just articulated.

And Matt, just to add to that. So we talked about $120 million of AI customer revenue. That $120 million, we had a slide last week. The Bare Metal part of that is 30% and shrinking, right? So 70% of that AI revenue is coming from higher order services, which, by definition, are much higher margin. So that's how we are able to get more, and that's only going to go up from here.

To clarify, shrinking in mix. Growing but at a...

Growing more and more.

I'd argue I'd love it to shrink and flip it. We've already had customers like [indiscernible] that came to us, wanted Bare Metal initially because that's what they were getting from everyone else. And then after experiencing our network and our capabilities, they migrated up to higher layer services. So I'd like that Bare Metal to -- I'd love for everyone who comes in to want to take advantage of our higher capabilities. Clearly, we'll take more if we need to, to win customers initially, but we'd be working really aggressively to migrate them up.

Great. Let's talk about agents and OpenClaw. Basically, how are you positioned? What are you seeing in the market? And how are you positioned around that opportunity?

Yes, we are positioned extraordinarily well. So last week, I talked in the earnings call about how just in a handful of days, we had more than 30,000 OpenClaw one-click droplets running on our platform. That was with 0 marketing dollar spent, right? We just overnight, became a natural destination for deploying OpenClaw agents for one primary reason, right? So the reason why I'm super excited is everything we've been talking about from an agent perspective, we were able to see and then some. What I mean by that is agents need a lot more than just an AI model to run, right? Of course, we have AI models. You can bring your own Anthropic key or we have a dozen or so open source models. So that's all fine. But agents need memory, Agents need storage. Agents need a way to orchestrate and need sophisticated API capabilities, and they need CPU compute to perform actions and things like that.

So they essentially need a full stack cloud. They need serverless inferencing. They need everything that we offer. And that's why we saw this massive explosion of open claw agents, and that has really not stopped over the last several days. It's going strong. And for me, more than anything else, it just establishes a blueprint for agentic applications of the future, right? OpenClaw was more of a personal productivity type of agent. But when you extrapolate that to agents that are going to deliver value in the enterprise, we're just getting started. And the beauty of our platform is it is ready-made for agents to deploy other agents. It is ready-made for agents to consume our APIs.

So for example, we shipped remote MCP server last quarter. So with remote MCP servers, these agents don't even have to talk to a human and log into the cloud console to create new artifacts and stuff like that. They can just spin up new instances. They can spin up new capabilities on the DigitalOcean platform without talking to a human using the remote MCP artifact. So we are set up beautifully for this, and that's why we have become a natural platform because we have all the underlying pillars that agents love to leverage. So I'm super excited. for a simple reason that this shows everyone what a blueprint for an agentic cloud looks like.

Perfect. I want to come back to the data center strategy. You all colocation and then also as you think about bringing on this capacity throughout this year, what kind of risk or how do you think about the risk around delivery timing?

We communicated at earnings that our first of the 3 data centers, which is the smallest, it's 6 megawatts, is going to start ramping revenue in second quarter. The other 2, which is 1 is 10,1 is 15, come on in second half. We've been, as you've hopefully come to learn, appropriately conservative in terms of our planning around when we actually get those turned up and when we start to generate revenue and the pace at which that revenue ramps. So we're very confident in the guidance that we provided. And the 21% for this year, exiting the year at 25% plus, that's a good -- we feel very good about that.

And we're working, again, because we work with kind of existing colo providers, these aren't new investors that are building from dirt and they have never built a building before and are kind of going through the first time. In most cases, these are data halls and existing facilities and with very experienced operators, and we feel very confident about their ability to execute and our ability to partner with them.

Great. I want to come back to the go-to-market. And as your focus shifts to AI-native enterprises, how do you approach that from sales-led growth, product-led growth? Where are you on building out sales team where are the investments needed?

Yes. So as you all probably know, we are very likely the company -- at our scale, we are probably the most sophisticated product-led growth company in the industry. And on top of it, we layered in a little bit of sales-led growth, primarily looking at our digital native enterprise customers. You don't have 0% churn in our $1 million cohort by accident, right? So it was very deliberate from a go-to-market point of view. We put our arms around our big customers. We are working very hard to expand them, making $100,000 customers $5000, $500,000 customers $1 million and $1 million customers, $5 million customers.

So that farming motion or account management motion is working really well. On top of it, from an AI perspective, our product-led growth continues to be a big top-of-the-funnel machine for us. On top of it, we are also adding -- we are becoming very, very active in the venture community, in the start-up community, ensuring that we have a good, steady top of the funnel of well-funded start-ups that are in the precipice of changing their respective domains. We are a a16Z alumni. We are a TechStars alumini. I'm actually keynoting at the TechStars Conference on Monday. So we are very active with these communities to ensure that we are able to pick and participate in the top companies of their respective portfolios and bring them to our ecosystem.

So those are all the things that we are doing. Our AI sales team is still fairly small. It is mostly just inbound. We have no problem generating demand given our product-led growth machine as well as the technical evangelism that we continue to do. It's all hands on deck. We don't need an army of salespeople to bring in this revenue. So we have a small but mighty AI sales team. And we will continue to expand that primarily in one direction. We are doubling and tripling down on forward deployment engineering, which is going to be super important for the next couple of 3 years in terms of working -- having our engineering teams working hand-in-hand with customers because there is a lot of magic that happens when engineering teams get together. So FTE teams is a big focus for us, and that's how we are evolving our go-to-market.

All right. Great. I want to round out the discussion just hitting on capital allocation. We talked through CapEx, finance leases, but what about buybacks and what about M&A?

Yes. So buybacks have always been an important part of our long-term capital allocation strategy. I'd say, given the priorities right now are 100% focused on organic growth and maintaining a healthy and flexible balance sheet, we're unlikely to do material buybacks in the near future. We still have an authorization of -- I think it's $100 million over 2 years, but I would expect that we'd be using our cash for growth, not for buybacks in the near future. M&A is always something that we're focused on. I'd say we're probably more focused on product -- things that advance the product road map or acqui-hires versus any kind of a scaled M&A. So we don't anticipate it being a material use of capital at this point. We think we could do it like little tuck-ins here and there.

Perfect. Paddy and Matt, thank you so much for the conversation. Really appreciate it.

Thank you, Josh.

Thank you.

Good morning, and thank you for standing by. My name is John, and I will be your conference operator today. At this time, I would like to welcome everyone to the DigitalOcean Fourth Quarter Earnings Conference Call. [Operator Instructions]. I would now like to turn the conference over to Melanie Strate, Head of Investor Relations. Please go ahead.

Thank you, and good morning. Thank you all for joining us today to review DigitalOcean's Fourth Quarter and Full Year 2025 financial results and an investor update.

Joining me on the call today are Paddy Srinivasan, our Chief Executive Officer; and Matt Steinfort, our Chief Financial Officer. Before we begin, let me remind you that certain statements made on the call today may be considered forward-looking statements, which reflect management's best judgment based on currently available information. Our actual results may differ materially from those projected in these forward-looking statements, including our financial outlook. I direct your attention to the risk factors contained in our filings with the SEC as well as those referenced in today's press release that is posted on our website.

DigitalOcean expressly disclaims any obligation or undertaking to release publicly any updates or revisions to any forward-looking statements made today. Additionally, non-GAAP financial measures will be discussed on this conference call and reconciliations to the most directly comparable GAAP financial measures can be found in today's earnings press release as well as in our investor presentation that outlines the discussion on today's call. A webcast of today's call is also available in the IR section of our website. And with that, I will turn the call over to Paddy.

Thank you, Melanie. Good morning, everyone, and thank you for joining us. We had a fantastic quarter and a very strong finish to the year, and I'm excited to share the details with all of you. We ended the year with 18% revenue growth in Q4, reaching $901 million for the full year. We delivered $51 million in incremental organic ARR, the highest in the company's history.

Our 1 million customers reached $133 million in ARR, growing at 123% year-over-year. We maintained financial discipline and strong profitability with 42% adjusted EBITDA margins and 19% adjusted free cash flow margins for the year. There is a lot to be excited about. And given this momentum that we are seeing and the progress we are making against our long-term strategy, we wanted to provide a more comprehensive update today rather than wait for a separate Investor Day. Our prepared remarks will be slightly longer than usual. We'll advance slides from our earnings presentation on the webcast as we go, and we'll leave plenty of time for questions.

AI is reshaping entire industries, and we are built for this shift. Software is being disrupted, not by incremental AI features, but by a structural shift to agentic systems operating at scale. Cloud and AI native disruptors are moving beyond AI [indiscernible] at a breakneck speed. We are deploying agents that reason, act retain memory and run continuously.

In this structural shift, we see a secular hyperscale size opportunity by serving AI and cloud native companies driving this disruption. When markets are disrupted like this, there is typically a short window to take advantage of the opportunity, and let me tell you how we are seizing it. First, our top customers are now our growth engine. We have turned what was once viewed as a weakness into a competitive strength.

Our top digital native customers or [ D&E ] which include cloud and AI native companies are now our fastest-growing cohort and in fact, growing significantly faster than the market on [ DL ]. In a nutshell, Scaling our top customers was [ 1 second ] train. Today, it's our growth engine.

Second, we are on the right side of software disruption driven by AI. Modern cloud and AI native companies are going after large markets with disruptive AI-centric software innovation. They are increasingly choosing DigitalOcean at their natural platform to build and scale their [ IdentiKI ] software. And when these companies disrupt and scale at unprecedented rates on our platform, we win.

Third, we put the cloud in Neocloud. These AI natives need more than just GPU rentals or inference APIs. They need access to optimized AI models, both closed and open source, production-grade inferencing and a full stack cloud for their software, all working together at global scale. We deliver all of it in one integrated agentic inference cloud.

And finally, we are building a durable and profitable growth engine. We are investing responsibly while driving balanced growth. Without chasing the GPU training arm [indiscernible], we expect to deliver 21% revenue growth in 2026, reaching 25% plus growth by Q4 2026 and 30% growth in 2027. We are on a path to being a weighted rule of 50 company next year on the back of our existing committed data center capacity alone.

Put simply, we are accelerating growth the DigitalOcean way. In December, we crossed a major milestone, surpassing $1 billion revenue run rate. This is a remarkable achievement for a company that was founded through [ Techstars ] in 2012. This success is a testament to our passionate team and the vision of our original founders. I also extend my deepest gratitude to all our incredible customers who have supported us throughout this journey. But what matters more than this milestone is where we are going.

We exited 2025 at 18% year-over-year growth and are on a path to deliver 21% growth in 2026 with an exit growth rate of 25% plus in Q4 of 2026. We are picking up momentum, and we have outgrown the old narrative. Let me elaborate. Our top customers are now our growth engine. For our first decade, we built an iconic developer cloud. That foundation still matters, and we have over 4 million active developers on our platform that absolutely love us.

Over the last several quarters, we have deliberately shifted focus towards serving our top DNE and eliminating any reason for them to leave DigitalOcean as their scale and that focus is working. In Q4, we delivered a record organic incremental ARR of $51 million and $150 million on a trailing 12-month basis, both surpassing even our peak COVID era quarters. This record trailing 12-month incremental ARR was balanced across AI and cloud customers.

ARR from D&E reached $604 million in Q4, which is now 62% of total ARR, growing 30% year-over-year. And our D&E NDR reached 102%, continuing to outperform developer NDR. And like I've been reporting for a while now, our largest customers in the D&E cohort are accelerating the fastest. Our $100,000 customers are growing at 58%, our $500,000 customers are growing at 97%, and our $1 million customers who reached $133 million in ARR are growing at 123% year-over-year, all well ahead of market growth rates.

And NDR also increases meaningfully as these customers scale. Q4 was 102% for our $100,000 customers, 106% for our $500,000 customers and 115% for our 1 million customers. Churn for our $1 million customers was 0 in Q4 and has averaged 0% over the last 12 months which clearly shows that our top customers are now scaling with us and becoming our growth engine. You should also effectively debunk any misconception that our most successful customers will outgrow our platform.

Recapping this section, we are accelerating past the $1 billion revenue run rate milestone and our top customers are driving this acceleration. We are no longer defined just by entry-level developers experimenting on our platform. We are defined by high-growth cloud and AI native companies running production workloads scaling revenue and building their businesses on DigitalOcean. Said simply scaling our top customers was once a constraint. Today, it's our growth engine.

On to the next point. We are on the right side of software disruption. There is a structural shift happening in software and DigitalOcean is emerging as a preferred platform for cloud and AI native companies that are driving this disruption. The last generation of Software as a Service or SaaS monetized per user per seat, value, scale with headcount.

This next generation of AI-centric software monetizes per token for inference request Value scales with intelligence delivered as AI model capabilities accelerate entire categories of horizontal and vertical software are being reinvented. Incumbents are reacting to transformational change by layering AI into their workflows, seeking to enhance their existing software.

But AI native companies are starting from first principles. For them, AI isn't a feature. It is the very engine that defines their product. Every time they deliver value, [indiscernible], tokens are consumed and intelligence is produced. DigitalOcean is uniquely positioned to serve these disruptors, and that is evident in the traction we are getting from leading AI native companies. We have signed and expanded production workloads with scale, cloud and AI native companies like character.ai, workato and Hippocratic AI, companies with product market fit, real revenue and rapidly scaling demand.

Our work with character.ai demonstrates this clearly. We delivered 100% throughput increase and roughly 50% lower cost per token. For character.ai on our production inference cloud powered by AMD Instinct GPUs at production scale. This is not a [ lab ] benchmark. This is on live traffic across tens of millions of customers. This demonstrates our ability to support production scale inferencing for leading AI companies with our differentiated performance cost efficiency and integrated AI and cloud platform built for inference first production workloads.

Another AI native with a proven product market fit is Hippocratic AI who builds health care-focused conversational AI, designed to support clinical workflows and patient engagement. Hippocratic AI selected DO's agentic inference Cloud to power HIPAA-compliant clinical AI workloads. This validates not just our performance but our enterprise-grade security and compliance.

For Hippocratic AI, we optimize their multimodal deployment on NVIDIA hardware, reinforcing the importance of vertical innovation from GPUs to networking, [ cortile ] optimization, cloud integration and inference software. These AI native also scale very differently. While traditional cloud customers may take years to reach $1 million in ARR, AI native can cross that threshold in months or even weeks. When inference is your product demand compounds quickly.

DigitalOcean is purpose-built for these disruptors. As software becomes more intelligent and AI-centric, we are building the vertically integrated inferencing cloud designed to power the next generation of AI natives, putting us squarely on the right side of this AI-driven disruption and our Agentic Inference Cloud is capitalizing these disruptors.

Next, let me explain how we are enabling this. We do this by putting the cloud in Neocloud. Over the last couple of years, the new category of Neocloud has emerged that is largely optimized for one thing, large-scale AI model training, dense GPU farms, high-performance networking, frontier AI model training workloads. This is an important layer of the AI stack, but serving inferencing is different.

As AI diffuses into every software company, workloads shift from training a handful of frontier models to running millions of real-world applications. and real-world AI-centric software needs more than GPU farms. They need compute, storage, databases, networking, observability, security, all working seamlessly together with predictable and transparent unit economics.

Over the past 4 quarters, we have evolved our Agentic Inference Cloud to meet that reality. We have combined specialized inference infrastructure with our full stack cloud platform, purpose-built for production AI while staying true to what defines DigitalOcean, simplicity, open standards, enterprise-grade performance and SLAs and predictable and transparent unit economics.

A good recent example of this in action is [ Open law], which recently took the world by storm by demonstrating the power of agentic software, giving us a glimpse into what AI-centric software future will look like. [ Open Cloud ] is an open source AI agent framework that allows developers to run real-world task-driven agents. When customers deploy [ open cloud ] on big solution, they need more than just GPUs, because AI agents are stateful. They reason, they take action, they retain memory. They interact with third-party APIs. All this requires more than just a GPU form. It takes a full cloud and AI stack working together side by side.

Customers increasingly understand this as inference is the heartbeat of modern AI native. It is their primary operating cost, their performance level and their competitive moat. Their production traction scales directly with model quality, inference performance and unit economics. As they grow, they don't build their products around a single close source model, but rather orchestrate multiple models in real time, often leveraging open source and a mixture of expert approaches to optimize both accuracy and unit economics.

Our platform delivers flexibility at every layer, from serverless inference APIs to dedicated clusters and GPU droplets, allowing customers to precisely match performance and cost to their workload requirements. We pair that with performance optimized open source models, delivering high accuracy, strong throughput, low latency and compelling unit economics. And this isn't a stand-alone inference platform. It is deeply integrated with our full stack cloud that we have hardened over the last dozen years so that customers can build, deploy and scale their entire AI application in one integrated environment with enterprise SLAs.

Our agent development platform takes them from experimentation to production with real-world AI agents. Underpinning all of this is a deep lineup of GPUs from NVIDIA and AMD, supported by rapidly expanding global data center footprint, built and operated with years of operational expertise supporting mission-critical workloads. This integrated platform and flexibility of choice is precisely what makes DigitalOcean a natural platform for agentic software.

Let me explain this again using [ open cloud ] as an example. Customers can build and deploy [ open cloud ] agents on distillation in 2 distinct ways, depending on their need for control, scale and operational complexity. The first path optimizes on simplicity and speed. Customers can launch a preconfigured one-click GPU droplet and have an [ open cloud ] agent running in minutes. This model gives full control over the environment. Ideal for experimentation, customization, performance tuning and for teams that want direct access to the infrastructure layer.

The second path optimizes for global scale. Customers can deploy [ open cloud ] on DO's managed serverless platform where DigitalOcean handles provisioning, scaling, security, container orchestration and operational management. This approach is ideal for teams that are scaling a global application. Both approaches run on the same integrated cloud with access to managed databases for agentic memory object storage for artifact, virtual private cloud networking, observability and GPU backed inference.

That's what vertical integration looks like in the inference economy, not just providing bare metal GPUs or even just generating inference tokens, but providing a secure, scalable and manageable foundation for intelligent stateful systems. Within days of launching [ Open Cloud ], nearly 30,000 native DigitalOcean, 1-click [ open cloud \ droplets were created, and that was just the starting point. thousands of other open cloud deployments were activated by customers, signaling the emergence of a new ecosystem almost overnight.

The success of [ open cloud ] is an early view of how the AI market will continue to evolve and can serve as a blueprint for AI native businesses on how a new generation of software will be built around autonomous agents that orchestrate complex multistep workflows across systems, continuously reason with data and context and execute tax end-to-end with minimal human involvement.

As these AI native companies move from proof of concept to production agents, the richness of the underlying platform, the security posture, manageability, scalability and predictable unit economics become mission-critical. And that is exactly where distillation is fast emerging as the natural platform for building and scaling AI agentic software.

