All right. Fantastic. We will go ahead and kick it off on stage with the CoreWeave session at the Goldman Sachs Communacopia + Technology Conference. I'm Gabriela Borges, and I'm delighted to have Mike Intrator on stage with me, CEO of CoreWeave. Thank you for coming.
Excited to be here.
Hey, Mike, I wanted to start with a little bit of a technical question for you, which is, we've all seen the data points from third-party industry. We even, selective cases, have been able to speak with customers, and they'll consistently tell us that CoreWeave is able to deliver a level of performance on GPU training that is unparalleled in the industry. Give us your best layman's explanation on why you think you've been able to go from a crypto miner to offering the best performance in GPU training in the industry bar none.
Sure. I'll do the best I can. One of the things that's important to understand is when we built the company, we had an opportunity to break it down to first principles and say, "If you were going to build a cloud, specifically built to address the needs of artificial intelligence, parallelized computing, how would you go about building that?" The clear answer was, you don't go about doing that by retrofitting a legacy environment to go ahead and create the solution. We jokingly talk about it internally, it's like the minivan versus the F1, right? It is a tool that was specifically built in every decision from when we founded the company in order to optimize for this specific use case, not to provide generalized computing. There are lots of great options for that.
What you have is an environment that has these compounding, incremental, thousands of decisions that have been built around making the environment as performant as possible for this specific use case. We select our clients carefully, in terms of finding folks that need this use case to be able to be supported. You find a software stack, an infrastructure stack, all of those things really are built to be able to provide that environment in the way that we do. It is great to hear that the clients that you are speaking to are taking advantage of the efficiencies that the tooling that we built provides them.
The next question I have for you is taking some of those performance decisions that have been made for training GPUs and then extending that advantage to something like post-training and something like inference. Talk to us about why those decisions that optimize for training also give you a performance advantage for post-training or inference.
Sure. One of the things that we believed very early on was that the space was going to evolve very quickly. When you are in an environment as dynamic as that, you want to build your infrastructure to have the most amount of optionality embedded in it as possible, right? We do not really think about our infrastructure as being, "Oh, this is our training infrastructure, and this is our inference infrastructure." We think of it as AI infrastructure, and we build to a standard that allows our clients to use that infrastructure in whatever way they deem to be most productive for their businesses. That is how we built it.
All of the tooling that we have within our Mission Control and all the other observability suites and all the tooling that we have provides for an environment that allows for incredibly performant environment for folks that are using the compute. We can see when people are using it in different ways. The consumption profile will look different. All of those things will follow with their use case. To varying degrees, different parts of our stack will provide for different advantages, depending upon what their use case is.
This all makes sense. I want to pick your brain on a couple of industry dynamics that we get asked about all the time that are very core to where CoreWeave fits in that stack. The first one is supply-demand, where CoreWeave has consistently said that we might not be at supply-demand equilibrium until the end of the decade. My question to you is, tell us a little bit more about how the demand environment has evolved this year because of agentic AI, and where do you think it goes from here? What do you think is being underappreciated about agentic AI demand over the next 12 months?
Yeah. We have been incredibly consistent in the way that we talk about demand, and that has been driven by our ability to talk to an incredible cross-section of compute consumers, right? That's everything from the hyperscalers, through the AI natives, the frontier labs, people that are productizing and delivering agentic AI solutions, all of those things. We have never flinched from our assessment that the world's capacity to deliver compute has been, and will continue to be, wildly overwhelmed by the delivery and demand for intelligence, broadly speaking. We've been righter than wrong on that one. I've got a bunch of really smart folks around me that come and email me every day that we don't have enough compute. I'll take their word for it.
One of the really interesting, I'm not sure this exactly answers your question, but one of the really interesting phenomenon that you're seeing in the market right now is the large consumers of our services of compute delivery of the software stack that we've built on top of that. Those consumers are demanding ever larger amounts of resources, right? We have the nine-month rule at CoreWeave. We look back at whatever we've done, and nine months later, we think it's cute. That's the dynamic that we've seen in the market. But that is being compounded right now because of new entrants into the market, and new entrants are enterprise clients coming in. I talked about this a little bit on my last earnings call where a company like Caterpillar comes and starts to use us.
It's a great example of enterprise consumer coming to us and saying, "Hey, the way that we're going to participate with our company in artificial intelligence, the way that we're going to train our models, the way that we're going to serve our models, is going to look different than it has historically. We're going to build our own clusters." That is transformational, right? That is really allowing CoreWeave to establish itself as an AI cloud participant, a hyperscaler within the AI system, and that's been really, really exciting for us.
