Good stuff. All right. We will go ahead and kick it off. Really delighted to be here at the opening company session, day one, Goldman Sachs Communacopia. I am Gabriela Borges, I cover software here at Goldman. My colleague, Maura Hager, 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 is wonderful to be here. It is 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. 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. At the time, the industry feedback on Paperspace was a little bit mixed. I fast-forward to today, and the business that you have built on what was originally an acquisition, along with the core IP of DigitalOcean, is really incredible. 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 a technology standpoint? Yeah. Thank you, Gabriela. That is a great question to set us up here. I think the first thing we had to figure out was what role did we want to play as an AI infrastructure provider, right? I think the first order decision was to figure out in the AI infrastructure space, there were two broad categories. One was training, one was inference. We made a bet, which at that time, a lot of people squinted at that decision saying, okay, we do not 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 that decision was, number one, self-reflective. 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. 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 the most part, versus the VCs or your investor money. That was the biggest decision we had to make. Then there were a lot of other decisions. Number one, we were fortunate to have an incredible talent density for building platforms. Over the last 18 months, we have added to that talent pool in a big way. 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 four or five very deep customers, because usually that takes you in a way that doesn't lend itself to building a broad platform. Having the luxury of now 680,000 customers, is a luxury that not many companies have, right? Then you fast-forward to where we are right now. 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 it. 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? That was one. 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. The cloud needs to be built for agents versus humans, and we call that the AI-native cloud. That's how I would summarize the last three 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. 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? You have commented a little bit on, well, we are actually starting to see real monetization versus just VC subsidies. Yeah. That are burning credit. Yeah. Talk a little bit about that dynamic. Yeah. 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 is very structured. There is a huge corpus of ground truth data you can feed into models. So there is 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 is relatively no surprise there, and we have a lot of customers that do that as well. 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 full-length feature films, changing the workflows of movie production. There is a lot of action there. It is also with OpenClaw and Hermes and other agent harnesses, personal productivity is seeing. GrokBot, I was very pleasantly surprised. I have been using it for the last 10 days. Yeah. It is an amazing product. All these personal productivity harnesses, and now I just heard about this company called Instinct. Instinct, over the weekend. 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. 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. If you take a step back and think about what our customers are doing, most customers start their journey with closed-source models, because you need to understand whether you have a product market fit. The best way to do that is, hey, give me the most expensive, most advanced models. Let me prove that I have a business. Once you are sensing and smelling that product market fit, typically two things happen. One, you have a real CFO, and you start looking at the cost of goods sold, and you're like, wait a second. The more we scale, the more this business model doesn't make sense. 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? The whole concept of owning your intelligence comes into play. 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, K3 was seminal in how advanced it was when it came out. A couple of weeks ago, we saw GLM-5.3, Qwen3.8, and the list goes on and on. I think the distance between absolute frontier and near frontier is collapsing every week. We are starting to see companies move more and more towards that. These companies are also starting to create more and more agentic workflows. That's just starting. If I look at this from my vantage point as an AI infrastructure provider, there are two slip streams. One, you have to be in the token flow, or you have to be in the agent flow. 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? Yeah. 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. You 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. Your tokens on Tuesday morning are of higher quality than the tokens you 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. 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. 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 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 bait 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. 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 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 ms, so it's instantaneous. You typically only pay for what you're consuming from a CPU cycle perspective. When an agent sleeps and awakes, it has persistent memory. When you look at many of the, 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. Then, of course, you need to have the ability to add security, unique identity management. You need to bring in providers like Okta or someone to make sure that your agents have a persistent identity. So this is what I mean by agent flow. 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 raw bare metal kind of infrastructure, the more you go from GPU economics to software economics. 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%? Yeah. The balance, the 85% is made up of inference services, which is basically all of the token economics that Paddy was describing. It is 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. 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 endeavor. You were talking about 70% gross margins, right? And 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 five 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 is really no upside to that, right? The 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 have abstracted the delivery of the value, which is tokens, from the underlying infrastructure. 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. Every single model that comes out requires different optimizations, and it is a daily game. You are constantly trying to increase your utilization. The more you can think about, well, what do I do in the off hours if I am 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. It becomes a price optimization and utilization game. It is a very different model with very different muscles that are required. And we have been in the 100% consumption-based business for close to 15 years. That is 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 are delivering beyond just, hey, it is X hours, X dollars per hour over a fixed period of time. 