The competitive landscape is crowded with companies speaking to their ability to address the inference market, but our differentiation from these competitors is very clear. Neoclouds rent out GPUs. Inference [ rapper ] providers stop at inference APIs and model libraries. We continue to effectively compete with hyperscalers who bring scale, but also come with complexity and cost structures that are aimed at traditional large enterprise companies.

While each of these competitors address a component of the inference value chain, real-world identic software requires a tightly integrated environment where inference, orchestration, persistence, networking and security are designed to work together with simplicity, global scale, enterprise SLAs and predictable unit economics. That is where DigitalOcean wins.

This differentiation is clear to our customers, but it's also very clear in our financial profile. As a full stack cloud provider, that has operated mission-critical workloads for cloud and AI native for over a decade, we look very different from a financial perspective than other players chasing the AI training market or components of the inference market. Where Neocloud has very high revenue concentration with just a few very large customers making up the vast majority of their revenue, [ dissolutions ] top 25 customers represent only 10% of our revenue.

While GPU rental providers own bare metal revenue and margins on their infrastructure, DigitalOcean drives higher revenue and margin from our full stack inference and cloud solutions. And when a growing number of Neoclouds are investing massive amounts of capital and burning near-term profits and cash for future returns, this solution is already profitable and generating cash.

Our traction with cloud and AI native is no accident. It is the result of relentless focused investment and disciplined execution. We recently strengthened our executive team by adding Vinay Kumar as our Chief Product and Technology Officer. As a founding member of Oracle Cloud Infrastructure, or OCI, Vinay brings deep hyperscale expertise and leads our product, platform, infrastructure and security teams, having built a hyperscaler from the ground up at OCI, he looks forward to scaling up another one at DigitalOcean, one that is purpose-built to meet the complex needs of cloud and AI native workloads globally.

In the meantime, our R&D team has been very busy continuing to ship products and features that are helping our customers scale on our platform. On [ GoreCloud], we launched remote MCP support embedding AI directly into the control plane, enabling secure 0 setup infrastructure management. On our AI platform, we introduced the age and development kit, an enhanced agent evaluation tools to help customers move from experimentation to production with measurable performance and reliability.

With GPU observability, managed NFS and multi-node GPU support, we significantly expanded our ability to run large-scale mission-critical inference in production. This is what vertical integration looks like, infrastructure, inference, observability, agent tooling, all built to seamlessly work and scale together. And we're just getting started. We'll share the next wave of innovation on our Agentic Inference Cloud at our next deploy conference in San Francisco on April 28, as we continue building the platform, purpose built for the inference economy.

Our differentiation is durable and will continue to grow as the market shifts from training to inference. To give investors clearer visibility into this momentum, we are introducing a new metric, ,AI customer revenue. AI customer revenue includes all revenue from customers leveraging our AI products, including both inference and core cloud services. Because AI natives don't just buy GPUs, they build, operate and scale applications which need a full stack inference cloud.

In fact, 70% of our AI customer ARR in Q4 2025 was already coming from inference services or general-purpose cloud products rather than from bare [ metal ] GPU rentals. And these customers are growing rapidly with Q4 AI customer ARR reaching $120 million, growing 150% year-over-year, now making up 12% of total ARR.

In summary, we don't just rent GPUs. We run production AI. We are not a GPU landlord. We are an AI cloud platform. We deliver hyperscaler grade infrastructure and reliability purpose-built infant services co-located and integrated with a full stack general-purpose cloud designed for the next generation of AI native. Or put simply, DigitalOcean puts the cloud in Neocloud.

Now on to my final takeaway. We are building a durable and profitable growth engine. At our Investor Day last April, we laid out a plan to return the business to 18% to 20% growth by 2027. On our last earnings call, we pulled that growth projection forward by a full year guiding that we would reach that 18% to 20% growth range in 2026. And just 9 months after setting that original plan, we've already reached the bottom end of the target range at 18% growth in Q4 of 2025, achieving it 2 full years ahead of our original target.

And the momentum we are seeing gives us even greater confidence. We now expect to deliver 21% revenue growth for the full year 2026 with an exit growth rate of 25% plus by Q4 and reaching 30% growth in 2027. As we ramp into our committed 31 megawatts incremental capacity this year, there will be measured near-term pressure on gross margin and adjusted EBITDA, but we remain confident in our 18% to 20% unlevered adjusted free cash flow margin guide for the year.

The near-term pressure is just a physics problem, given the start-up cost timing and revenue ramp characteristics of quickly adding new capacity. It is the natural result of pursuing high-return growth opportunities, but we remain disciplined operators. Demand continues to far outstrip supply. And we will take advantage of opportunities to further accelerate growth when they present themselves. We will do so responsibly and we'll continue to pursue investments with attractive returns match investments with revenue timing, maintain a strong balance sheet and allocate capital with trigger even as we accelerate. Growth and discipline are not trade-offs for us. They are both operating principles.

With that, I will turn it over to Matt to walk through the quarter and the year in more detail and to provide additional color on our updated outlook. Matt, over to you.

Thanks, Paddy. Good morning, everyone, and thanks for joining us today. As Paddy just shared, we're a very different company today than we were just a few years ago. It's an exciting time at DigitalOcean. We are a rapidly growing and profitable company that is incredibly well positioned to take advantage of the hyperscale sized inference market opportunity. This excitement is clearly evident in both our recent financial performance, and in our higher near-term and long-term outlook.

Revenue growth has reaccelerated. We've reversed declines from our top customers, turning them into a key driver of our growth. We have scaled our AI customer ARR to $120 million, growing 150% year-over-year. And we've done this profitably, growing adjusted EBITDA and adjusted free cash flow on both an absolute and a margin basis. While we are pleased with our progress over the past several years, it is our recent momentum that gives us the confidence to further increase our near-term and long-term outlook.

Fourth quarter revenue was $242 million, up 18% year-over-year and we closed '25 with full year revenue of $901 million. We delivered sustained acceleration through the back half of 2025, driving a 500 basis point increase in Q4 growth from the same period just a year ago. We delivered the accelerated revenue growth with strong margins and growing profits even as we increased our investments.

Fourth quarter gross profit was $142 million, up 13% year-over-year, with a gross margin of 59%. For the full year, gross profit was $540 million, up 16% year-over-year, with a gross margin of 60%. Adjusted EBITDA in the fourth quarter was $99 million, an adjusted EBITDA margin of 41%. Full year adjusted EBITDA was $375 million, a 42% adjusted EBITDA margin. Trailing 12-month adjusted free cash flow was $168 million in Q4 or 19% of revenue.

We maintained our attractive free cash flow margins in '25, in part by expanding our financial toolkit to include equipment financing. This better aligns infrastructure investment timing with the revenue that it supports. We will continue to utilize a combination of upfront asset purchases and equipment leasing as we invest to fuel our growth. We continue to be disciplined financial stewards for our investors. We prudently use stock-based compensation to attract and retain our critical talent while repurchasing shares to mitigate to [indiscernible].

[indiscernible] declined to 9% of revenue in 2025, down from 12% in the prior year. To put that number in context, we have a 33% margin if you subtract [ SBC ] from adjusted EBITDA. At 33% margin, we are just above the 80th percentile of a broad software comp set on an adjusted EBITDA less SBC basis. And we are well above the 13% median of that group.

Non-GAAP weighted average shares outstanding increased slightly from 103 million to 105 million over the same period. To reduce dilution, we repurchased 2.4 million shares in 2025 for $82 million at an average price of approximately $35. Note that we ended 2025 with our full $100 million buyback authorization in place and that authorization continues through July 31, 2027.

While we continue to view share repurchases as an important long-term tool, our near-term capital allocation priorities are squarely focused on organic growth and balance sheet flexibility. GAAP diluted net income per share in the quarter was $0.24 and $2.52 for the full year. 183% year-over-year increase. Non-GAAP diluted net income per share in the quarter was $0.44. For the full year, non-GAAP diluted net income per share was $2.12, a 10% year-over-year increase.

As a quick reminder, recall that our 2025 net income per share metrics were impacted by the actions we took in '25 to strengthen our balance sheet. In 2025, we proactively addressed the upcoming maturity of our 2026 convertible notes. We did this through a series of successful financing transactions that have given us significant balance sheet flexibility. These transactions included the establishment of an $800 million bank facility, the issuance of $625 million of 2030 convertible notes and the repurchase of the majority of our then outstanding [ 26 ] convertible [indiscernible].

Excluding the effects of these financing transactions, non-GAAP diluted net income per share would have been $2.29 for the year and $0.53 for the quarter. With our 2026 notes largely addressed, we ended the year with a strong balance sheet. We have sufficient liquidity and projected cash generation to address the remaining $312 million balance of our outstanding [ '26 ] convertible notes.

Having drawn down the remaining $120 million on our Term Loan A in February, we will repurchase or redeem the remaining [ 26 ] notes for cash before or at the maturity in December of '26. Beyond this, we have no other material maturity until 2030, and we entered 2025 with approximately 3.2x net leverage.

Before I get into guidance, I want to highlight an action we are taking to further concentrate our investments on our key growth levers. We are sunsetting a small legacy dedicated Bare Metal CPU offering. We expect approximately $13 million of ARR to roll off by the end of Q1 2026. As this revenue is noncore, we have excluded this legacy product revenue from our customer-specific year-over-year growth metrics.

0

Shifting back to guidance. We entered 2026 with tremendous momentum and confidence. Paddy spoke of the material demand we're seeing for our Agentic Inference Cloud. We also continue to improve visibility on our net [ term ] revenue growth as we increased RPO in Q4 to $134 million, up 121% sequentially, up close to 500% year-over-year. With this growing demand and visibility, we are again increasing our near-term growth outlook.

For the first quarter of 2026, we expect revenue in the range of $249 million to $250 million, which is approximately 18% to 19% year-over-year growth. We expect first quarter adjusted EBITDA margins in the range of 36% to 37%. We expect non-GAAP diluted net income per share of $0.22 to $0.27 based on approximately 111 million to 112 million weighted average fully diluted shares outstanding.

For the full year 2026, we expect revenue growth between 19% and 23%. This is 21% at the mid -- beyond the 18% to 20% growth outlook that we shared just last quarter. And it is important to highlight that this would be 21% to 24% projected growth if we exclude the impact of our discontinued legacy Bare Metal CPU offering.

We will deliver this accelerated growth while maintaining attractive margins. We project full year 36% to 38% adjusted EBITDA margin and 18% to 20% unlevered adjusted free cash flow margins, which is $207 million at the midpoint. We expect non-GAAP diluted net income per share of $0.75 to $1 on 111 million to 112 million weighted average fully diluted shares outstanding. This growth outlook is based on the incremental data center and GPU capacity investments that we have already committed that will come online over the course of 2026.

As we look at the quarterly progression within 2026, it is important to understand the timing of this incremental capacity and how that timing impacts our financials. We are bringing 31 megawatts of new data center capacity online and 3 new facilities in 2026. The smallest of our 3 new facilities will start ramping revenue in the second quarter. The remaining [indiscernible] start ramping revenue in the second half of 2026.

Aligned with this capacity ramp, we expect second quarter revenue growth to remain around 18% to 19%, with revenue growth then ramping in Q3 before exiting the year at 25% plus in Q4. While there are always supply chain and implementation timing risk to manage, we believe our implementation time line is realistic.

Increased data center lease expense and equipment depreciation expense will both hit our financials several months before we generated our first revenue in these facilities. Given this lag between expenses and revenue, cost of goods sold from higher GPU-related depreciation and operating expenses from new data center operating leases will increase in the early part of the year as we ramp into the new capacity. These increased costs will cause the expected upfront drops in gross margin and net income that we have seen when we turned our previous data centers.

The initial impact will just be larger as we are turning up more capacity at one time than we've done in the past. Near-term adjusted EBITDA margins will also be impacted somewhat from these dynamics although the impact is less as adjusted EBITDA is only impacted by the higher data center operating. Net leverage is projected to be above 4x and in short term as we add finance lease obligations to fund our GPU and CPU investments, this increases net debt several months ahead of revenue and adjusted EBITDA ramp.

We anticipate returning below 4x net leverage over the medium to long term as we increase utilization in these data centers and ramp revenue and adjusted EBITDA. We will achieve these growth targets by focusing on our 2 primary growth levers, scaling our top D&E customers and expanding our base of AI native customers. We will focus our investments on meeting the needs of our top D&E customers so that they can continue to scale on DigitalOcean as they grow their own businesses.

We will continue to invest both in our differentiated Agentic Inference Cloud and in the data center and GPU capacity required to support AI native. While we are excited by our growth potential in 2026, we are just getting started as we reach full utilization on our existing committed capacity, we expect to reach 30% revenue growth in 2027. We will drive this growth while delivering projected 20% plus unlevered adjusted free cash flow margins, which would make us a rule of 50-plus company in 2027. We will achieve this while making smart investments, earning attractive margins and maintaining a healthy balance sheet.

We have both the tools and the discipline in place to continue to take advantage of opportunities as they arise. We will continue to share details on our leading indicators and our progress as we execute. We are increasingly confident in our ability to build a durable and profitable growth engine. With that, I'd like to turn it back over to Paddy to close this out before we get to Q&A.

Thank you, Matt. Before we move to Q&A, let me leave you with a few thoughts. We crossed $1 billion revenue run rate in December, but that milestone is not the headline. The headline is where we are heading. We are no longer a niche developer cloud, with a platform that high-growth cloud and AI natives are increasingly choosing to run production AI workloads at scale. We are projecting to exit 2026 at 25% plus revenue growth with a clear path to 30% growth in 2027 with the existing committed data center capacity alone.

Our top customers are accelerating and are growing significantly faster than the market on DO. We have outgrown the old DigitalOcean narrative. Scaling our top customers was once a constraint. Today, it's our growth engine. Our $1 million customers are at $133 million ARR, growing at 123% year-over-year. The world of software is shifting from seats to tokens from experimentation to production for model training to inferencing at scale. And in that shift, the winners in inference will be more than just GPU landlords.

They will be vertically integrated AI cloud platforms that deliver performance, great unit economics and simplicity that embraces open source, exactly what we have and what we continue to build. Our AI customer ARR reached $120 million in Q4, growing 150% year-over-year with 70% of that coming from inference and core cloud products, not from Bare Metal. And we're doing it without chasing the GPU training arms race.

Without sacrificing discipline, without compromising profitability, we are building something durable. AI is reshaping entire industries, and we are built for this shift. I'm incredibly excited to be part of DigitalOcean at this critical inflection point where a new era of software is being ushered in. I take incredible pride in building a platform that AI pioneers are increasingly leveraging to disrupt software. I thank all of you for your partnership and support, and I hope you will join us in San Francisco on April 28 to learn about our platform, our innovation and our customers. With that, let's open it up for your questions.

[Operator Instructions] Our first question comes from the line of Raimo Lenschow with Barclays.

Congrats from me. That's amazing how a company is transforming right in front of my eyes. Paddy, can you talk a little bit about the customers that you're seeing? Like the talk in the market, a lot of that is just opening on trades, maybe Google and they are basically doing everything and nobody else really comes up. When you talk looking at your customers, looking at the pipeline of customers out there. How do you see that inference market evolving in terms of how broad that will be? Is it just unproper doing everything? Or what are you seeing out there in the field? And then I had one follow-up for that.

Yes, Raimo, thank you for the question. It's a very thoughtful way to get started. Of course, OpenAI, Gemini and Anthropic get all the headlines in the mainstream news coverage. But as we talk to AI native companies and even examples that I was using in my script, and you will hear a lot more about this at our deploy conference with very specific benchmarks and data.

But what we are hearing from these AI native companies is that while these close source models are really, really good, the open source alternatives are extraordinarily important to manage the unit economics as these companies came because the cost per token for the open source model is about 90% cheaper, right? So with a very comparable accuracy as these open source models mature.

So we have many AI native customers that are using as I mentioned, a variety of open source models at real time when they're doing inferencing, they want us to manage a multitude of open source models and even route their request intelligently to these open source models, and of course, use close source expensive model on a case-by-case basis, it could be for certain prompts, which are better served by these close source models and route everything else to these open source models so that they can have a balanced unit economics.

So it is by no means -- and if you look at data from open router, 30% of the traffic already today is served by open source. That is without a lot of optimization that is without companies like DigitalOcean really stepping up and taking full ownership and guardianship of these open source models. So we are doing a lot of work in this regard over the next couple of months, and you will see it in our [ deploy ] conference.

But this 30% is only going to grow as these real-world AI native workloads explore we are going to see a lot of open source adoption. Even in the open deployments that we are seeing, there is a very healthy adoption of open source model serving these open class agentic -- agent farms. So it is really interesting to see how this is evolving. And I want to say there is definitely a world beyond these closed source models. The open source ecosystem is thriving, and it is only going to grow in strength from here on.

Yes. Okay. Perfect. And Matt, one question that comes up a lot at the moment is on the weighted rule of 50 numbers. If you look at your [ waiting], and then there's a lot of questions about the free cash flow margins that you think about in 2027. Can you maybe kind of go a little bit deeper there because that comes up a lot here at the moment?

Yes. Thanks, Raimo. The weighted rule of 50 is pretty simple for us. We multiply revenue growth by 1.5% and add 0.5x the free cash flow margin. And that's effectively saying that you're counting revenue growth 3x as valuable as the point of free cash flow margin. But the important thing to note is while we talk about weighted rule of 50, if you look at the growth projections we provided were actually a regular weighted -- a regular rule of 50 as well with projected 30% revenue growth in '27 with 20% unlevered free cash flow margins.

So that is, I think, a very big testament to the growth opportunity that we have in front of us. But also the disciplined financial discipline that we've been employing with the ability to accelerate revenue growth while still maintaining very attractive EBITDA margins and very attractive free cash flow margins is kind of part of the model, and it's the benefit of us not chasing the GPU training kind of arms race.

It's -- we believe that we'll differentiate based on software and a differentiated platform, and we see a tremendous opportunity to drive really attractive margins as we expand and invest appropriately.

Our next question comes from the line of Kingsley Crane with Canaccord Genuity.

Congrats to the whole team on the results. I think you've done an excellent job with the investor update. I actually want to circle back to the inference cloud dynamic with open source models. We've been looking open router data as well. I mean some of these models come and go pretty quickly, have many [indiscernible] communicate to. How are you thinking about quickly providing support for those classes and models? Is there any operational tax to quickly provide support? And then just how to think about them driving growth, both from a revenue and profit standpoint. Could there be more of a [ Jevons ] paradox dynamic there with the lower cost models?

Yes. Thank you, Kingsley. That's a good question. So you asked 2 different questions. One, from an operational overhead in terms of days or support to these models, obviously, we've been extending day support for a majority of these open source models as they come out. And there are a couple of things there.

One is, obviously, there's a little bit of manual overhead in supporting these models. But a large portion of this test and readiness harness is automated. And it is only going to grow in automation, and you will see a lot more details around this at our deployed conference. And the second part of your question was really around the [ Jevons ] paradox of as these open source models proliferate, how should we think about the growth profile of not just our platform but also these companies, I think it is only going to aid in the deployment of AI native software in pretty much every segment of the market.