I am curious what you think changed that now has enterprise customers wanting to build their own AI clusters in-house versus outsourcing.
I think it is a lot of things. I think part of it is the experience of building cloud-based systems over the past 20 years has given companies some perspective about how important compute was going to be. Then with the advent of artificial intelligence and the degree to which that kind of raises the stakes, there is an interest in having more control over that resource that they consider to be fundamental and important to their businesses than perhaps they did historically. It is the boiled lobster problem, right? The cloud kind of came out of nowhere, and people began to use it, and people were using it before they realized how incorporated and embedded it was going to be within their critical systems.
This time, they are going in with their eyes wide open, which really provides us with a wonderful opportunity to go to them and say, "Hey, we can really assist you beyond what has been an oligarchy of three massive companies that have delivered this service. We have some alternatives. We have a different way of looking at it. We have a different way of providing the compute that will be most performing to you." That drives down cost. It does all kinds of great things for them.
Very good. My second industry question is world models. When do you think world models start to have an impact on the demand curve to CoreWeave? I know you have already announced customers in this space.
Yeah. Look, it's the same but different is the way I look at that. The value propositions that a company like CoreWeave provide to that space are really quite similar to the value propositions that we provide across the entire space. Yes, their data's a little different. All of those things are true. I do think that really when you're thinking about the world models or agentic AI, it really comes down to have you built your infrastructure, have you built the software rails and the stack to provide an environment that is going to be incredibly performant, incredibly well-controlled, incredibly cost-effective from the amount of compute that you're able to squeeze out of every dollar you invest. All of those things are the North Star for CoreWeave. It's how we've built our business from day one.
We're about to go into midterm elections here, and we've seen on both sides of the aisle a number of debates around data centers that in aggregate have made ease of scaling and securing a data center harder than it was a year ago or two years ago. How do you navigate at CoreWeave the stuff that's within your control? How do you plan for what's not in your control? Where do you think this debate goes over the next year?
Okay. Somehow in a relatively short period of time, I turned into a bad guy, according to my son, because of the data center dialogue that's out there. Look, we have a method and approach to how you go about building and scaling data center capacity that is really built around entering the markets early, really working with the local communities, being sensitive to their needs, and building infrastructure that is useful to our clients while not being unreasonably irritating to the local communities. That has been a very productive way to approach building infrastructure. The world is definitely evolving, right? There is increased resistance to data centers going into communities. There are pieces of that debate that I think are completely nonsense, and I think there are pieces of the debate that warrant real introspection by the industry.
Water usage, these are closed-loop systems and as more of the data centers come online that are able to support this AI infrastructure that is so power intensive, they must be closed-loop systems. They don't function unless they are. I think that going into the elections is going to be a maximum volume. I think that the approach of not allowing data centers to be built is kind of a fool's errand in some ways, because I think the demand for the compute doesn't change. It just moves in terms of where it's going to be built and how much it's going to cost to build. But I don't think it changes the overarching demand cycle. We have been really forward-thinking around the idea of diversifying our portfolio of data center infrastructure.
I talk a little bit about this on my earnings call because it's such an incredibly high-profile issue. I talked about the idea that we have over 1 GW of power contracted outside of the U.S. as you look to build a portfolio of infrastructure to be able to support your clients with. I think that's the right strategy, and I think we'll continue to focus on making sure that we can build both within the domestic U.S., but abroad as well.
Maybe I can bridge this to a horizontal and vertical integration question. We'll talk a little bit more about your ability to go up into the software stack.
Sure.
What I want to ask you is why now you feel the need to also own the underlying data centers. I think you have your first data center coming online by the end of this year. What's driving that decision on vertical integration beyond the margin stacking argument?
This is not a new issue for us. We've thought a lot about this. We've spent a lot of time thinking about which parts of the vertical integration we want to be involved with. There are two reasons that you want to vertically integrate downward, and both of them are very important. First one is that by vertically integrating down into the physical data centers, you can recapture some margin, right? I would say that's the lesser of the two. The second thing is, this stuff is difficult, right? It's not, "Hey, I've got some megawatts and I'm going to turn that into a supercomputer." There's a lot of steps along the way there.
Entering into an environment where we have more control, a deeper understanding, a better sandbox to work on the flywheel of new innovations that we want to bring to our own data centers, that makes us stronger, right? More control over pieces of our infrastructure make us better. It is not a new part of our strategy. We have thought a lot about it. Data centers take time to build, so you are starting to see the first two to three of our self-builds spin up, which we are incredibly excited about. But you will see a continued effort on our part to control the physical infrastructure both up and down the stack. We think it is important. We think control is important, learning is important.