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 are in versus the actual software differentiation driving those pricing increases? Yeah, I would say that, 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 are constantly managing and optimizing the price optimization, and that is not something that we have to wait three, five years to do because we are under long-term contracts. So it is inherent in what we do and how we price our services. So we have 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 are just increasing their prices or we are 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. As Paddy said, once you start getting into the token economics, it is not a supply-demand, how many GPUs do you have? It is how many tokens can you give me at what level of quality? We control the economics on the back end. So it is a very different model that we are playing. Mm-hmm. As we think about the supply constraint environment that we are in, how do you go about securing incremental megawatts of capacity? Yeah, we have been really effective. People say this sometimes in maybe a pejorative way, but we do not play the gigawatt game, we play the megawatt game, right? We are out there looking for tens and 20s and smaller amounts of megawatts. The value for that is we are 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 de-risks 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. You have seen that in evidence. The three data centers we turned on this year, we turned all three of them on ahead of time, and are doing quite well in terms of ramping those up. As we are out looking for incremental capacity, we will take bigger locations than we have looked at today. 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 are profitable, we generate cash. A lot of what they are 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 very different credit profiles than what we would have. So we have been doing quite well, I would say, in terms of securing incremental capacity. In any given quarter, there is this dynamic where the capacity coming online is sort of baked into the model already. How should we think about upside in any given quarter to kind of the metrics you laid out? You want to take that or? Yeah, I can take it. I think that is not really true in our case. For most cloud providers, that may be true because the quantity is kind of limited, right? For us, when we look at the value equation, we look at it as P times Q. When 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 or driving higher utilization with all the techniques that Matt talked about in terms of bin packing, follow the sun, and things like that. That is 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 is just in the minute we put something on demand or spot, it is gone. On-demand, we never get it back. That is the thing. We have to call up customers and say, are you really using it? Or, we would love to have it back. Anyway, and then storage pricing, right? I mean, everyone is doing storage pricing. That is fine. We are also doing storage pricing. That is fine. These are all the levers we have in the queue. 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 three 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. 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. 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. The other thing that people miss is, for every one of our new data centers, we are not just deploying GPUs, right? We are deploying a full stack AI-native cloud. 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. We have the ability, the more AI traffic we drive, the more core cloud consumption we drag through and attach. That is another lever we have. 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 three weeks ago, because we have a fungible on-demand fleet of H100s, and 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 is 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 stand-up. We look at the price side of the equation really closely to maximize the fleet utilization. At the end of the day, for us, it goes back to we fundamentally believe that software makes megawatts more valuable. Mm-hmm. As you are 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? Yeah. We have not changed our views of what an acceptable return on investment is, nor our payback. We are very excited by the opportunities that we have in front of us to invest. You have seen this. We are, I would say, appropriately conservative. When we underwrite a new data center, a new GP investment, we underwrite it with very conservative revenue assumptions. We assume price compression, which is, in fact, we have been wrong, right? Prices have been going up, but we are underwriting it with price compression. We target paybacks that you would expect. We talked about in the 3-ish year kind of a range. What you are seeing now in the market is the CapEx per megawatt is increasing, and it is 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. 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. 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 pick your capacity. As you think about the next 18 months, how much is baked in terms of the queue 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 as is 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. 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 similar return characteristics. Right. If you said, hey, how big could you get, how quickly? 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. 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. I can tell you we've been working on capacity for the last 18 months. Right. We feel good about our ability to- Right. Really flex that, the queue as well. We're all on the edge of our seats waiting for whatever you'll eventually tell us about the 2027 guide. Yeah. 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? Yeah. 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. Think of all things that have changed since we gave the 50%+ guidance for 2027. One, we have some nine-figure deals to give us visibility that we didn't have when we had that. Two, we've added some incremental capacity. We announced 20 MW 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. The P times has now got a big lever. We've also increased the guidance and outlook for exiting this year to 35% +. We're already going to start a decent amount higher. A lot's gone really well relative to and we're turning on data center capacity on time and even ahead of schedule. Prices are going up, not down. There's a lot I'd say that's embedded in that as you think about next year. It's one, we said 50% +, that's a full year number. 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. Our goal is to take advantage of this massive opportunity that's in front of us. It's a generational opportunity. 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 2027 and beyond. I want to end here on a comment where Paddy talked about reassessing the go-to-market and using some more AI tools. At the same time, Matt, you have commented on the deals getting bigger. We remember two years ago, actually, it was probably three years ago now, where DigitalOcean said, Look, we are going to do direct sales reps. We are going to land larger customers. It was really hard to get off the ground back then. You have actually gotten it off the ground. How does that go-to-market motion evolve from here, and where do you start running into more of the neocloud and hyperscalers? Yeah, that is a great question. 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 is hard for go-to-market, honestly, to keep up, right? Which is a great problem to have. We have a new CRO now, Kevin. Yeah. who came from Vercel. He definitely speaks the AI-native language. 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 is probably 120 days or something. We had 6,000, 7,000 customers already on it, and it is just a luxury that most companies do not have. Building on top of that, we are increasing our direct hand-to-hand customer acquisition strategy with the top, I do not know, 300, 500 AI-native companies. It is 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 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 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. We are fortifying our ability to go have these conversations. We have two 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. We are trying to throw out any existing playbooks. We must invent a new playbook because there aren't too many companies that have figured this out yet. Okay, fantastic. Please join me in thanking Paddy and Matt for their time.