And I think we should also not think about AI native workloads at open source or close source. What we are seeing is the mixture of both for the same use case, for the same inference call even, some parts of the application stack -- based on the pumps, we do intelligence routing. Right now, it's fairly manual, but we are working on different types of algorithms to route it in a much more intelligent and smart fashion.

So you will see a universe going into the future where prompts are going to get routed to different models all working together at the same time to deliver high throughput, low latency, acceptable accuracy with great unit economics of token throughput. So this is coming. We are already seeing it from many of our AI native workloads. And that is how I see the market evolve as open source model continue to catch up with these closed source systems. The close source systems are really important to be on the bleeding edge of innovation, but a vast majority of these long-running agentic software like [ Open Cloud ] can very materially run on these open source systems.

Thanks, Paddy. That's really helpful. And then for Matt, obviously, $22 million for ARR per megawatt is a clear differentiator. I'm curious now that Atlanta is close to full utilization. Any insights you have on just what a full utilized megawatt can look like in terms of a revenue efficiency standpoint for AI.

Yes, that's a great question, Kingsley. If you look at the public data that's available to like a Neocloud, which is more of a bare metal model, they show like, what, $9 million to $12 million, I think, in ARR per megawatt. Clearly, we believe we can deliver more than that. And if you look at the guidance that we've given, what you'll see is that while it's [ 22 ] now, that's, again, with a small less than 10% or right around 10% of our ARR in AI.

So as we grow AI, it will come down. We'll add incremental ARR per megawatt greater than what you're seeing from the Neoclouds, but the drop from a bigger mix of AI by the end of '27 once we're fully ramped with the incremental [ 31 ], it will only drop by a couple of million. That will be around $20 million. And so if you think of us as not having, okay, we've got AI investments and we've got core cloud investments, but we have more of an overall AI cloud platform that has GPUs, it's got CPUs. It's got core compute and bandwidth and all the capabilities that you need, we still expect to deliver materially higher ARR per megawatt than what you're seeing in the Neocloud space.

So we feel really good about the returns that we're getting and the margin that we're able to drive. And this is only going to increase. I mean, you saw the chart in the deck about how many -- how much of the AI customer revenue is coming from non Bare Metal that's 70%. That's only going to increase. And that smaller lever of core cloud is only going to increase as customers become entrenched on our platform and they start putting in database and storage and some of the other higher-margin capabilities that are sticky. We're very excited about our ability to serve the kind of full addressable wallet of the AI native.

Our next question comes from the line of Josh Baer with Morgan Stanley.

Congrats on the strong results and impressive targets. Just wanted to clarify, the incremental 31 megawatts that all comes online by the end of '26 driving that 25% revenue growth exiting the year. But then as utilization increases, the capacity is enough to reach the full 30% growth in 2027 revenue?

That's absolutely right, Josh. You nailed it. as we said in the call that the smallest of the 3 facilities, which is 6 megawatts, is going to come -- start ramping revenue in the second quarter. But the other 2 start ramping in the second half. And just with the -- what we believe is appropriate assumptions around the timing and the ramp of that we'll hit 25% in Q4 as an exit growth rate, 25% plus. And then if all we did was kind of fill those -- continue to fill those up, we've hit 30% for the full year in 2027. And we feel very good about, again, the returns that we would generate there and the growth trajectory that we would be on at that point.

Okay. That's helpful. And I was just hoping you could sort of review some of what Vinay Kumar's top priorities are at this point, there's been so many positive changes from a product and innovation perspective over the last couple of years. What are his priorities? What changes should we expect going forward?

Yes. Thanks, Josh. So as I was mentioning in my prepared remarks, given his background at RF Cloud, he has really hit the ground running. His top 1 or 2 priorities are going to be building -- continue to build out the inference cloud. And you will see a lot of very detailed announcements on April 28 at our deploy conference on how the next generation of this inference cloud capabilities is going to look like. The team is super [ head down ] and busy working on it now.

We also will continue to raise the bar on our core cloud capabilities because our cloud native digital native enterprise companies are also scaling tremendously on our platform and they require continuous innovation from our side on advanced things like different types of databases and different scale aspects, scalability aspects of our database as a service and various parts of our core cloud infrastructure like high-performance storage, network file systems.

So one of the things that Vinay is working on is delivering innovation in our core infrastructure that is applicable to both AI native and cloud native. So there is a huge intersection that when you look at companies like the AI native that we are rapidly scaling up on our platform, they require very similar things from, say, high-performance storage as an example. Like I don't want to preannounce stuff that we are working on, which we will come out on April 28 with.

But a lot of those things are very similar to what our cloud-native companies can also benefit from. So there is a quite a robust lineup of capabilities that we are working on for both the inference cloud as well as some of the underlying infrastructure enhancements that will be applicable to digital native enterprise companies. So that's what he's focused on delivering. And as I mentioned, given his background, he's almost hit the ground running in terms of ramping up the innovation on the core inference cloud.

Our next question comes from the line of Wamsi Mohan with Bank of America.

Great to see this growth acceleration here. Firstly, maybe, Paddy, just visibility around the 30% growth. How should we think about that in terms of -- I mean, historically, obviously, dilution is a very different company today. But historically, you really not had like long-term contract, long-term visibility. You're talking about very meaningful acceleration as you go to 30% plus.

Maybe if you could dissect some of the underlying drivers of what you're looking at, which gives you the confidence? And maybe despite [indiscernible] that between Infrastructure as a Service and Platform as a Service, that would be maybe a different way to slice and give people a view over there. And I have a quick follow-up.

Thank you, Wamsi. The -- I think Matt broke down some of the physics of the acceleration, right? So we have new capacity that is ramping up throughout this year and going into next year as well. So that gives us a lot of visibility. So first of all, maybe I should take a step back and talk about the fact that our -- the demand that we are seeing now is very, very robust. And it far exceeds the supply that we currently have from an infrastructure point of view.

So we are being super responsible in ramping up our capacity. We are being super aggressive in the time lines. We are working very closely together with the data center providers and [indiscernible] to get this capacity online and the fastest possible speed of flight scenario as much as we can. So given the schedule that we are currently working on, we feel very confident that as we bring this capacity online, we have enough demand in the pipeline to be able to fill up this capacity with very responsible unit economics.

So that's what is giving us the confidence to provide the outlook of 25% plus exiting this year and 30% for next year. And also, our RPO has been going up steadily, and that is one leading indicator. But also, I should add the fact that inferencing is very different, right? I mean this is a -- these are real-world workloads. As opposed to training where a company can just raise venture capital money and just commit to a 2-year, 3-year contract to burn dollars to build a frontier model inferencing workloads are typically paid by end customers. So for us, that is super exciting because we are typically working with post product market fit companies that have real revenue, working with real consumers or business to business like hypocritic AI.

They're deploying in some of the world's largest health care providers. So we know that as their demand picks up, they're going to need more and more inference capabilities. So our confidence really stems from the visibility we are getting into our customers and the real-world influence demand. So I feel if you look at it from a customer perspective or you look at it from a capacity point of view, those are the one -- those are the data points that we used to triangulate our guidance for exiting this year and next year.

Okay. And then maybe one quick one for Matt. So can you just talk a little bit about the margin progression? I guess you mentioned some near-term margin compression given your capacity ramp should we expect that will persist through all of 2026, given the timing of the ramp? And then as you ramp into '27, we should be back to 2025 levels.

Thanks, Wamsi. Yes, there's certainly going to be some near-term pressure, as we said, on the gross margin, for example, but the metrics that we think are the best indicators of profitability for us are continue to be adjusted EBITDA margin and free cash flow margin, both on an unlevered basis and on a levered basis.

And if you look at the margin guidance that we provided for the full year '26 and the ranges for '27, you see exactly what you just described, which is we'll have a little bit more pressure this year as we ramp. But then as we grow into that, the utilization increases, that kind of catches back up and you should see an upward trajectory on the margins. The mix of AI services versus the core cloud, that's certainly -- that's a longer duration impact because as we add more AI capabilities and more AI revenue, margins are lower than the core cloud margin.

So you'll have a little bit of a mix impact in addition to the timing impact, but all of that is net out in the very, very strong adjusted EBITDA margins that we're projecting and the very strong adjusted free cash flow margins and unlevered adjusted free cash flow margins.

Our next question comes from the line of Gabriela Borges with Goldman Sachs.

Congratulations to the DigitalOcean team. Paddy, I have a little bit of a longer-term question for you. If I think about DigitalOcean core value proposition, on democratizing access to cloud. That has been true for many years now. My question for you is, what do you think is structurally different with the AI compute cycle that will allow digitation to essentially capture and hold on to a higher share of wallet in AI inference compute relative to the cloud cycle.

And the reason asking is because there are 32 companies that show up in the semi analysis across the [ MAX ] benchmarking airport. We know that [indiscernible] is early. We know the inference cycle is early. How do you think about DigitalOcean's ability to durably capture higher share relative to the 31 of the competitors in the long term?

Thank you, Gabriela. And I'm sure if semi analysis was around in 2011 or 2012 when cloud was taking off, they would have been 32 BPS providers as well. And we went from that to a $1 billion run rate in 12, 13 years. And [ Vinay ], if I take a step back and think about how durable our mission is in the world of AI, I think I hit on a few different things, right? I fundamentally believe that inference workloads are also workloads or real-world applications as well as the application scales, you need a variety of different things all working together, right?

AI native, [indiscernible] do want to just use one provider for token generation, go to another provider for database, go to a third provider for their application experience and go to a fourth provider for some of the other core storage and other artifacts. They want an integrated cloud that is co-located and all of these perimeters to work hand-in-hand together so that they can focus on building their business are not mess around with infrastructure.

The other part that I feel very confident is something that we are going to be talking and dealing with a lot in our deployed conference on April 28, which is this emergence of a mixture of AI models that is required to run efficient unit economics in inferencing mode. So the difference in the unit economics between closed source models and open source models is 90 -- so open source models are 90% more cost effective than it compares to close source models. And it already has 30% market share with just a handful of open source models on the market.

So I feel this is only going to go from strength to strength. And that has been a big differentiator for DigitalOcean throughout the years as well. So we talk about 32 companies showing up in some of these market landscapes. But when [ Open Cloud ] became viral a couple of weeks ago or a month ago, we were one of the natural places where developers started deploying it.

As I mentioned, we have more than 30,000 of these agents running, and we barely did anything from a marketing point of view. In fact, we did no marketing. All we did is scramble our jets to make sure that developers have first-class experience deploying these agents on our platform and we were such a natural choice for running these long-running agentic software because they need a lot more than just access to GPUs or just access to inference tokens.

So I feel very good that we -- our product strategy is working, and we are able to serve the needs of inference workloads running in production. So we're already starting to see the proof points for where different parts of our inference cloud are getting lit up. And the slide that I walked through in terms of our AI customer revenue, 70% of our revenue already is from non Bare Metal. And that should give us a lot of confidence that our platform services, higher-margin services are resonating with our customers. They're increasingly coming on -- coming to us as they recognize that bare metal is not going to be sufficient for them.

Yes. Really good color. I'll stay on this one -- 70% non Bare Metal data point. And I'll ask the question to Matt. Payback period on GPUs. The last time we talked about this, I think you've told us it was around 3 years. But that before your [indiscernible] head focused on maximizing or improving the ARR per megawatt of capacity. So my question is, how is payback periods in GPU engine?

Well, that's a great question, Gabriela. And one of the things that I want to make sure everybody understand is if you think about why did we lease gear, like why are we doing equipment leasing, it's to address exactly this challenge. If you said, okay, well, you're going to spend hundreds of millions of dollars on GPUs and you're going to have to wait 3, 4 years to pay them back. That's a model. That's not the model that we're pursuing.

Our model is we're leasing the gear, which means we're earning more ARR per megawatt and pretty associated GPU investment than what a Neocloud would earn. But we're also earning cash on that within months of actually deploying it. right? As soon as we deploy that and we start earning revenue and it ramps, we're paying on a monthly basis for that gear over 4 or 5 years, and we're earning more than 2x that in revenue.

So from a payback period, we still have the same kind of payback hurdles that we've had before. You'd like to see 3-year paybacks on most of your investments. You might be willing to extend that to win some early customers. But if you actually think about the mechanics, that there isn't -- that's a little bit of an intellectual exercise because we're paying -- we're already paying our gear back within a month or 2 because we're earning more cash than we've spent on that gear. And that's the reason you align your investment with revenue.

Our next question comes from the line of Param Singh with Oppenheimer.

First of all, Paddy, I wanted to get a sense of your [indiscernible] AI platform, obviously, that's driving a lot of growth. But where do you think some of the missing pieces are in terms of your technology given that the Neoclouds are starting to get a little bit more aggressive, do you think you have a sustainable competitive advantage? And how do you plan to sustain that?

Yes. Param, not only do I think we have an advantage now and our lead is increasing compared to other Neoclouds because they are coming from a training world, which is which is totally different, right? The needs all the way from the way GPUs are network and the cluster sizes, everything is so different. Inferencing is very different, as I explained, and if you look at the slide that shows the richness of our inference cloud stack, each one has taken us years and years to perfect.

And as we work very closely with cloud AI native companies, we are -- we are understanding and getting an appreciation for their real challenges, right? The example that I was talking to you about where customers need orchestration across different AI models at real time when they are trying to parse out a prompt and so that query or make real-time decisions.

So we are getting so much intelligence just working hand-in-hand with our customers. I feel like our lead is only going to increase from here on. And it's not to say that we won't have competition, but I feel very confident in our ability to out-invent these other companies in terms of our inference cloud. And the durability is for you to see are we have 0% churn in our $1 million per customer. So something is working, and that is our agent inference cloud.

And as my follow-up, do you feel you're constrained by the availability of power and physical location at this point? Or put conversely, given the opportunity to invest even heavier and grow faster, given the demand from the AI natives, what would you prefer at this time? Or would you rather have a slower pace of investment? Happy if you could give me some insight on really appreciated.

Yes. As Paddy said, we have more demand than we have supply. But we're also making, I think, very prudent and appropriate investment decisions. We don't want to go all in with like a single customer. We don't want to go all in on a single generation of GPU technology. We believe that building a diverse set of customers that are very heavy in the inferencing workloads and not chasing training, we'll build a durable model for us.

And so we'll continue to evaluate opportunities to continue to accelerate our growth and we'll make good appropriate financial decisions, and we're doing it in a very balanced way across a diverse set of customers. But we're very, very highly concentrated on what we're good at and where we're differentiated and where we can earn a good return, and that's what's driving our investment decisions.

Our next question comes from the line of Radi Sultan with UBS.

First one for Paddy, kind of on a similar line of questioning, just sort of that longer-term capacity add framework. As you guys think about sort of how much capacity you want to procure and maybe stretch it out over the next several years? Like what are you looking at specifically to inform that decision? And then what gives you confidence in being able to fill that capacity over the next several years?

Yes. Thank you. We look at many, many factors, but the dominant one is we look at our customer demand, look at what they're dealing with, how they are projecting their needs. So that is a big, big input factor for us.

The second one is we look at the footprint from the perspective of for inferencing, obviously, we need to have a really good geographic spread co-locating and for all of our new data centers, we have both core cloud as well as AI capacity, all running on the same server stack. So that's an important aspect for us to have all of these things colocated.

The third thing we always look at is how we are going to keep up with the generational leap frogs of OEMs, including AMD and NVIDIA and perhaps others in the future. So these are all important factors that we take into account as we consider how our footprint is going to look like over the next several years. And we are always making this evaluation, we are looking at various options as we build out our long-term plan. And as I said, primary driver is always looking at our customer needs, customer demand, what kind of workloads are they ramping up. The demand for their application is a big driver for us. So those are some of the input factors that we use to plan our capacity.

Got it. Just a quick follow-up for Matt. Does the [ 27% ] EBITDA margin and free cash flow guidance contemplate any additional capacity investments next year? Or is that just reflect some of the 31 megawatts you're bringing online this year?

It's just reflective of the 31 megawatts that we're bringing on this year.

Our next question comes from the line of James Fish with Piper Sandler.

Maybe just following up on that. If AI is growing as fast as it is, you guys are needing to bring on capacity now to meet all this demand, aren't you going to need more capacity then? And Matt, additionally, it looks like you're excluding finance leases and your free cash flow metric. Why treating it like this as if it wasn't finance, you'd still have CapEx, and it does seem to imply, I'm getting a lot of this question premarket here.

It seems like you're implying about 10% reported free cash flow on '27. So can you walk us through that? And I know this is a loaded question, but A lot of those that are providing lease servers are implementing memory cost increases. So I guess how are you thinking about what commitments you actually have from them and potential pass-through of memory costs?

Yes. Just I'll take that in reverse. So yes, we've seen increased component costs, the same as others in the industry, and that's all reflected in our guidance. And again, it hasn't changed our return expectations or the economics that we'd see. It's just it means that there's more costs associated with some of the service that we're bringing on. But this is -- I'm glad, Fish, you brought this up, which is you got to think about our free cash flow in tiers, right?

So you say, okay, well, you got unlevered free cash flow, which, again, people should be using from a valuation standpoint, and that we're talking about being in the 18% to 20% range. When you add the interest expense, you get the levered free cash flow, which is what we've historically -- that's our adjusted free cash flow margin. And you're only you're only giving up a couple of percentage points there. And that interest right now is half like the TLA and it's half equipment leasing.

And then as you point out, you have the principal payments that are more of a financing transaction. That's why they're not captured in either the adjusted free cash flow or leverage adjusted free cash flow. But if you take those financing transactions and if you're going to lump everything in it and you say, what about the mandatory prepayment of $25 million a year on your term loan, okay? We'll throw that in there.

If you take all of the cash payments, including the principal payments, including the prepayment of the Term Loan A, so that's all financing stuff. So again, you're mixing metaphors here. So if you throw that all in, we're still generating cash. So you're saying, hey, well, it's [ 10]. I'm like, hey, it's [ 10]. I think we're generating cash. while we're accelerating the growth of this business into the 30s and on an unlevered free cash flow basis, it's 18% to 20%.

So it's a testament to our ability to dramatically accelerate growth. We've taken growth from 11%, 12%, 13% to guiding to 30%, and we're generating incredibly strong unlevered free cash flow. We're generating very strong levered free cash flow. And if you throw the kitchen sink in there and all the payments that we have to make, we're still generating cash. I mean that's an incredibly strong position to be in, and we have a very flexible balance sheet. So we feel very good about the cash generation that we're setting out while we're delivering this growth.

