DigitalOcean Holdings, Inc. (DOCN) Earnings Call Transcript & Summary

All right. We will go ahead and kick it up. Really delighted to be here at the opening company session day one, Goldman Sachs Communacopia. I'm Gabriela Borges here at Goldman. My colleague Maureen to 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 our wind back to when you first came in as CEO. And at the time, the digital vision strategy in AI, in shut an asset called paper space, which was applied 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 a 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, much more deep from the technology standpoint?

Yes. 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 and 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 highsight, the reason why we made that decision was, number one, self reflective, like what are we good at? We are really good at understanding development we are really good at building platforms. We're really good at managing global scale infrastructure for production workloads. So that was a big part of it. The second was inferencing, we believe 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 the 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 -- I mean, 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. It cannot assemble it -- this is not an application application portfolio like Salesforce. There is the 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 second -- another important decision was the order of operations. The sequencing really matters. We build software first. Now we are adding scale, right? A lot of companies went for scale first and now our building software. We'll see where we all end up, but we like our chances and our rough 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, coating has been one of the big agentic use cases, maybe customer experience to 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 building. And you've commented a little bit on, well, we're actually starting to see real monetization. This is just VC subsidies that are burning credit Talk a little bit about that.

Yes. So when you look at our customers and what they are 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 model. So there's a lot of reasons why coating has really taken off. And coating 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 film, changing the workflows of movie production. So there's a lot of action there. It is also with OpenFlow and Hermes and other agent harnesses, personal productivity is seeing in 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 Instant right, something like. Yes, yes. So there are a lot of these personal productivity age and 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 close 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 late second, if we the more we scale, the more this business model doesn't make sense. So they start looking at open rate models. The second thing is then you start thinking about, "Hey, are we just one feature update from being completely disintermediated by the close 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 at the near frontier space gets pushed, K3 was seminal in how advanced it was when it came out. A couple of weeks ago, we saw GM 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 to explain, this is 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. We 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 1 example of post training. So you're talking on Tuesday morning or 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 as 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 monetize at a completely different rate versus a Deepseek flash. DeepSeek flash is great for 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 rate category. Agent flow on the other hand is how do you build and scale agent 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 affirmal, but they need persistent memory, right? Agents are very short-lived. So last week, we announced a new product called Agent harness open harness run time. And this open harness one time enables customers to bring any hardness, whether it is Hermes or codecs open clot, 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 is 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 awake, it has persistent memory, right? So if you look at many of the -- like, for example, let's say, you're automating and identifying an STR 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 persist 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 understanding 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 front times, 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 more in that into the conversation at some numbers around that? Yes. Let's talk about this progression from bare metal GPU to more of the managed services and comics. You're at around 15% of bare metal AI revenue and the rest of the 85% is these higher values. Can you talk to us about the unit economics of these higher-value services relative to the bare metal 15%?