Yes. I mean the growth acceleration looks good. And Paddy, for you on Slide 20, I got asked a couple of questions ago to a degree. But probably you point out the difference between you guys and Neocloud and inference wrappers. And maybe being humble about it, you point out that you're about 75% of the way in the first 3 categories. And so is this something that we should be expecting to hear about at the April event? Or what do you guys need to do to get to that full 100% difference?

Yes. Fish, I don't know if I will ever call myself 100% in those things because that market is changing so fast. Like if we ask 5 of our customers today, what they want versus what they thought they wanted 3 months ago is meaningfully different, right? Because as they are going into their customer base and deploying their solutions, new things come up all the time. The capability of AI models evolve all the time.

So this is going to be a moving target for the next couple of years. But the first part of your question, absolutely, that is where our R&D team is super heads down, inventing new technologies, inventing new parts of the stack. So you will hear a lot more about this on April 28. But I would say this is where I feel very confident that we already have a lead and that lead is only going to grow over the next few quarters.

Next question comes from the line of Thomas Blakey with Cantor Fitzgerald.

Congratulations on a great quarter and a great outlook here. Maybe some follow-ups to my peers. Paddy, you mentioned, I think it was to a previous question about demand outstripping supply and giving you great visibility that you've kind of alluded to in this call, not expecting you to give calendar 28 commentary if you wanted to because like you looked out 2 years on the April 25 call, that would be great. But in addition to that, I'm interested in what you're seeing in a pricing dynamic. If demand is outstripping supply, you're lining up these new AI natives. Just maybe some commentary on pricing would be helpful.

Yes. Thanks, Tom. So I think we have already talked about what we are going to talk about in -- for 2027. But in terms of the -- yes, the demand is clearly there, and we are moving as fast as we can to first deliver on these 3 data centers that Matt talked about.

From a pricing point of view, it is -- we have competition from all kinds of different players and the pricing is holding. And in some cases, it has gone up. And we are very, very attuned to what is going on in the market. And there is a lot of scarcity of supply across the board. So we are also in a position where we work very closely with our customers to ensure that we are calibrating the prices that we have, both on demand as well as contractual prices to keep pace with what the market dynamics are at this point.

But I would say nothing has materially changed. And the pricing is also a function of the generation of the GPUs that we are talking about, right? At the lowest level, if a customer wants access to GPUs, it is priced GPU dollars per hour. And at that layer, it really depends on the generation of the GPU, whether it is [ Blackwell ] or the [ hopper ] series for NVIDIA or the 350, 355 from AMD or the 300 or 325.

So it really depends on the nature of the generation. There are also other dependencies like the cluster sizes, the cluster configuration, what kind of networking they want and so forth. And as you move up stack, if you look at my Slide 19, as -- and each -- the one thing that I did not mention in Slide 19 is that customers can enter our stack at pretty much any layer of the stack, right? So the higher up you go in the stack, you're not pricing by per GPU hour, but your pricing per token.

And there, we have a lot more degrees of freedom in terms of how we price versus competition. Because there, you're doing dollar per token, but also you have the flexibility of running it in different types of hardware. You can also change up the AI model that is servicing this token request. So we have more degrees of freedom in customers, some customers need that flexibility, and they are willing to live with the higher orders of the stack rather than dictating which generation of hardware they want to run in.

Right. That's super helpful, Paddy. And just maybe as an extension of that flexibility, it was impressive to hear about the 0% churn in the large $1 million-plus cohort with 115% NRR. I'd love to know what the overlap there is with regard to the AI native exposure, if you could maybe kind of talk about just those customers and how much of that is from AI and for Matt relatedly, is -- and you're improving -- are we finally including AI and ML revenue there? And if not, when can we expect that?

Yes. Thanks, Tom. So it's about -- on a customer account basis, it's about half of the $1 million customers. Our AI customers and half are core cloud or general purpose cloud only. It's a little bit more on a revenue basis or an ARR basis, a little bit more AI, but not a lot. It's not too far off of 50-50. And as you saw in the materials, 3 or 4 -- 48% of the trailing 12-month incremental ARR is coming from those -- from AI customers. So that's kind of how the split is.

In terms of the -- no, it's not in there yet. And the reason that we disclosed the AI customer revenue, and we will continue to disclose that as a metric in the growth rate. And also looking at the RPO, which is, again, a decent chunk of that, not all, but a decent chunk of that is also AI. We're trying to give you better leading indicators of the performance of the AI customer base.

The [ MDR], if you look at some of the charts that we showed with some of the bigger inferencing providers, those -- they just got started on the platform in kind of the June, July time frame. And there's a big difference in the, I'd say, the size and caliber of the customers that we've been winning in the last 6 months on what now 7, 8 months, I guess, on the AI side.

Those, we think, will have more of your traditional kind of NDR like characteristics where they grow and expand on the platform using inferencing, which is more of a production workload versus a lot of our earlier customers were smaller customers doing experimentation, doing projects, and they just don't look like revenue was growing like crazy because we would be adding a ton of those customers.

But if you look at any of the individual customers, it was it was hard to see a pattern. And what MDR is a SaaS metric is it looks for, okay, it looks for patterns where you bring on a customer and you can expect them to do XYZ over the next 12 months. And we just didn't see that. There's no noise in our AI customer revenue kind of lumpiness early that we see changing. So we'll continue to evaluate that every quarter and at the appropriate time, we'll contemplate rolling that in. But it's probably still 12 months away.

Our next question comes from the line of Patrick Walravens from Citizens.

Congratulations on the quarter. And I have to say congratulations on the slide deck. It's fantastic, and I'm sure all of your investors are going to appreciate it. So Paddy, I was looking back at my note from 2 years ago when you joined and at the time, one of the things you said was that our durable competitor differentiator for us long term is going to be in the software layer.

And you said you were focused on bringing simple, easy-to-use AI ML capabilities on both hardware and software to developers. So what I'm wondering is, as you look back -- and you've got -- and you're growing 11% when you joined and decelerating, right? So as you look back, what parts -- which of the growth drivers that have caused you to accelerate, now we're talking about 30%. Did you anticipate and which were fortuitous is probably the wrong word, what favors the prepared, but which were sort of unexpected?

Yes. Thank you, Patrick. I would say what was pricing -- and maybe I'll take some creative liberty in answering your question. So what took a few quarters for us to get right was -- as I mentioned several times during this call, we had -- we had a constraint in keeping up with customers that were scaling rapidly and scaling big on our platform that I joined.

So it took us a few quarters to really understand, get to the bottom of their needs -- and there was a lot of work that had to be done for us to get to the 0% churn that I was so proud to share with all of you this morning. So that took a lot of engineering effort. And I'm super proud of my team and it's a lot of very complex technology work all the way from advanced networking to fortifying our storage to inventing new things in our database offering and so forth.

So that took a tremendous amount of heavy lifting and that job is not done yet as we get to -- we started with 100,000 customers, then we focused on 500,000 customers. Now we are focused on million-dollar customers. And who knows in the next couple of years, we'll be talking about $5 million and $10 million customers. So that bar racing is an ongoing endeavor for us.

And on the more fun side of things is literally participating from the starting point with the AI native ecosystem. So we are learning at their learning, and we are inventing alongside them, and that is a great luxury to have because we feel like we can write their growth curve and as their needs increase, and they're learning the right way to do this from a workload perspective. We are just trying to keep up pace and they're super appreciative of us inventing on their behalf to make their lives easier so that they can focus on their domain and invent new things for their customers. So we'll share a lot more of this on April 28, but that's how I would answer your question, Patrick.

Next question comes from Mike Cikos with Needham & Company.

Congrats on the strong growth [indiscernible] you're providing us. Matt, if I could just come back, and I know that the free cash flow topic has come up a couple of times here, but you can see as well as anybody, just how sensitive the investors are in this market to the AI CapEx investments that are required or different financing vehicles that are out there.

Just to be clear, when we look at the calendar '26 versus the calendar '27 guide, that unlevered to just free cash flow or just the free cash flow guide the 3-point delta is expected to widen to about 10 points in calendar '27. If we take that one step further, and I know that your guidance for -- or those guardrails for '27 currently don't contemplate additional capacity coming online.

But it seems fair that we should be assuming more capacity. And if that's the case, would those different -- would that delta between the unlevered and the levered free cash flow margin widen further from there? Is that fair?

The way I think you got to think about it is, again, if you're looking at the levered free cash flow, it's got other stuff in it besides the equipment leases. It's got [ TLA ] interest. It's got other things. If you look at the -- as Fish was saying, if you look at the other cash, there's mandatory prepayments of the term loan A.

So you got to be real careful about what you're using for what purpose, right? So if you said, hey, what's the steady state cash flow generation capability of this business. Again, because we lease equipment, we don't have an upfront capital requirement that makes it super lumpy. We can make that smoother and we can grow.

However, when you're growing a business even with that model and you're adding data center capacity, you have a couple of months where you're actually taking data center lease expense and you haven't generated any revenue when you lease year unlike if you buy gear, you put it in your warehouse, you actually don't expense it until you actually deploy it. When you lease gear, you start that lease expense as soon as it's shipped. So you have front-loaded cost that don't catch up the revenue right away.

But because you didn't have a big giant slug of capital, as soon as revenue starts generating, you're immediately generating cash and you're improving your margins with utilization. So the steady state, like if you said, that's why we've been very crisp about what's included in the numbers. It's to give you a sense of what the margins look like on a steady-state basis. If we just continually assume, well, we're going to add incremental capacity, which I can't tell you how much incremental capacity we're going to have because we haven't contracted it, and we haven't committed anything to incremental capacity.

So what we're showing is when we add 31 megawatts as an example, and you roll that forward a year, you have incredibly strong cash flow characteristics to that. And there's going to be a short-term impact on gross margins and net income because of the timing thing I described, but that works itself through relatively quickly. And so you would expect that as we saw other opportunities to accelerate our business with similar economics, that we would make similarly good decisions and that engine will keep going.

And so it's -- I view it in a very different way than what you're describing. I view it as we're going to commit to more capacity. It's because we have more growth opportunities and the returns are incredibly compelling, and we're doing it in a way where we match the revenue and the costs, and we're not going out above our skis beyond our skis and making massive commitments chasing the data center and GPU arms race.

We're doing it methodically. We're doing it where we have an advantage, where we earn a good return and we're able to do it. Well, again, taking 13%, 11%, 13% revenue growth to $30 million while still maintaining really good margins. So we're really excited about the potential we have and the economics that we're delivering.

Maybe for a quick follow-up here. Understood on the accelerating growth you guys are looking at throughout calendar '26, just based on the megawatts coming online. One thing I wanted to ask, again, I'm sure that you guys have your own models as you're looking at the AI customers ramping, but to drive that 25%-ish growth exiting calendar '26. Can you provide any additional color for what you're assuming in terms of ARR directly from those AI customers, if I'm thinking about the $120 million that we see today exiting '25?

The only thing I would say is what we said is that AI customer ARR in Q4 was $120 million grow at 150%. We have more demand than we have supply. We're bringing on supply. You should expect that it doesn't slow down.

Our next question comes from the line of Mark Zhang with Citi.

Just given the strong demand environment, should we see more capacity comments coming, I guess, like you announced today? And this -- if that's not the case and is there enough incremental capacity or [indiscernible] capacity in your current footprint to support continued growth. Just any insights there will be appreciated.

Sure. So Mark, as we said, there's enough capacity in committed capacity. There's enough growth potential in the committed capacity to get us to 30% growth in 2027. Clearly, we're very cognizant of the data center market and very active in terms of the evaluation of that. We haven't made any commitments at this juncture to share with the market.

And if we get to a point where we make a commitment, we'll certainly share that. But at this point, again, we thought it was incredibly important for people to understand how to digest capacity as we bring it on. And that's why we've guided to what we have based solely on the 31 megawatts we've already committed and it gives you a good sense of how it ramps and what the economics are. And should we bring down the incremental capacity, you'll have a good model to add on to the growth ramps that we've already articulated.

Okay. Great. And then maybe related to that, can you -- is there a sense of utilization of your current estate, maybe like given in terms of the what the current capacity is we know the current capacity, could there maybe any sense of the contracted capacity that you have on the books?

Yes. So from a contracted capacity standpoint, again, if you're talking about data centers, we've got 31 megawatts that we're adding to our roughly kind of, call it, 43 or 44 , which will put us at 70 -- just about 75 megawatts when we're done. So the 6 megawatts -- so we're sitting at, call it, sitting at 43 and we're adding 6 that will come online or generating revenue in the second quarter and the balance of the incremental 31, which is about $25 million will come on and start ramping revenue and in the second half.

And we expect to reach -- whether we are at full utilization as a function of whether we decide to fill them all with GPs right away or we do it over time because we'd like to strike out the generations of GPUs. We don't like to go all in on one type of a generation of GPUs but we'll be at a very healthy utilization in -- at some point in 2027, which is enabling us to get to that 30% growth.

At this time, we have no further questions. That concludes our Q&A session and today's conference call. We would like to thank you for your participation. You may now disconnect your lines. Have a pleasant day.

Awesome. Thank you, everyone, for being here at the UBS Global Technology and AI Conference. My name is Radi Sultan. I cover the SMID-cap infrastructure software stocks here at UBS. Next up, we have DigitalOcean, Paddy, CEO; Matt, CFO; and Melanie runs IR. So first of all, thank you very much for being here today.

Great. Thank you so much for having us here. It's one of those conferences I really look forward to coming from Boston or living in Boston like that a couple of days, get out and enjoy some warmth in December.

Awesome. Awesome. Yes. So maybe just get started. You recently put out an impressive 18% to 20% growth outlook for next year, a full year ahead of schedule from the guidance you gave at the April Analyst Day. So maybe just to level set, I'd love to hear from both of you on what's driving the recent strength? What's changed since April that you felt comfortable pulling that forward? And maybe we'll kick it off there.

Yes. So thank you for a great lead off question. So April just feels like an era away. It's only been 7 months, but a lot has happened in the market since then. So when we look -- when we put out the 18% to 20%, we were coming off of a year where we had a couple of different priorities. One is on the core cloud, as we discussed in our Investor Day in April, our biggest priority was to take care of our customers with the largest workloads, right? So that was priority #1 for us. And for that, we had to address some of the core product gaps that we had as well as build a go-to-market that was complementary to what we believe is the industry's best product-led growth machine.

So those were the top priorities to fix or address the needs of our customers with the largest and the most sophisticated workloads on cloud. The second one was to incubate the cloud business. So if you fast forward 7 months, not only did we address the issues that we had with some of our largest customers or the biggest workloads defecting from our platform and going into hyperscaler clouds, we took that weakness and turned into one of our biggest strengths. And that's why we have started reporting our 100,000-plus customers, which grew at 41% last quarter. And on top of that, we also talked about our $1 million-plus customers growing at 72% and becoming a big part of our customer base.

So we took what was once a weakness of our portfolio and now have started turning that into a strength of ours. So that's number one. And the cloud business at that time was fairly small. But when you string together 5 quarters of 100% plus growth every quarter, it becomes a fairly sizable part of our business. And the combination of these 2 things is what is giving us the confidence to say, hey, we'll take the outlook that we provided for 2027, and we are very confident that we will get there a full year ahead of schedule. And the additional thing that gives us confidence is the fact that now we have also announced that we are going to take 30 megawatts of extra capacity from a data center point of view to accelerate our AI deployments. So it's a combination of these strong business fundamentals that's giving us the confidence.

Awesome. Paddy, you called out multiple 8-figure deals just in the month of October following the most recent earnings call. So very strong relative to the size of the deals you've signed in the past. So maybe you could just speak to what sort of been the biggest driver of that traction? What's the mix between core cloud and AI? And maybe any trend in sort of the nature of those end customers and how you landed those deals?

Yes. So this is a relatively new muscle for the company. As those of you will appreciate that have followed the company for a long time, we have barely had any RPOs to report, primarily because of the fact that we grew to just shy of $1 billion of run rate in the shoulders of 640,000 paying customers. So you can do the math. There's a lot of small customers that have really built up the business to date. But over the last 3 or 4 quarters, that has started changing quite appreciably. So we have these large 6-figure, 7-figure and now 8-figure deals, both in AI and core cloud. So let me circle back to last month. In the earnings call, we talked about the fact that Q3 was the highest organic net new ARR add in the company's history at $44 million.

And we also mentioned that less than half of that came from AI. More than half of that came from core cloud customers. And when you think about the 7-figure, 8-figure commitments we have, it is really a blend of all of these things, right? So we have some AI native customers that are willing to commit themselves to multiyear 7-figure, 8-figure deals with us for -- and the vast majority of that comes from AI infrastructure AI platform, but also some core cloud consumption as part of their commitment, but also from customers that are driving our core cloud consumption. So in the last earnings call, I talked about a long-time customer of ours called Bright Data, which is a leading web data provider, have significantly increased their footprint on us because their business is thriving as being one of the leading providers of web data to the LLMs, the leading frontier models of the world.

So we are getting a lot of good tailwinds, both from AI native companies that we are building relationships with and nurturing them, especially in the inferencing world, but also with some of the product gaps that we have closed in the core cloud over the last 12 to 18 months is helping some of our cloud customers to increase their footprint on us, especially repatriating some cloud workloads from hyperscalers to the DO platform.

Got it. Got it. Maybe just drilling down into the AI side. You talked about AI revenue reaching mid- to high teens by the end of next year. Can you just talk through where you're seeing the biggest uptick in AI demand and maybe how broad-based is that?

Yes, it is quite broad-based. So if you think about our footprint now, it is just shy of 10% of total run rate. And if we keep stringing together these 100% plus year-over-year quarters, it is not inconceivable that we will get to that number that you just mentioned by next year. So we feel really good about the kind of traction we are also getting. So as I just mentioned, a vast majority of our AI revenue comes from infrastructure, but most of that comes from inferencing workloads. I would say most, if not all of it comes from inferencing workloads.

The reason why that is important to us is twofold. One, it enables us to build direct customer relationships with these AI-native companies. Number two is by the nature of inferencing, these companies are in post-product market fit. So they are no longer burning venture capital money trying to find a good niche that they can fit into. By definition of an inferencing workload means that an end customer, whether a B2C consumer space or a B2B enterprise, an end customer is paying for these workloads. So that gives us the confidence that the investment we are making in time and money in these advanced AI native companies is going to help us build a very durable business for the long haul because unlike some of the other neoclouds, we are not just taking excess spillover bare metal as a service capacity from the hyperscaler because we think that's not the most durable way to build our business given how rich of a software stack we have, we feel like building that durable relationship with end customers will give us the most runway in building a very, very strong business for the long haul.

Got it. Is there any type of sort of end customer use case driving an outsized portion of that AI demand? Maybe how much is AI native versus sort of more traditional?

So most of it is AI native. When I say AI native, these are new generation of silicon -- mostly Silicon Valley companies that have emerged in the last 12 to 24 months disrupting a B2C or B2B space. So one example that I gave in the last earnings call is a company called Fal.ai, which is emerging as the hugging face of generative media models. So they have some of the world's most bleeding-edge generative media models used by companies like Shopify and others to improve e-commerce shopping cart conversions. And the reason why that is really important is it's a real use case that is being consumed by large digital native and other e-commerce companies.

And that is one example of a B2B use case. There are also other B2C examples where companies are building digital characters or AI characters that consumers want to interact with either stand-alone or in the context of a gaming system and things like that. These are all real use cases that customers are spending money on. So for us, it is a good combination of some B2C use cases, but also some B2B real enterprise traction.

Got it. And maybe just -- I know you mentioned this, but most of the AI revenue is coming from the infrastructure layer today. You have a full stack AI offering. So maybe you could just talk through how you expect the mix of that AI revenue to sort of change over time?

Yes. Right now, we have -- even in the infrastructure space, we have multiple points of monetization. So 6 months ago, I would have said most of our revenue comes from Bare Metal as a Service, just like how most of the Neo clouds today only have a Bare Metal as a Service offering. But for us, a lot of the infrastructure consumption has moved from bare metals to what we call as GPU droplets, which is a layer of abstraction that we have built on bare metal. And we charge a premium for it, and most of our customers prefer to use GPU droplets because of some of the performance enhancements we provide.

In fact, the performance is on par, if not better than bare metal access, but it also takes away a tremendous amount of headache associated with building the image, managing the image, managing the life cycle of the infrastructure, the observability, the resiliency and the availability of these droplets is significantly above what they would have to do managing bare metal. And they're willing to pay a premium for that. So that's another point of monetization for us. The next one is we have a class of AI native companies that say, yes, I'm tired of just using GPUs directly. Can you give me serverless endpoints for these types of models, whether it is an open source model like DeepSeek or Qwen or something like that or Llama or a closed source model like Anthropic or OpenAI. So we have serverless endpoint as a point of monetization.

On top of that, we also have a variety of building block services in our platform layer like guardrails, knowledge base, observability, agent evaluation, agent templating. We have a bunch of middleware modules to help companies build and run agentic software. And in the last earnings call, we announced that we have 19,000 agents in just like 6 or 7 months of being in production that companies are using to deploy agents in their enterprises. So we have different levels of monetization. And most of this also dragged through or starting to consume our core cloud services all the way from Kubernetes to object storage to database storage and everything in between. So we have different ways of monetizing our AI stack.

Got it. Got it. Maybe just turning to Matt. AI financing has obviously been very, very topical lately. You guys introduced equipment leases for the first time this past quarter. A big investor question I get is around the role of those leases going forward when you think about the financing side of things. So maybe you could just talk through how you think about financing that AI build-out, the role of those leases going forward and how you see that?

That's great. Yes. We've been talking about this for multiple quarters now. And we've got, like Paddy said, almost $1 billion of ARR. We got really good margins at like 60% gross margins and low 40s EBITDA margin and high teens free cash flow margin. And one of the questions that we would get from investors is, okay, well, what happens when you need to accelerate? And if you're going to lean into AI, is that going to reduce your free cash flow margins? And what we've always said is, no, there's a lot of different ways you can finance gear and get access to capital without messing your free cash flow margins. And we've been able to tap the equipment financing market, both working with some of our relationship banks some of the OEMs. We're getting really good terms, and we're very, very comfortable with that. We also have access to additional capital sources that are interested in putting capital to work in this space around equipment financing.

And for us, it's straight equipment financing. Literally, we just pay over time, and we own the gear at the end. It's dollar cost buyout, and we're able to better match our outflows with our revenue. So it's a great way of accelerating the growth of the business. And the evidence of that is when we pulled in the free cash flow -- or sorry, when we pulled in the revenue guide by a full year to 18% to 20% next year, we were able to also guide that, hey, we'll still be in mid- to high teens free cash flow while we're accelerating our GPU investment, we're bringing on 30 megawatts of new capacity and still generating very, very good margins across the board.

Got it. And you don't think any of the recent sort of AI financing concerns that sort of led through to your ability to access capital and sort of those range of options?

No, we've demonstrated that we have a lot of the market available to us. We just did a $625 million convert earlier this year. We raised a very, very attractive bank facility, $800 million earlier this year. We've tapped the equipment leasing market. And yet we still have really good leverage and generating really healthy free cash flow margins. So I think unlike some of the other players that are involved in the space, we have an existing business. We have almost $1 billion business generating free cash flow. We're a slightly different credit profile than maybe some of the other folks in the space who are leasing just GPUs to rent them to hyperscalers or someone else. It's a different credit profile.

Got it. Got it. Maybe just on the capacity addition side. You've talked about adding 30 megawatts of capacity next year, a big step up from the 40 to 45 megawatts that you guys have today, which supports just under $1 billion of ARR. So maybe just how should we think about your ability to monetize that incremental 30 megawatts and how that compares to sort of how you're monetizing the existing footprint today?

Yes. Well, it will be a healthy mix of GPU in that investment in the 30 megawatts. But we tend to think about it as a portfolio. We've got 43 megawatts of existing facilities. The 9 most recent megawatts we added in Atlanta had both GPU and core cloud. We'll have GPU and core cloud in each of the 3 facilities that make up the 30 megawatts that we'll bring online in first half. And so if you again, look at us and you compare us to, say, a neocloud, and I know we've had a lot of conversations today about, okay, how do I do the math? You're adding 30 megawatts, how many dollars of revenue are you going to get from that?

I'd say, look at our dollars per megawatt across the portfolio and just based on the 2025 consensus. And what you'll see is that clearly, because we have core cloud and AI and AI is still a small portion of our revenue, that we have dollars per megawatt in revenue that's materially higher than what you're seeing from some of the public numbers that you'll see around the neoclouds. And when we expand that, clearly, we'll mix in a bigger blend of AI than we have historically. So the number won't be the same as it is today. It will be lower. But if you look at the yield that we'll get relative to what some of the other folks are getting on a dollar per megawatt because we've got higher layer AI services and higher margin, we've got the pull-through of the core cloud, we expect to get a healthier dollar per megawatt than perhaps what you're seeing elsewhere in the market.

Got it. Got it. And maybe just are there any other big puts and takes when you think about maybe CapEx per megawatt and then maybe anything else on the revenue side that would longer term prevent you from trending more towards where you're monetizing existing capacity today?

Well, I think if you look over a very long time, you think that the reason that we can get so high of dollar per megawatt on the core cloud is it's a very established market, and there's a lot of density you can pack into that. And clearly, there are some characteristics of AI that are going to make that lighter in the near term. They take up more space and they take more power, and so you're not getting the same density. But again, the goal for us is we're not in this just to lease GPUs to people. If that's all it was, that's not our game. That's a scale game that we're not participating in and it's mostly training oriented.

For inferencing, to be an effective inferencing provider, you need to provide the full suite of capabilities because all the inferencing applications need storage and bandwidth and database and they need higher layer services, as Paddy talked about, they don't all want raw access to GPU. So we believe that our investment in GPU infrastructure is going to bring a much bigger kind of revenue pie than, again, what you may be seeing elsewhere in the market.

Got it. Got it. Maybe just when we think about supply constraints have been a big issue in the broader industry around bringing capacity online. Can you just talk through how you were able to secure capacity? Maybe any big constraints you see to bringing that capacity online and maybe how you're thinking about that?

Yes. The capacity that we've taken under contract, like we said, it's 30 megawatts. It's across 3 different facilities. And this is a very similar dynamic to what we saw with broad GPUs, 6, 12 months ago, people were talking about GPU shortages and how long it was taking people to get it. We didn't really have a problem, and it's because we are buying at a smaller quantum than what maybe the hyperscalers or some of the larger projects were undertaking. And ours were a lot more flexible with respect to the space. For inferencing, you don't need giant clusters. You don't need to have all of your GPUs in one place.

So we're able to take 30 megawatts, again, in dispersed kind of form across 3 different data centers, all in the U.S., but that gave us a lot of flexibility. And when you get above 50 megawatts, it starts to get a bit more competitive because the hyperscalers are taking down capacity in those kind of chunks, but they're typically not taking down a 6 or 10 or 15 or those sized facilities. And there's enough colo activity in the space right now where there's third-party colo providers that are still putting money to work and have capacity available that we've been able to get the supply. And then on the back end, GPU capacity has not been an issue for us either. We've got a variety of global OEMs that we leverage and we buy both NVIDIA gear and AMD gear. And you certainly have to order in advance, a couple -- you got to order like 4 or 5 months out in advance, but there's no restrictions in terms of the amount of capacity that we could get at this point.

Got it. I mean we talked about 30 megawatts of capacity for 2026, but maybe how do you think about that longer term? What do you look for when you're adding capacity? How much do you look to have a secured commitment before you make that investment? And maybe you could just talk through sort of the longer-term capacity planning algorithm.

Yes. Well, as Paddy said, the reason that we were able to make the 30-megawatt commitment now and to accelerate revenue growth is we have better visibility into our demand than we've ever had, right? And we've got a number of large committed contracts. We've got really strong pipelines for both AI and some of our large core cloud customers. And that gave us the confidence to secure the incremental 30. Given the pace at which data center capacity is being taken down, clearly, we had to order enough to not only serve the capacity demand we have today, but give us room to grow into 2026 and early 2027.

But I can tell you, we're already out in the market looking for, okay, what are we going to do in 2027? What are we thinking about in 2028? And we're certainly paying a lot of attention to the capacity -- projected capacity requirements and to what's available in the marketplace. So it's certainly causing, I think, the whole industry to think years in advance rather than 6 and 12 months.

Got it. Maybe just on the margin side of the equation, Matt, at the Analyst Day, you talked about a few levers you had to maintain margins even at the gross margin layer. So maybe just how are you managing margins going forward? And maybe in the event that you do get more AI revenue coming online sooner, how you sort of manage that margin offset?

Yes. That's a great question. So we still have, I'd say, margin capability inside gross margin. Gross margin, think of it as the 2 biggest cost structure elements there are depreciation and space and power. Depreciation is just cost of equipment. So as those prices become more competitive, you'll get better margins, but that's not something that we control in the near term. The one thing that we do control, though, is we can optimize our data center footprint while we're expanding our capacity. So we're going to take down an incremental 30 megawatts of capacity in bigger chunks than we ever have, right, in kind of 9, 10, 15-ish size. If you looked at our existing data center footprint before we implemented our first AI-focused data center, it was a bunch of 2.5 megawatt facilities, and they're in really expensive markets.

So they're in New York, they're in San Francisco, they're in Toronto. And then globally, they're in really pricey locations. We can optimize out of some of those smaller facilities with high prices and into -- just consolidated into some of the bigger facilities in the Midwest or in kind of second-tier cost markets. And so that's on the gross margin side. On the operating expense side, I think we've demonstrated a really good ability to control costs. We have been investing in R&D and in sales and marketing, but we expect to get operating leverage over the coming years. We're certainly driving a lot of operating leverage in SG&A. in the G&A side of that.

And we certainly see operating expense improvements that we can drive. We've hired, I think we're up to about 150 people engineering in Hyderabad in India. And so we've been leveraging our global cost structure for that. And then as always, we're very protective of our free cash flow margins. And I think the evidence there is the mid- to high teens guide for next year while we're accelerating our revenue growth.

Awesome. Paddy, maybe just on the hyperscaler migration opportunity, like it seems like that's really picked up recently. So can you talk about why that's seeing an uptick? And specifically sort of what's changed around the product suite that's driving those customers to migrate from the hyperscalers over to your platform?

Yes. So there are 2 primary reasons why we are seeing an uptick in this. One is all the product features that we have shipped over the last 4 or 5 quarters. And broadly, you can think of some network enhancements, security enhancement, observability and manageability enhancements and things like that. So that's one reason why. The second one is we have completely overhauled our go-to-market machine, all the way from having technical account managers who are emphasize on the word technical, like they're really technology folks who are managing the relationships with our large customers, giving them the assistance that they require to move migration workloads. We're also expanding our systems integrator partnerships where some of them are building practices around helping migration of workloads to DigitalOcean, investing in the appropriate tooling required to help accelerate some of these movements.

But if you take a step back and think about this, multi-cloud is a thing that is here to stay, right? Most companies, whether they're digital natives or brick-and-mortar enterprises, everyone has a multi-cloud posture at this point. And until about 1.5 years ago, we were not even in the conversation to be an active participant in this multi-cloud world. But now we are. There aren't too many public clouds. There may be exactly 5 public clouds with our posture and footprint. There's Microsoft, Amazon, Google, OCI and DigitalOcean, right? That's pretty much it. If you want to have a reasonable footprint on the public cloud, it's only 5 clouds that you can deploy to.

And most companies will pick one of the big 3, and then we become a very natural second or third cloud for these large deployments. And one specific feature that has been a really big enabler of this is the direct virtual private cloud connection. Now we enable between our data center and Google and Amazon Cloud enables our customers to have very sophisticated workloads split across clouds. And that is a big unlock. And I can keep going on these things, but it's that one-two punch of fixing some of the product gaps that we used to have and investing in the appropriate type of go-to-market function to give our customers a good, smooth on-ramp to increasingly adopt DigitalOcean as a proper multi-cloud option.

Is there any particular workload or maybe customer type where you're seeing the most success today? Or is this sort of you think broad-based?

I think it is mostly -- it is very broad-based, but we are still staying true to our focus on digital native enterprises. So we still don't go after on-premise deployment and try to move them to the cloud because there's just so many other things that come with it, like legacy workloads have a lot of center of gravity that we are -- at this point, we are not focusing on. They also have other compliance and privacy issues and things like that. So we are focused on digital native companies. And there are enough of them, as I explained in our Investor Day in April, there are enough of them in the world that are big and thriving. And these are also companies that are in the bleeding edge of AI adoption. So if we are able to focus on this over the next couple of years and really nail this, we'll have plenty of market share to take.

Got it. And you said something in an earlier meeting around the pitch for DigitalOcean has really changed since you joined. And I was wondering if you could elaborate on that and maybe just talk us through that journey.

Yes. So the pitch for DigitalOcean, especially when we are talking about attracting world-class caliber leadership leaders to come and join us in our journey, it has completely transformed in the last 20 months, right? 20 months ago, my pitch was around, "Hey, this is broken, you have to fix that." We have a deficit in the leadership bench. You have to come and recruit, blah, blah, blah. And now it is a pitch that is built on a very, very strong foundation. It is a very positive pitch to say, "Hey, look at all the things we have accomplished over the last 18, 20 months." And now we can build on top of this foundation and take advantage of this generational opportunity we have ahead of us with AI. And even in AI, we have some of the largest AI native companies running direct live traffic on the DigitalOcean platform.

So that's a huge plus in terms of our ability to be attractive for world-class talent at all levels of the organization. So it's a really positive shift in making ourselves attractive both to prospective employees, but also to other customers. Nothing is as attractive as showing real-world traction, and that has also really helped us with presales qualification and helping move pipeline through the different stages.

Awesome. And maybe just to wrap it up, Paddy, when we first met a little while back, you mentioned that product is really where it all starts for you. So I would love to get a sense of the next 12-month product road map, what you're most excited about on the product side and then how you're planning on pulling that through on the go-to-market side of things?

Yes. So on the product side in cloud, for example, we have accomplished pretty much most of the things that we set out to do from -- for making our platform complete for the type of digital native enterprises that I talked about in the Investor Day. Now we have raised the bar again, right? So we are now going after more sophisticated worldwide workloads that are in the 7-figure, 8-figure range. So the bar shifts in terms of what we need to build out in Database as a Service, storage, performance and things like that. So that is a very, very attractive challenge for our principal engineers and the architects that are working on our platform.

On the AI side, we are very proud to say that we have one of the best architected inference infrastructure in the world, whether it is bare metal services or our GPU droplets. It is second to none in terms of its resiliency, scalability and performance throughput. And we are winning AI native workloads purely on the back of our performance or flocks per second and things like that. On the agentic gradient AI agentic layer, we have the most comprehensive agent development life cycle, but we are just getting started. There's a lot of other product road map items. We are co-inventing with our AI native customers. So that's, in a nutshell, what the road map looks like over the next 12 to 18 months.

Awesome. Look forward to checking back in this time next year. Thank you so much, guys.

Thank you, Radi. Appreciate it.

Ladies and gentlemen, thank you for standing by. My name is Krista, and I will be your conference operator today. At this time, I would like to welcome everyone to DigitalOcean's Third Quarter 2025 Earnings Conference Call. [Operator Instructions] And I would now like to turn the conference over to Melanie Strate, Head of Investor Relations. Melanie, you may begin.

Thank you, and good morning. Thank you all for joining us today to review DigitalOcean's Third Quarter 2025 Financial Results. Joining me on the call today are Paddy Srinivasan, our Chief Executive Officer; and Matt Steinfort, our Chief Financial Officer.

Before we begin, let me remind you that certain statements made on the call today may be considered forward-looking statements, which reflect management's best judgment based on currently available information. Our actual results may differ materially from those projected in these forward-looking statements, including our financial outlook.

I direct your attention to the risk factors contained in our filings with the SEC as well as those referenced in today's press release that is posted on our website. DigitalOcean expressly disclaims any obligation or undertaking to release publicly any updates or revisions to any forward-looking statements made today.

Additionally, non-GAAP financial measures will be discussed on this conference call and reconciliations to the most directly comparable GAAP financial measures can be found in today's earnings press release as well as in our investor presentation that outlines the financial discussion on today's call. A webcast of today's call is also available in the IR section of our website.

And with that, I will turn the call over to Paddy.

Thank you, Melanie. Good morning, everyone, and thank you for joining us today as we review our second quarter 2025 results. We continue to make meaningful progress on the strategy we laid out during our April Investor Day. Our performance this quarter was very strong. We exceeded our Q3 guidance on both revenue and profitability metrics delivering 16% revenue growth and the highest incremental organic ARR in the company's history, while generating 21% trailing 12-month adjusted free cash flow margins. We continued innovation in our comprehensive agent cloud to support the needs of scaling AI and digital native enterprise customers making sure there is no reason our highest spending customers ever need to leave our platform. We augmented our industry-leading product-led growth engine with focused direct sales motion driving customers to migrate workloads from the hyperscalers to our platform and building traction with direct AI-native customers.

This progress is evident in the rapid growth of our largest customer in their increasing willingness to sign committed contracts with us, with customers having more than $1 million in annualized run rate reaching $110 million in ARR, growing 72% year-over-year. and with multiple customers signing 8-figure committed contracts after the quarter closed. The demand for our agent cloud has exceeded our supply. Our performance and the visibility we have into demand gives us the confidence both to increase our 2025 and 2026 revenue and adjusted free cash flow outlook and to also increase our investment in data centers and GPU capacity to further accelerate growth while maintaining attractive margins. I will now dive deeper into all of this, starting with our third quarter financial results, as highlighted on Slide 10 of our earnings deck. Q3 revenue hit $230 million up 16% year-over-year, marking the highest growth in Q3 2023. We delivered our highest organic incremental ARR in company's history at $44 million.

This growth was driven by a balanced performance across our comprehensive agency cloud platform as direct AI revenue more than doubled year-over-year for the fifth consecutive quarter and our general-purpose cloud products saw the highest incremental organic ARR since Q2 of 2022. We delivered this accelerating revenue growth in Q3 while exceeding our profitability guidance. and materially strengthening our balance sheet. Adjusted EBITDA and non-GAAP earnings per share were both well above guidance on the back of strong execution, and we delivered a strong 21% and trailing 12-month adjusted free cash flow margin as we introduced equipment leasing into our financial toolkit in Q3 to better align the timing of our investments with our revenue. To give us further flexibility to invest in growth, we also repurchased the majority of our 2026 convert in the quarter, strengthening our balance sheet.

The primary drivers behind our accelerating top line growth are threefold: number one, the increasing momentum we are seeing with AI native customers. Next, the material traction we continue to generate with our highest spend digital native enterprise customers; and finally, the continued strength we are seeing in revenue from new customers. Our unified Gradient AI identic Cloud, which is outlined on Slide 7 of our investor presentation, is getting increasing traction with larger well-funded AI native companies that are in inference mode. These scaling companies increasingly leverage our unified agent cloud with many of our top customers already leveraging both AI and general-purpose cloud capabilities and with many more having at least starting to test and experiment with AI on our platform. Evidence of this traction is in the growth rates of our highest-spending customers.

Revenue from these customers who were at $100,000 plus annual run rate grew 41% year-over-year, increasing to 26% of total revenue. Growth is even higher for our largest digital native enterprise customers as the more our customers are spending, the faster they are growing on DL. The charts on Slide 11 shows that our customers with greater than $500,000 and greater than $1 million in annualized run rate grew revenue 55% and 72%, respectively, providing clear evidence that are increasing the ability to not just attract but also retain and grow our largest customers, demonstrating that customers can keep scaling on our platform and never have a reason to leave. Let me now dive deeper into this fraction using Slide 12 as the backdrop to illustrate just how much progress we have made since the last earnings call. I will start with our AI infrastructure on the bottom right, which is a full stack inference platform targeting AI native customers that have their own models that they want to tune, optimize and run in inference mode.

These customers select our platform for our full set of capabilities, where we combined a powerful lineup of GPUs that are available in both bare metal and droplet configurations, including inference optimized droplets with advanced inference performance optimization like page retention, flash attention, FP quantization, speculative decoding model operations management, reduced time for first token and compelling TCO economics. Our AI infrastructure provides comprehensive hardware plus software infrastructure for AI native companies that are scaling up real-world inference workloads globally on D. fal.ai or fall, a generative media model platform that provides text to image and text to video models for major customers such as Cana Shopify, perplexity and more is a great example of a customer that is taking advantage of our unified agent cloud.

They leverage a range of our AI infrastructure solutions, including GPU droplets, both to host their media models in production, serving their end customers and to do research and fine-tuning all is more than just an important customer as we have come together in a strategic partnership to accelerate generative AI concentration by making image and audio generation more accessible to start-ups and enterprise. Through this partnership, Paul will host and run hundreds of its models on digital oceans infrastructure, powering applications across creative and enterprise use cases. This means customers can create agents that understand and generate not only text but also images, data and other forms of input, significantly expanding the range of real-world problems our customers can solve.

New break is another example of an AI-native customer, leveraging our unified at cloud driving the next generation of digital media, new speak delivers timely and relevant local news and information to 40 million monthly active users newsbreaks AI-powered infrastructure mix sophisticated personalization accessible to mainstream users nationwide. They utilize our AI infrastructure to train and deploy complex recommender systems and natural language processing models that are foundational to their product. Our AI infrastructure high throughput and memory capacity are critical for running inference at scale, which allows them to perform real-time content franking and ad placement for millions of concurrent users. Trading AI Agency Cloud unifies our integrated AI capabilities with our full stack general purpose cloud, which we've been optimizing for over a decade, enabling news break to preprocess their work on our CPU droplet and run their vector search service in advance of running their AI workloads, optimizing both cost and performance.

Network File storage or NFS, which delivers high throughput performance for both GPU and and non-GPU droplets is an example of a unified agent cloud capability. Customers can now attach and provision storage in just minutes accelerating time to value by eliminating idle time with seamless integration into our Kubernetes engine, NFS makes it easier than ever to scale applications and workloads while maintaining speed, reliability and efficiency across environments. Moving up the stack outlined on Slide 7. And the AI platform layer on the middle right is typically leveraged by companies that are users or consumers of that are looking to build agentic applications without having to directly manage the infrastructure. As we know, the future of AI is an agent and agentic workflows, which is a natural evolutionary step for all SaaS and other applications. We continue to evolve our AI platform as the foundation for building and deploying these intelligent agents empowering complex enterprise Agentic workflows.

It now supports serverless inferencing across the most popular model, including open AI, Entropic, Mistral, LAMA, Deep Seek and others, including new generative media models from fall. We've added a powerful knowledge-based service that lets customers bring their own data and improve accuracy along with built-in guardrails for safety, visual agent orchestration and enterprise-grade features like observability, gift integration and auto scaling Together, these capabilities make our Gradient AI agency cloud platform, 1 of the most intuitive and complete platforms for taking AI agents from prototype to production. These key capabilities help companies develop and operate AI agent fleet and manage their full life cycle of these agents seamlessly from a single platform while leveraging the best-of-breed AI models from various providers. We are particularly excited about a major customer we signed for our AI platform after the Q3 quarter close.

This customer is a global digital systems integrator who signed an 8-figure per year multiyear contract to leverage our Agent Cloud to drive AI transformation for its digital native enterprise customer base with a specific focus on identifying the full software engineering life cycle, including planning, backlog and road map management release planning, release execution and customer support. We'll provide more information on this exciting customer after we formally announced the partnership in the upcoming days. The AI platform layer continues to also gain broader momentum with over 19,000 agents created so far, of which more than 7,000 are already in production. One specific customer, Sakima an Italian leader in GDPR compliant, ethical and secure AI solutions across Europe chose to leverage the Gradient AI agency cloud over the hyperscalers.

By using our platform they are now able to create and roll out agents to automate customer support, knowledge management and content creation while reducing development time and costs associated with the agent life cycle. This quarter, we also expanded our AI ecosystem with the launch of the Digital Ocean AI partner program with several of our partners outlined on Slide 13. We -- this is a major step in empowering AI and digital native enterprises that are building and scaling their businesses, leveraging AI. These companies don't have time for a fragmented infrastructure. They instead want a unified cloud and an AI platform that lets them seamlessly build and scale intelligent applications using agents. This new partner program brings together AI native companies, integrators and the venture ecosystem to help these builders reach more customers accelerate innovation and amplify their global reach.

Combined with our AI platform and infrastructure, this ecosystem makes digital ocean, the go-to destination for these AI native businesses who want simplicity scalability and reach without the hyperscale complexity. In Q3, we continued to deliver product innovation in our core cloud stack, to support our highest-spending customers by meeting their needs as they scale their business on GO. One such example of a digital native enterprise customers failing rapidly on DO is bright data. a leading provider of web data sets to global Frontier LLM labs for training AI mod. Right Data leverages various components of our Agentic cloud to scale high-volume global workloads on our platform. VPN Super, who develops trusted VPN and security solutions is the most downloaded VPN app in the world is another digital native enterprise growing on our platform. VPN Super empowers millions of users across the globe to browse securely and privately regardless of their location.

They signed a 7-figure D to migrate multiple workloads to distillation and they selected do for our ability to handle large traffic spikes, platform reliability and our global scale. These growing customers require general-purpose cloud capabilities that grow with their business and we delivered a number of these new features during the quarter, as you can see highlighted on Slide 12 of our earnings presentation. For example, we recently introduced Pass cold storage an enterprise-grade object storage solution designed for customers managing data at massive scale with support for hundreds of terabytes and billions of objects per bucket, it offers free receivable, predictable low-cost and immediate access to data, eliminating the trade-off between affordability and performance. This coal storage is secure, reliable and resilient, providing our customers with the confidence to store and access mission-critical data sets seamlessly as their needs grow.

During the quarter, we also enhanced our managed Databases offering with automated storage auto scaling, enabling customers to scale seamlessly as their data needs grow. When capacity thresholds are reached storage automatically scale in 10 gigabyte increments are higher with 0 downtime and no disruption to workloads. This feature is available across -- all major database engines, including MongoDB, Postgres, MySQL and is fully customizable, allowing customers to set thresholds starting at 20% utilization. With a simple Pasco model, autoscaling eliminates the burden of manual intervention, ensuring that applications fail reliably and cost effectively. The steady stream of new features is resonating with our AI and digital native enterprise customers. Over 35% of our customers with more than $100,000 in ARR have adopted at least One of our new features released over the past year, and those customers having adopted at least one of these new products have seen a several hundred basis points increase in their growth rate after adopting the new product.

Our strong performance are growing momentum through the first 3 quarters and the visibility that we now have into demand gives us the confidence to raise our near- and medium-term growth outlook. We are raising our full year 2025 guidance on both revenue and margin and we now expect to achieve our 18% to 20% 2027 revenue growth target in 2026, a full year earlier than we had projected. It has also given us the confidence to accelerate our investments to drive growth in 2026 and beyond. When we outlined our 2027 growth objectives this past April, we indicated that we would increase our investment as we saw opportunities to accelerate our growth. We are now seeing more demand than we can support with our existing capacity, which is evident by us having signed multiple 8-figure committed contracts after the quarter ended that will materially increase our RPO in Q4.

With this increased conviction, we began to put the foundational elements in place in Q3 to even further accelerate our growth. We started ordering more GPU capacity to meet the growing inference demand we are seeing from our AI native customers. We also secured around 30 megawatts of incremental data center capacity to support growth in 2026 and beyond. We added equipment financing to better align our investments with revenue. We ramped engineering resources to accelerate our unified agent cloud road map and continued our targeted investment in new sales and marketing initiatives to complement our industry-leading product-led growth engine. These investments will build on the success we have seen to date and will set us up for a strong 2026 and 2027. Our Q4 and 2025 full year guidance implies a 16% exit 2025 growth rate. And while we won't provide 2026 guidance until our February earnings call, we expect to comfortably deliver 18% to 20% growth in 2026, achieving our 2027 growth target a full year earlier than previously projected.

We will deliver this growth while maintaining strong adjusted free cash flow margins in the mid- to high teens. Matt will provide further color on these investments and the projected impact on our growth and profitability in his remarks. As I said in my opening, we delivered strong performance in Q3, beating our guidance on both revenue and profitability. We are seeing momentum with our unified agency Cloud. And this momentum is evident in the rapid growth of our higher-spending customers. and demand is exceeding our current capacity. All of this gives us a conviction both to raise our 2025 and 2026 revenue and adjusted free cash flow outlook and to increase our investments to take advantage of the opportunity in front of us. We look forward to sharing more on our progress and our outlook for 2026 over the upcoming months. Thank you, and I'll now turn it over to Matt.

Thanks, Patty. Good morning, everyone, and thanks for joining us today. As Patty discussed, we are excited about our strong Q3 2025 performance -- we are gaining traction with our unified Agente Cloud, which is resulting in strong revenue growth from our highest-spending customers, and we are seeing more demand and satisfied with our current capacity. This momentum and visibility gives us conviction both to increase our 2025 and 2026 revenue and adjusted free cash flow outlook and to put in place the foundations to further accelerate growth in 2026 and beyond. In my comments, I'll walk through our Q3 results in detail, share our fourth quarter and updated full year financial outlook and also provide an update on our 2026 expectations. Starting with the top line. Revenue in the third quarter was $230 million, up 16% year-over-year, the highest revenue growth since Q3 of 2023. The -- this growth was balanced across our unified Agente Cloud and was primarily driven by increasing traction with our higher spending AI digital native enterprise customers with steady contributions from our product-led growth engine.

We continued to see strong AIML revenue growth in Q3 with AI revenue more than doubling year-over-year, which is done every quarter since we launched our AI platform. We also delivered the highest incremental organic AR in company history at $44 million, bringing ARR to $919 million. With rapid product innovation across our unified Agente Cloud platform, and with the strategic go-to-market investments we made earlier this year, proving to be effective, we are having increasing success attracting and growing larger, well-funded AI and digital.

This is evident in the revenue from our customers whose annualized run rate revenue in the quarter was greater than $100,000, which now represents 26% of overall revenue, growing 41% year-over-year. 20 basis points higher than the growth we saw from that cohort in the prior year. Adding to this growth, revenue from general purpose cloud customers in their first 12 months on our platform continues to be strong. and we have stabilized NDR as net dollar retention remained at 99% in the quarter, up 200 basis points from 97% in the third quarter of 2024. Turning to the P&L. While we accelerated revenue, we also delivered strong performance on all of our key profitability meters. Gross profit was $137 million, up 17% year-over-year, with a 60% gross margin for the third quarter. 100 basis points higher than the prior year. Adjusted EBITDA was $100 million, a 15% increase year-over-year and an adjusted EBITDA margin of 43%.

The non-GAAP diluted net income per share was $0.54, a 4% increase year-over-year. This result was impacted by the $625 million convertible note we issued in August the repurchase of $1.19 billion of our 2026 notes and the interest expense from the $380 million drawn on the Term Loan A component of our existing credit facility. The net impact on non-GAAP net income per share of these balance sheet activities was a reduction of $0.05 in Q3. And excluding these charges, non-GAAP diluted net income per share would have been $0.59, GAAP diluted net income per share was $1.51, a 358% increase year-over-year. This increase is primarily driven by the onetime reversal of our tax valuation allowance and gain on debt extinguishment which is slightly offset by the impact of our new debt structure. Q3 adjusted free cash flow was $85 million or 37% of revenue, which is up significantly from the prior year's $19 million or 10% of revenue and on a trailing 12-month basis was 21% of revenue.

This increase was driven in part by the equipment financing in Q3. During the quarter, we entered into an equipment financing arrangement with a third-party financial institution for $28 million to better align our investments with the future revenue that they will generate. Absent the leasing of this equipment, our adjusted free cash flow would have been 25% of revenue in Q3. Turning to the balance sheet. We strengthened our balance sheet by repurchasing approximately 80% of our 2026 convertible note for a combination of the issuance of a new $625 million 2030 convertible note offering, the $380 million drawdown on the term 1 component of our existing credit facility and approximately $230 million of cash. Following these actions, our cash and cash equivalents balance remained healthy at $237 million. and the combination of cash, remaining term loan A capacity and projected cash flow generation is collectively more than the remaining balance of our outstanding 2026 convertible notes.

We repurchased $2.9 million of shares in Q3, buying back approximately 101,000 shares, bringing our cumulative share repurchases since IPO to $1.6 billion and 34.9 million shares through September 30, 2021. These Q3 repurchases completed our 2024 buyback program, and we will operate our repurchase program through July 31, 2027, and under the new $100 million authorization we announced during the quarter. During the quarter, we repurchased a portion of our 0% coupon 2026 convertible notes in part with an interest-bearing term loan A that is initially at SOFR plus 175 basis points or roughly 6.1%. As a result, we now project to have modest interest expense in the near to medium term, where interest expense was previously immaterial. Given this, we have added a new disclosure metric for unlevered adjusted free cash flow, which we will provide in addition to the current adjusted free cash flow metric, which is a levered adjusted free cash.

We believe that unlevered adjusted free cash flow is an important metric as it provides a clear view of our cash generation before the impact of financing decisions, and many investors and analysts use this unlevered adjusted free cash flow as the basis for their enterprise value calculations. Our Q3 unlevered adjusted free cash flow was $85 million or 37% of revenue. The strong demand we've seen across our unified Agente Cloud and the traction we are seeing with our higher spending AI and digital native enterprises. Coupled with the increased visibility we have from having signed multiple 8-figure committed contracts after Q3 close, gives us the confidence to raise our outlook on both revenue and adjusted free cash flow margin for both 2025 and 20.6%. For the fourth quarter of 2025, we expect revenue to be in the range of $237 million to $238 million, which is approximately 16% year-over-year growth.

For the full year 2025, we project revenue of $896 million to $897 million, representing approximately 15% year-over-year growth, an incremental 100 basis points higher than our prior guidance. For the fourth quarter of 2025, we expect our adjusted EBITDA margins to be in the range of 38.5% to 39.5% with an adjusted EBITDA margin of approximately 41% for the full year. For the fourth quarter of 2025, we expect non-GAAP diluted earnings per share to be $0.35 to $0.40 based on approximately $111 million to $112 million in weighted average fully diluted shares outstanding. For the full year 2025, we expect non-GAAP diluted earnings per share to be $2 to $2.05 based on approximately $106 million to $107 million in weighted average fully diluted shares outstanding. The Q4 and full year non-GAAP diluted earnings per share guidance includes the projected impact of a range of about $0.05 to $0.10 reduction in Q4 and and about $0.15 to $0.20 reduction for the full year from the net impact of our Q3 refinancing actions.

We project a full year adjusted free cash flow margin of 18% to 19%. Looking further ahead, I would also like to provide a brief update on our 2026 outlook. While we will provide more fulsome details on 2026 expectations during our earnings call in February, we have already begun to put the foundations in place to further accelerate growth. Given our momentum and the increased visibility into demand on the back of several recent customer wins, we have committed investment in additional data centers and GPU capacity that will come online over the course of 2026 that will accelerate growth ahead of our previously communicated time line. We have signed leases for approximately 30 megawatts of incremental data center capacity across several new data centers that will commence over the course of 2026. These new data centers our corresponding investments in incremental GPU capacity will enable us to comfortably deliver 18% to 20% growth in 2026, achieving our 2027 revenue growth targets a full year earlier than we had projected.

And while our COGS and operating expenses will increase in early 2026, as we ramp into our new data center capacity, we anticipate delivering high 30s to 40% adjusted EBITDA margins. while maintaining mid- to high teens adjusted free cash flow. We also remain committed to maintaining a healthy balance sheet, and we anticipate that our net leverage will end 2026 in the mid-3s range. including the impact on net debt from any incremental lease-up. We look forward to sharing more on the traction we are getting with our unified agenetic cloud, the growth we are seeing from our highest spending customers the investments we are making to further accelerate our growth and our outlook for 2026 and beyond when we get together again in February. That concludes our prepared remarks, and we will now open the call to Q&A.

[Operator Instructions] Your first question comes from the line of Gabriela Borges with Goldman Sachs.

Congratulations on a really exciting 20 preliminary forecast. How do you at, I want to ask you about the multiple 8-figure committed contracts that you're talking about. Tell us a little bit about this cohort of customers. To what extent does it overlap with some of the AI revenue that you're talking about? I know you've been working with the private equity community as well, you've talked about migration. So maybe just a little bit about the type of customer that's signing the fit contract and the extent to which I know in the past, the AI cohort has been much less sleepy and its ability to ramp up and down. And so I'm trying to understand the infection between those 2 cohorts.

Yes. Thank you, Gabriela. Great question. So the 8-figure commitment contracts that we just talked about come in different forms. Primarily, these are AI native companies that are looking to take advantage of our infrastructure as well as the other customer that I was just talking about for our AI platform, looking to build a series of agentich experiences for software engineering, taking advantage of our Gradient AI platform layer. So it's a combination of all of these. And as I was explaining in my prepared remarks, it is increasingly getting difficult to just separate out where AI starts and stops and where for cloud begins because most of the customers that are starting their experience with digital lotion from the AI side, are increasingly using our various storage artifacts like network file system or the the VPC capabilities or a number of the networking capabilities and things like that. So the crossover is becoming more and more between the AI side of our platform and cloud.

So that's why we are now starting to see a more unified cloud platform from us, which we are calling as the agent cloud. So A lot of these commitments and contracts that we are starting to take on now typically start with AI, but also spill over to our AI cloud side as well. So what is exciting for us is that some of these customers start their journey with DO using a fairly small proof-of-concept type of footprint. And now they're starting to scale. And this is also one of the many reasons why we are expanding our data center footprint so that we can keep scaling with these customers. And the other attribute I want to call out here is that these AI workloads are predominantly influencing, if not all, on the inferencing side. So it is durable, it is predictable, and we also have a great opportunity to keep scaling with these customers as they find real-world fraction and scale globally. So that's why it's really important for us to start looking at our capacity as we place bets on some of these real marquee AI native companies that are finding traction with real end customers, both on the consumer side as well as on the enterprise side.

Your next question comes from the line of Radi Sultan with UBS.

We good to see the platform traction coming in ahead of schedule. I guess for me, just AWS and Azure both had some pretty high profile outages recently. I'm just curious, like is that having any near-term impact or catalyzing more migrations from the hyperscalers that driving more traction for Partner Network Connect or some of your other multi-cloud offerings. And then just curious, how many of those 8-figure deals were migrations from the hyperscalers.

Yes. Thank you, Radi, for the question. So we've been seeing a steady increase in migration workloads since we made it into an explicit go-to-market motion. And as you know, migration of sophisticated workloads is always a combination of factors, right? It's seldom we see a disruption from a cloud provider, and we're just going to move a fairly sophisticated global workload. But it is a combination of factors. Some are driven by dissatisfaction with an incumbent. But mostly, it is driven by something they find attractive in a new cloud provider like digital ocean. So as we have started building out our cloud capabilities, especially the ones that I described in Slide 12 of our deck like more advanced networking, various flavors of our droplet configuration or storage, like cold storage is a really, really important capability that many of our large customers with sophisticated workloads have been asking us for the auto scaling of our DBA, I mean these are all very fundamentally building block type of capabilities for attracting more migration workloads not to mention some of the stuff we did in the last couple of quarters, like virtual private cloud and Direct Connect and things like that.

So even though a single incident doesn't necessarily precipitate major shifts in workloads, these are all paper cuts and us having these other digital native enterprise-ready capabilities just makes us all the more attractive to incoming migration. And the AI native workloads, typically are new workloads that are starting on our platform. But many of the workloads that we are seeing that I talked about during my prepared remarks on the cloud or migration coming from various other hyperscaler clouds.

Your next question comes the line of Josh Baer with Morgan Stanley.

Great. And congrats on the acceleration. I wanted to follow up on the 8-figure contracts signed after quarter close. I guess I'm wondering do you have capacity to serve some of those in the coming weeks? Or is that all really what the 30 megawatts of new data center capacity is geared toward and then hoping to get a little sense of like the ramp in that capacity basically, any context for the go-live times for these big contracts and like how to think about the shape of 2026.

Yes. So I'll get started, Matt, and then you can chime in. So from the ramp-up perspective, some of these customers are already doing business with us. And as we think about the capacity, there's some capacity we are bringing online in our existing data centers. And then, of course, a lot of the the visibility that we now have with these customers and their inference scale-up is the reason why we have taken up expanded data center capacity. And these data centers will come online progressively through 2026, right? So we do have a build schedule from these providers. And we work very, very closely with them to make sure that we get the warm shell and then we move in and we start racking our servers and there's a lot of moving parts in terms of bringing this capacity online. But we don't have to wait for this new capacity to start lighting up these workloads. As I said, we do have some capacity in our existing data centers. So it is a combination of these things. And you're absolutely right that this visibility into the inference adoption of our AI native customers is the reason why we are expanding our data center footprint.

Yes. I'll just add that most of the capacity is going to come online and call it the first half of next year. In fact, you'll see in our fourth quarter financials, and this is included in the guidance for the adjusted free cash flow we have for 2025, that we're going to be paying some of the NRCs for some of these build-outs in the fourth quarter. And so we expect that the capacity will become online in the months and quarters following that. So it will be an early ramp of the data center capacity. But then clearly, you have incremental time to deploy the GPUs and for the customers to wrap. So we expect the revenue ramp to be relatively smooth over the course of the year, but we'll be bringing on a fair bit of capacity in the first half.

Your next question comes from the line of Kingsley Crane with Canaccord Genuity.

I want to echo I can add some sure it's gratifying for the team -- with a larger peer New cloud peer has acquired a handful of past capabilities over the past 6 months, including more recently Python notebook. I think reinforcement learning for agents -- what's your take on that just to give credence to your strategy? And how do you see competition evolving as you continue to cater towards customers at market?

Yes. Thanks, . Thank you for the question. So we are approach when we laid out our strategy in April, we start our strategy with a deep understanding of who our customers are and what it would take for us to serve them well. And of course, that understanding has deepened over the last 6 months as we have started working very, very closely with these customers. In terms of the Python notebook, as this has been a capability that we have had for a couple of years now. in some of the storage enhancements that we are seeing in the market is also something that has been long a part of our very rich and deep software stack. So if you take a step back and think about our strategy, our strategy is now from an AI perspective, targeting AI native companies that are building real businesses and running models in an inferencing mode and these are real-world applications that require not just GPU and inferencing capabilities, but they need agentic workflow capabilities.

They need storage, databases, authentication authorization, they need orchestration from a Kubernetes perspective. So essentially, they need a unified agentic cloud stack, which is what we provide. So we have been executing on our strategy. And if you look at Slide 7, you see the richness of the stack that we have built all the way from infrastructure on both cloud and AI to middleware with Platform as a Service or the Agentic development life cycle. And Slide 12 shows how much we have enhanced that since just the last earnings call. So we've been super busy and every orange box you see on Slide 12 has been a result of feedback that we are getting from customers real time. Some of these features have been laid up in just a matter of days. And that is the power of really coinventing some of these pieces working hand-in-hand with our customers. And I feel building on our strength, which is software differentiation, and leveraging the strength of our 12-year-old full stack general-purpose cloud really puts us in a very favorable position when it comes to being attractive to these scaling AI data companies.

Is hardware a part of it?

Absolutely. But we think as companies become more and more sophisticated and they start serving real-world enterprise needs. The center of gravity is going to shift from hardware and networking and move more and more towards the software stack as we have seen in every wave that we have encountered over the last 2 or 3 decades. So we feel really good about where we are and we'll be aggressive in adding new functionality into our platform as we start seeing opportunities for those -- from our customers and in the market.

Your next question comes from the line of Patrick Walravens with Citizens.

This is [ Nick ] on for Pat. Paddy, one for you. You kind of answered this already, but are there any other factors that you guys take into account when deciding what to build next? Like you mentioned that it's primarily customer driven, but are there like competitive positioning or how do you balance the idea of what a customer wants with competitive positioning and long-term revenue opportunity, especially in something that could be seen as more experimental.

Yes. Nick, thank you for the question. So just for the avoidance of doubt, we are a competitor of air, but customer obsessed. So we -- and that strategy has really worked for us, especially in the fast-evolving AI landscape, where we have been very, very disciplined in not chasing the bright shiny training workloads and trying to be somebody that we are not. But we've been patient and now we see an opportunity to decisively move to take a full position in the world of inferencing and offer a software platform that combines the the raw power of having the best of breed flexible AI infrastructure and combine it with where the puck is going, which is there's a whole generation, like 20 years of app developed applications that have been developed over the last 20 years will need to be replaced and modernized with agents and agent workflows so managing that whole life cycle is starting to already create a tremendous amount of problems for companies to build, operate and manage.

So we think there's a phenomenal opportunity for us to be one of the first movers into the world of agent development life cycle, and that is exactly what we are focused on. So while it is important to be aware of what our competition is doing, especially the Neo cloud, I feel very confident that we have unmatched software expertise and depth of our platform, as you can see from Slide 7 and 12. So if we keep doing what we are doing, our strategy is resonating with the AI native. And these are customers that are doubling and tripling their footprint on a month-by-month basis. So if we can keep pace with them and keep shipping at their speed, I think everything else is going to take care of itself.

Your next question comes from the line of James Fish with Piper Sandler.

Nice quarter. Just on the 2026 starting point, kudos on bringing that forward, understanding what's driving the confidence and visibility at this point, but how should we think about the parts underneath between core managed service and AI and do we still expect that AI business to continue to double given you've done it for 5 straight quarters now?

Thanks, Fish. We think that if you look at the success that we're having, it's coming from a combination of things. It's coming from our success with our largest customers regardless of whether they're AI or core cloud growing very, very rapidly with the $1 million-plus customers growing 72%. We think that continues. And a lot of the migration workloads that we've been working on and some of these longer-term committed contracts are also on the core cloud side. So growth of our biggest customers is kind of the lever #1. Lever number two, as you said, is AI growth. It has been doubling every quarter since we launched it, and we're expecting that to continue. So we're going to get a big chunk of growth from AI that will become a more material part of our business. It will get into the mid-teens, maybe even high teens as as a percent of revenue.

And then we're still generating very good incremental revenue from our product-led growth engine. Our customers in M1 to M12 are still at at very high levels relative to historical. And it's really those 3 levers that give us confidence. And as you said, we have better visibility now than we've ever had. You said how many how many 8-figure committed contracts have we ever had in the company. It's I don't know how many, it's not that many. It's maybe 1 or 2. And now we've got multiple that we've signed just in the last whole month or so. So we're very confident in the 18% to 20% revenue for 2026 and are excited to be able to pull that in a whole year.

Your next question comes from the line of Mike Cikos with Needham & Company.

This is Matt Calitri on for Mike Cikos and great to see the strength of large deals. I know AI is not included in net dollar retention. But are you thinking about starting to include it as it becomes a larger part of the business and more predictable? And what other puts and takes are you considering as you look to drive NDR back over 100%?

Yes, it's a great question. We are looking very hard at how to incorporate the resilient growth of inferencing into our metrics is 1 metric, and it captures it used lit in a SaaS environment that captures the ongoing growth of a customer to date, we haven't included it in MDR because a lot of the early traction that we were getting, and I think in a lot of folks we're getting in the industry was more project-based and more experimental as we're seeing customers like some of the ones that Patti mentioned like fall and others, they're bringing scaled workloads to us where they have more predictability in their demand and growth we believe it's appropriate to start to figure out how to include that.

We'll likely revisit this once we get into the beginning of next year, and we have a better sense for the 2016 outlook I mean we're providing more specific guidance. But what I'd say the key takeaway is the AI revenue that we're seeing now, that we're getting committed contracts for is behaving more like the traditional cloud where customers come in with scale production workloads and then they have more visibility into the growth of that workload over time. Hence, a metric like NDR becomes more relevant. And we're confident that in doing that, that would be additive to our communications of the resilience of the growth that we're seeing.

Your next question comes from the line of Mark Zhang with Citigroup.

Maybe just on NDRs, obviously still at the 99%, but we're seeing good expansion momentum and portfolio momentum. So can you maybe just walk through some of the key puts and takes there? And traditionally, I think expansion activity has been nested in that metric. So can you maybe speak to some of the behavior there? And any discernible changes in customer behavior versus last quarter, they were 12 months ago, whether it's on pace up expansion or how they're expanding.

Thanks, Mark. It's a great question. That expansion is definitely the driver of the growth. If you look at the big customers the customers that spend $100,000 or more in ARR as we showed, it gets better even as you get bigger spenders. They're driving a lot of our growth. And as you would expect, the NDR is better. The thing that people lose sight of, I think, when they think about the Digital Ocean business, they forget that we have 640,000 plus customers and 450 or so or more thousand of those are effectively a paid premium. They're small customers, they spend $10, $15 a month, and many of them stay on for a long period of time. The average age of that cohort is something like 4.5 years.

But you also have a lot of customers come and go. They experiment. I mean so the MDR of that paid premium segment of our customers is below $10 -- and so that weighs down the overall NDR of the company, and it masks the fact that with our largest customers, we're actually seeing very, very strong growth, driven by increased expansion. And that's another follow-up to the question prior to this. That's another metric that we're thinking about is because we're blending this -- where the real growth engine is for the company, we're blending that MDR with the NDR of a giant cohort, which is fantastic for us because it gives us access to a lot of developers. But it's just by its nature, it's going to have like slightly below 100 DR we think that's masking a lot of the underlying performance that we're seeing.

Your next question comes from the line of Wamsi Mohan with Bank of America.

Nice results here. Just given the strong growth trajectory and the confidence, can you just talk about where you think the net of both sort of explicit CapEx plus your equipment leasing what the sum total of that could be as you look over the next couple of years in dollar terms? And how large could you see that delta growing between adjusted free cash flow and your adjusted unlevered free cash flow margins over the next few years?

Wamsi, it's a great question. I think as we said, we're trying to give preliminary guidance for 26 to give the directional kind of growth rates. And we feel very confident that we can deliver that 18% to 20% revenue growth while still delivering the kind of the mid- to high teens in adjusted free cash flow. And when I say that number, that's the levered number that I'm referring to. And I think at this point, the market is evolving so fast, but it's hard to say what the CapEx would be or what the impact on adjusted free cash flow margin would be even in the second half of next year, much less or beyond.

What I can tell you is we -- and you've seen us in terms of our behavior, we didn't chase the training opportunity. We didn't pursue what we viewed as revenue that we weren't sure how durable that was going to be for us, and we didn't know if we had a competitive differentiation there. But what we said is where we see opportunities to deliver durable revenue growth with a differentiated product that has good returns that we'll make investments to pursue that. And you've seen that with our willingness to take down additional data center capacity and secure new GPUs. So I'd say what you can expect is continued disciplined behavior. -- where we're trying to drive durable revenue growth while maintaining attractive free cash flow margin.

Your final question comes from the line of [ Robert Galvin with Skip ].

I want to ask about the transition to leveraging leasing and how the gross margins of data centers and equipment you own and operate compared to gross margins from leased capacity?

Great. So just to clarify there, we don't -- we lease all of our data centers. We're a co-location tenant in each of our data centers. We don't own any. So the taking down of additional data center capacity that we've just referred to will behave the same way that it did when we took down the Atlanta data center earlier this year and when we took down the Sydney data center, several years ago. And what that does is our cost if you think about our cost of goods are variable with revenue over the long term. But in the very short term, they're somewhat lumpy. You take down an incremental data center, you have -- not only do you have incremental upfront costs and NRC that happens before you're generating revenue.

But the day you turn on the data center, you start paying for the space and some of the power, then you build it out and kit it out with gear and you fill it up and it takes some period of time to generate the fully utilized revenue in that facility. So there's always a lump of higher expenses in the beginning when you turn on data center capacity and you grow into it. And growing to it over a series of even just a couple of quarters and your gross margin gets back to what's more of a steady-state gross margin. So we expect that to happen in the beginning of of next year of 2026. But we factored that in and incorporated that into our guidance for the year in terms of the free cash flow margins that we expect to generate. So it's just normal course for us. It's nothing different than what we had done previously with respect to the data.

Ladies and gentlemen, that concludes today's call. Thank you all for joining. You may now disconnect.

All right. Good morning. Thank you so much for joining us at the DigitalOcean Session Conference, and special thanks to Paddy. Thank you for joining us.

Well, thank you for having us.

Can we get the door closed, please? If someone please. Thank you. So Paddy, I wanted to ask you about a topic that's been very much in the news [indiscernible] post Oracle's earnings report. I think there's a fair amount of industry discussion on whether investing in inference and training, training specifically is actually a good.

All right. Let's try that again. So Paddy, I think the news flow this week has been very much focused on the unit economics of training and inference. Now the beauty of the DigitalOcean business is you do a little bit of both and increasingly inference. So how do you think about as a CEO for DigitalOcean, how do you think about whether investing in training and inference is a durable, healthy business for the long-term, and maybe take training and inference separately?

Yes. So first of all, thank you for having us here. It's wonderful to be here as always. It's a great leadoff question in the sense that this has been a dominant theme for us over the last several quarters, but also this week where I've spent a lot of time with AI native companies that are ramping up their footprint on us.

And I would say we made a bet a couple of quarters ago that we're going to be focused on inferencing for a number of strategic reasons. Number one, that's very close to the DNA that we have had over the years. And a couple of really interesting data points from a unit economics point of view.

For training, it is all about GPU dollars per hour. But for inferencing, there are a lot of interesting patterns that are emerging where if you think about inferencing, it's all about the throughput that you can get, like the flops measured, right? And within reason, like within a family of GPUs, increasingly, our customers really don't care about whether we are servicing it with -- like I'll take an older example of H100s versus H200s, like they're like, okay, as long as I can get this kind of throughput, I really don't care how you service us. And so it's all about the dollar per flops versus GPU dollars per hour, so which is a pretty big shift.

And interestingly, there are a couple of other types of inference use cases that are emerging where I spent time with 2 startups -- 2 different start-ups this week, both of whom have freemium business models. So for them, the free tier, they're okay or they want us to serve the free tier customers using an open source model that we offer through our serverless inferencing fleet.

And for their more "premium customers," they want to have a closed source model that they have fine-tuned and they're hosting on raw GPUs on our infrastructure. And they want us to do the load balancing, and the routing dynamically based on whose request is coming in. So if you look at the price performance on these 2 types of requests that we are getting, very different, right? One is about 1/4 or 1/5 the cost profile of the other.

So from an inferencing point of view, it's a totally different ball game in terms of how our customers perceive the unit economics. And that's why it is really important for them to not only have a provider that is just racking and stacking GPUs, but have a full stack agentic cloud that can do all of these things dynamically.

And level set us on the mix that you have today, are there exceptions to the rule way, will accept to train workloads as the part of [indiscernible] on the customer journey?

It's getting smaller and smaller part of our fleet. And part of our resource allocation philosophy, and that's one of the big things we're looking at for next year's capacity, how should we allocate it across our customer base. It's going to be predominantly inference-based workloads for a number of reasons.

One, for me personally, as a CEO, I look at who's actually paying the bill, not the start-up, but who's eventually paying the bill? Is it a venture capitalist? Or is it actual real customers? And I get super excited, obviously, by -- like whether it is a consumer or an enterprise customer that is paying the bill because then there's more durability of revenue, both for the company that we're doing business with and for us eventually.

This is a great observation. So out of the companies that you have on DigitalOcean today, I know your visibility is not perfect. What percentage of them have customers paying the bill versus these?

So increasingly, a lot more. So we have companies that are doing, for example, generative media. And they're selling to B2B customers who may have consumers at the other end of this spectrum, but these B2B customers are looking at generative media as a way to increase their conversions or increase their engagement and the depth of product usage for them. So these are great use cases for us because we know that this use case, if the customer is -- or if the start-up is able to prove the workload, it is only going to go up in usage.

The other interesting comment you made there, inference is closer to DigitalOcean's DNA. I think you expanded on that a little bit already. But what do you mean by that?

Yes. So there are multiple reasons why we feel inferencing is a place where we have a big right to win. For a number of reasons. I'll start with the most obvious one, which is inferencing is a lot more than just GPUs. Yes, GPUs are a big part of inferencing. But when you talk about inferencing, you need the raw horsepower to leverage an LLM, whether it is a closed source model or an open source model or a serverless inferencing of either of these 2 models.

But more important than that is in an inferencing mode, you need to pump in some custom data. You need to process -- preprocess the data and post-process the data. You need a way by which you can build some of the other higher order services around it, like you need to have guardrails. You need to evaluate both in real time and also offline, which model is the best suited to serve the needs of your customers.

I gave the example of a free customer and a freemium customer, but also different types of use cases might require different parts of the same LLM even, right? We are now starting to see from a same LLM provider, different models are great at doing different things. So how do you do that in real time?

For many companies which are not super sophisticated AI native, they also want the ability to start building agentic workflows from a template. They also want the ability to have multi-agent orchestration. So you need -- you need a way to have sophisticated routing agents and traceability and observability. So all of the things that we have done on the traditional cloud are very important when you're running inferencing at scale.

So for all those reasons, a lot of our customers have started -- they come for the inferencing needs, but they stay because we are a full stack cloud because they start leveraging the other stuff. They don't have to go to multiple cloud providers. And the fact that in our new data center, these 2 stacks live side-by-side in an integrated fashion is a big, big deal for them.

Turning to the broader business. You've had a lot of success growing your Scalers+, your largest end cohort, now 25% of the portfolio and growing 35%. What are the specific product features and enhancements you've invested behind to accelerate this momentum? And where do you see the gaps that you need to fill to continue?

I'll start with the boring answer. There's not 1 or 2 features. It's a collection literally of about 250 features we have released over the last 4 quarters or so, right? And if you look at the number of business days in any given quarter, we are releasing a major product update almost every business day.

And we can categorize some of these things to say -- and for those of you who are new to the DigitalOcean story, when I came on board about 20 months ago, one of the biggest themes that was highlighted was the fact that customers grow to a certain size of footprint on the DigitalOcean platform, and they're forced to graduate because we didn't have certain types of functionalities, right? So you can group those functionalities into core compute.

We had a couple of types of "droplets" but we didn't have a lot of different flavors of them in terms of some use cases need memory-optimized droplets. Some use cases need compute-heavy droplets or storage-heavy droplets. We have fixed a lot of those things. We have a variety of different droplet options, even to the extent now we have an inference optimized droplet powered by GPUs. So that's on the fundamental level.

Our storage was also lacking a lot of high-throughput storage, input output, different types of network attach storage and things like that. So that was another big hole. Our networking stack was fairly basic, and that was a big area of focus for us over the last 6 months. We have added several features, including virtual private cloud and also direct connect between our data centers and the hyperscaler data centers.

And that's been a big hit because we don't charge extra for that, but that has been a big hit with our large customers, especially Scalers+ because now we can go make a pitch to migrate a part of their existing workload, like it doesn't have to be all or nothing. And our big customers absolutely love it because they love many parts of the DO platform, but not all parts of it. They want to have a multi-cloud strategy even within a given workload. Now they can tastefully pair us with an existing workload running on GCP or AWS. So that's another one.

And final thing is it's an evergreen area where we are investing a lot of bandwidth, which is our database offering. And I think that's going to continue over the next year or so. In a couple of weeks, we have our product conference in London. We're going to make a series of big announcements, and it will be in compute, storage, networking, database and everything in between.

And you've traditionally relied on a PLG motion. As you're expanding to these larger customers, how are you leveraging partnerships like with Hugging Face or channel partnerships to grab a larger portion of the market there?

Yes, it's a good question. So our product-led growth has been a major driving force in -- from the founding days to now, right? But over the last couple of years, it is -- it was showing signs of fatigue, if you will. But last quarter, we talked about how we had one of the best quarters ever in terms of our month 1 to month 12 cohort.

The reason why I get super excited by that is today's M1 to M12 is tomorrow's NDR. And I obsessively look at the quality of customers coming in, in terms of their ARPU and their velocity of progression through the M1 to M12 and how quickly they're attaching other DigitalOcean products.

So on top of that, we are now -- we have never had a proper sales-led growth motion. When we had good sales, we didn't have great products. And when we had good products, we didn't have good sales. But now we have a good 1-2 punch, and we are now packing a good amount of wood behind the sales-led growth arrow. So that will be a big theme for us next year.

On top of it, we are starting to open new front doors through which customers can come in. One obvious one is our AI front door has been great in getting more customers attach our traditional DO stuff.

But you also mentioned a couple of other partnerships. One partnership that hasn't gotten a lot of publicity is because we haven't really launched it yet, but we talked about it at the conference, a company called Laravel, which is the most popular PHP framework in the world. Their founder talked about how they're launching their BPS offering exclusively on DigitalOcean.

We have, I don't know, how many thousand people in the waitlist for that. We are going to be releasing that in the next couple of weeks. So we expect that to be a massive front door. So our partnership is not restricted to 1 or 2 companies, but we are looking at the open source community at large to drive a lot of traction to us. So that will be an evergreen motion in terms of our partnership, both on the core cloud as well as on the AI side.

And on the AI side, can you talk to us a bit about the breakdown of the platform currently? What percent of your customers are leveraging the infrastructure versus the platform and eventually the agentic? Where do you see this going over the medium term?

Yes, absolutely. And I'll maybe just do a continuation of the previous answer, which is on the AI side, the AMD Developer Cloud is another big front door, which is powered by DigitalOcean. So we continue to expand the footprint of how companies and developers can enter the DigitalOcean family.

Specifically talking about the AI stack, I think we have talked about -- and we largely borrowed the inspiration from the Goldman Sachs framework of IPA, infrastructure platform and agents or applications. And from our infrastructure point of view, it is both NVIDIA and AMD are the big GPU offerings.

But on top of these GPU offerings, we have a couple of layers of abstraction. Of course, we offer bare metal compute, but increasingly, a large percentage of our customers have started using our droplet architecture. And our droplets are very sophisticated in what they provide in terms of taking away all the brain damaging work of putting the right frameworks and making sure that they are working and stuff like that, but also more sophisticated orchestration and life cycle management of these instances, which if you worked with GPUs, it's not very sophisticated, right? Then there's a lot of breakage and you have to do a lot of baby sitting in terms of the life cycle management.

In addition to all of this in the infrastructure layer, we are also building some inference optimization logic, both ourselves as well as working with some partners. We are building inference optimization, including the GPU inference droplet that we created. So that is an ongoing R&D effort. So we are building a lot of IP on that.

The next layer is our Gradient AI platform, which is -- starts with serverless inferencing of both closed source as well as open source models and also all of the other building blocks that I rattled off all the way from a model playground to TCO calculations on different LLM throughputs to agentic building blocks of multi-agent workflows or agent evaluation, agent traceability and so forth. So all of these building blocks is what we call as Gradient AI platform.

So typically, the users of these 2 layers are very different. AI native start-ups typically want access to GPUs. SaaS applications and traditional software companies want access to serverless endpoints or the agentic framework directly because they're not trying to take GPUs and start building from scratch, but they are introducing AI as a feature into their platform.

So to finish my answer, most of the revenue today comes from AI infrastructure layer, but most of the mind share adoption and thought leadership is in the middle layer. Will that invert sometime in the future? Yes. When? I don't know. But we already have 6,000 unique customers using our platform, more than 15,000 agents deployed at this point. But a lot of them are in proof-of-concept mode. But at some point, over the next few quarters, that will invert, and we are really looking forward to that.

And you announced the general availability of your Cloudways Copilot. Any early feedback from customers, what you're hearing on adoption and...?

Yes, it's been a big hit. It's been a big hit. It's just amazing. See, our -- you have to realize our Cloudways customers are -- some of them are technical, but generally speaking, are not very technical, right? They're hosting websites or their digital agencies and things like that. And they typically have shared IT resources. They're not babysitting websites day in and day out.

So for them, having an agent that is doing the job of a human is a welcome addition to their fleet because they're not there looking at the varnish cache of their WordPress deployment all day long. So the more automation we can provide in terms of just the observability and monitoring of the health of their website is super welcome.

But the next step of actually taking remediation on stuff before it actually goes wrong is an absolute winner for them. And we are getting more than 95% accuracy in terms of our ability to predict that something is about to happen. And in addition to the cloud-based Copilot, we're using the same technology internally because I mean, obviously, we have a massive cloud footprint and with any massive infrastructure at this scale, things go wrong all the time.

And it has reduced our time to respond and mean time to respond and mean time to remediate by like 30% to 40% in most cases. And we've still only opened up 3 or 4 use cases internally and the productivity gains are just staggering.

So you made a really interesting comment there. AI native on GPU access, traditional SaaS companies on servers and more edge compute. The question for you is there's a debate in the market today on whether AI native companies can disrupt traditional SaaS [indiscernible]. And one of the parts of the argument as well, the tech stack is fundamentally different for an AI native company [indiscernible]. Would you agree with that? Or is there a valuability here?

I think I'm more on the side of over time, the AI natives are going to disrupt the traditional software companies, and there is a parallel stack emerging. And even in the AI native companies, and I should probably qualify this a little bit more.

AI native companies that are more infrastructure-oriented need raw access to GPUs. But I was talking to an AI native company that is building contact center software. They don't want access to raw GPUs. They want serverless endpoints, because they're like -- yes, but they want cheaper serverless endpoints, but they don't need access to raw GPUs.

What they need is our high-quality tokens and tokens out because they're doing voice to text and things like that. They want us to do the heavy lifting of, hey, I'll give you the model or I'll point you to the model, you host it, you manage the life cycle and just give me API access to it.

So I feel there is a parallel like observability, for example, right? Age-old problem. We've been doing it since the mainframe days. But the way you do observability for a pure end-to-end agentic stack is very different. Like what you observe and what you take remediation on is very different from what you observe and take action on for a traditional cloud stack.

So I think there is a parallel stack emerging. And it is also nuanced in the sense that the more sophisticated AI natives that are building raw infrastructure or doing media manipulation and those kinds of things need access to raw GPUs, but AI natives that are more in the business realm or building business workflow software are tilting more towards getting access to endpoints in a serverless manner.

Yes, fascinating. And you talked about this at the beginning how more of the mix is now customers that are actually paying for the business model [indiscernible]. This is a question about the quality of the revenue of some of the AI in the startup. To what extent are you still seeing a lot of stopping and starting a lot of experimentation such that it's not -- you can't measure NRR the way you typically measure, or are you starting to see the quality of that revenue is not [indiscernible]?

Yes. So I would say maybe about half of our revenue is very -- it's getting to be predictable because of this inference workloads.

Of the AI native companies?

Of the AI -- yes. So the AI native companies, and we are trying to get slightly longer-term commitments from these customers because we have limited capacity in terms of our fleet -- inference fleet. And part of my business development activity with these companies is, hey, give us more visibility. Can you give us 6 months? Can you give us 12 months or 18 months?

And the more mature these AI native companies are with their inference workloads; they are willing to now give us visibility into their 12-month run rate because they know what the price performance or number of tokens required are to serve their customers. And they have -- they have a prediction in terms of how that is going to look like in terms of their end user adoption and the number of tokens required to service and are able to give us some visibility in terms of what their needs are. So I think it's still early days, but we are starting to see that for sure.

And I know we debate this with Matt every quarter. The question on the demand environment, the visibility that you have in the company because you do tend to service more of an SMB developer around the customer. Would love to hear your thoughts, how do you feel about the health of the demand environment for both that -- look, you [indiscernible] what about the cloud?

I think they have been more resilient than I initially thought during the very turbulent April time frame. It has been fairly resilient. And I would also say that it's less about some of the macroeconomics going on globally, but there's a lot of microeconomics from a country-by-country perspective that we see. But in terms of the demand environment, we are not seeing anything unusual at this point.

How about from a competition standpoint? I really appreciate that at the beginning of this conversation, you were talking about the inference workloads being close to your DNA as a company and the unique value proposition that you have. As investors, we spend a lot of time hearing from new cloud providers that are addressing inference workload. So maybe drive that for us, are you seeing a change in competition for that AI native cohort versus the cloud [indiscernible]?

I think we're -- I don't know if it has really changed that much in the last 6 months. It's the same names that we keep seeing. So it hasn't really changed, but these are the same NEO clouds that you're probably hearing from as well. But I think there is definitely a more nuanced appreciation from our customers in terms of some of the other building blocks that they need, whether they -- we are seeing a lot of companies approach us.

So there's -- the concept of multi-cloud inferencing is also picking up. So it is not -- so we have many customers for whom we are not the only cloud. So they are -- they may start their journey from a hyperscaler for whatever reason, they don't have capacity or they don't have certain things available and they come to different NEO cloud, they come to us. So I think the -- the concept of going from a single cloud to multi-cloud, probably in classic cloud took 10 years for multi-cloud to really come to fruition. But in AI, we -- it feels like it's already there from -- for inferencing.

Turning to the balance sheet. You've historically invested around 20% of revenue in CapEx and 15% of that towards growth and 5% towards maintenance. How are you thinking in the medium term as you look to your Investor Day targets to reaccelerate revenue? Is there any near-term influx that's needed on the growth side of things to support this investment? And how are you thinking about additional funding tools to do so?

Yes. It's a really pertinent question given what we're seeing. So in the -- on the Investor Day, we said this is -- this has been our historical run rate, and this is the split that we are used to. And I think that's largely still true. But we're getting more and more confidence that we will not be afraid to invest behind durable growth backed by companies that are seeing real customer traction.

Now we also talked about the fact that just like other companies in the market, even some that came out earlier this week, there are multiple tools that we will leverage as part of our tool belt to support our growth aspirations. And the key thing here is to say it needs to support our growth aspirations. If there is a way to accelerate our growth or get to our growth aspirations faster, or any combination of those 2 things, we will absolutely invest behind it, and we are starting to look at some of those things as we are picking up momentum and traction with these AI native companies.

And not to sound like a broken record, but the more true inferencing workloads that we can see with durability attached to them, the more conviction we will have to invest behind those workloads and those companies. And part of our mandate is to go allocate our resources from a compute perspective behind these companies that have real enterprise and consumer use cases behind them.

So just by the nature of inferencing, it just takes out a huge piece of uncertainty behind like are they going to be viable in 6 months? Because if they're doing inferencing at scale with thousands of GPUs today, somebody is paying money in exchange for value. And that is a big validation for us, and that gives us more conviction to invest behind this.

What does the pipeline look like for the $20 million plus multiyear type deals like the one that you announced recently?

The pipeline looks healthy. Pipeline looks healthy. Part of it is with companies that we seeded and are finding traction. We are getting very active in the start-up community with the concept of, hey, this was something that DigitalOcean was great at. And you won't believe the number of people who come up to me regularly when I'm walking around with the DO T-shirt in an airport saying, hey, I'll learn how to code on Ruby on wing from DO. Yes. So we are trying to get back to that DNA with AI companies. Even last night here, we sponsored an event called Founders You Should Know, which is a very small, curated set of founders. And these are really successful serial entrepreneurs. And we are getting back to that route of seeding DigitalOcean as the place to start their AI journey, not just their cloud journey. So I feel really good about that.

And in August, you completed an offering of the $625 million convertible notes to -- in part retire the 2026 notes. Can you just bridge the gap between the remaining 20% to retire and how this is going to impact DigitalOcean's funding structure in the...?

Yes. I think we are in a really good place. I was just joking that I don't -- I'm so glad I don't have to take more calls on converts. So glad to have that. So I think we have a little bit of stuff left over, but I mean we are in excess of 40% EBITDA. So we throw out a lot of cash, and it gives us a lot of optionality to do all of the things that we just talked about, and we have a significant amount of runway between now and end of the next year to take care of this. So we have a very high degree of conviction that this is behind us, and we have multiple degrees of freedom to pursue.

I wanted to ask you a little bit more about product-led growth in [indiscernible] we had Canada, we have Versal, we have HubSpot conversations recently, where they both said that you sort of evolved from search engine optimization to AI engine optimization. I'm curious what you're seeing in terms of lead generation from LLM inferences and how you think about what positive growth looks like under that pyramid?

Yes. That's a great question. So we are -- obviously, we spend a lot of cycles tracking this. So the movement from SEO to GEO is real, and we are seeing it every day with tweaks to the Google algorithm and things like that. It makes a huge difference for us.

I'll start by saying our M1 to M12 has never looked healthier. It's an amazing engine for us. And SEO and our SEM spend is fairly miniscule, really small, like single-digit millions is what we spend. So it's not a big driver of our PLG motion. Our PLG motion is driven by community, our open source involvement, organic search, branded search is a very small part of our overall strategy.

We're also starting to get a disproportionate amount of our sign-ups from LLMs. But it's still early stages. We are getting a disproportionate amount of our sign-ups from LLMs, but their conversion rate and their ARPU is something that we are monitoring and tracking. Like are they coming here to do something serious? Or are they kids and students that are here -- yes. Sorry?

Experimenting.

Experimenting, yes. So we're looking at all of that. It's still very early days. And even for our product-led growth motion, we have multiple front doors, right? And the open source community is a great example where we get customers coming in from different open source frameworks into our PLG motion and then they become super entrenched customers of ours.

So unlike some other companies where SEM -- like even in my previous job, we used to spend like tens of millions of dollars to acquire customers through Google. That's not the case with us here at DigitalOcean. So I feel it is an important part, but it's not the most important part of our PLG motion. We have multiple feeding points into our PLG motion. But it's a really fascinating place where we're starting to see a significant impact of the Google Search algorithm and how we are bringing in customers into the funnel. Luckily, we have multiple bites of the apple, like we don't rely on search engine marketing to drive the top of our funnel.

Really fantastic color. Thank you for your time. [indiscernible].

Thank you very much. Appreciate it.