Well, let me stay on this point on bringing more capacity online is very hard operationally to do. We have seen so many announcements from new entrants in the past several months where we will get data point on companies bringing on 10 GW over the next two years, 7 GW over the next two years. How do you think this all shakes out from an industry structure standpoint?
I think it is hard. I think building infrastructure is difficult. I think that some will be more successful than others. I think that it is capital-intensive industry, and capital-intensive industries will tend towards periods of proliferation, and they will tend towards periods of consolidation. I think that during a time where the demand is so incredibly overwhelming, it is easier to launch a company. But the difference between launching a company and talking about a contract you sign, and actually being able to operationalize and deliver and maintain that infrastructure, the more experience you have with it, the more you realize how difficult it is to do over time. So I think there is going to be some interesting outcomes that are going to be associated with delivering this new infrastructure.
Absolutely. Okay. Let me ask you the supply demand question in a slightly different way. So we have established that we are in a period of incredible supply constraints. At some point, and we can debate when that is, the industry will come back into supply-demand balance. Right now, CoreWeave gives us these amazing data points on contract renewals to H100 and pricing dynamics. How should we think about longer term, what happens when supply-demand normalizes? The bear case we hear from investors all the time is, well, you have hyperscaler customers who can simply then pull back what they have "outsourced" to CoreWeave and return back in the stack. So maybe just would love your thoughts on the longer term for CoreWeave in a normalized supply-demand environment.
I think my last answer was the most politically correct answer I could have given to that question.
Very fair.
Look. CoreWeave has been leading the AI cloud space now since its inception. I would argue that largely CoreWeave has really built the space. What started out as a cost-plus environment where we were able to win contracts because we would go to large consumers of compute and say, "Hey, we'll build this for you." And they would say, "Well, we're willing to pay cost plus for it." That kind of built our business and allowed us to scale. With scale in this business comes all kinds of incredible ways of driving economics and driving margins, which we're really excited about, and the market will see that continue to accrete to us over time. The other piece of it is just these companies, they entered into contracts with us, and when we IPO'd they were like, "Oh, you're never going to do that again.
It was just a one-off." Then they came back and did it five more times with us, and then they did it across the industry. The argument continues, "Oh, they're just going to pull it back." That doesn't really Well, there's two things, right? It doesn't really match with the fact pattern, right? They built data centers, and they could have pulled those back, but they didn't, right? They go through these periods where they build their own data centers, and then they go through these periods where they go to third-party data center providers. So that's one thing is we do not believe that they are going to pull back, right? We just don't think that that is the way that it's going to play out. But the second piece of it is what these contracts did for us was provide us with time, right?
And with that time, you have seen us build a cloud that is specifically built to be able to support AI use cases. With the software stacks, with the infrastructure, with the reputation for delivering this incredibly high-quality product, we have begun to win a broader and broader universe of clients. Some of those contracts may come back to us, some of those contracts may not. But the size and scale of the business, our ability to support enterprise, our ability to support governments, our ability to support our own products internally, our managed inference product, all of those things, those internal demand things, those are incredible paths to building a sustainable cloud.
And we believe that is a necessary part of exiting this period of disequilibrium in a way that allows you to be successful and a hyperscale provider of the infrastructure that is required by the world for artificial intelligence. And we're really well-positioned to do that.
Let's talk about building the cloud. There are a number of pieces that I think are interesting. We could talk about managed inference, for example. There's a CPU pull-through piece to this as observability. Maybe just take a step back for us. What are the pieces that you think are most strategic to going from AI Neocloud to full-blown long-term hyperscaler winner?
The answer to that is we meet our clients where they are, right? When we go and we work with a client, they tell us what they want, right? And we can be extremely flexible about integrating, "Hey, we've got this particular type of storage we want to use." "Okay, we'll build it into our system. Not a big deal." So it really depends on who the client is, what the workload they're trying to support will define what we bring to market. The attach rates, and I think I spoke about this maybe two earnings ago. 75% of our clients are using three different services, right? Three different silos within our product suite already being consumed, right? That's exactly what you want as you're building the cloud. That is exactly the way that you measure how people are going to consume what you're doing.
I would argue the percentages are even higher than what they look like because some of our clients are these really early-stage ventures that are not up to product number two and product number three yet. They're still kind of working it out. We really think that you've got to build a suite of software services, whether it's storage or memory or networking and all the different things that people are going to consume, then go ahead and put it in front of them and say, "Look, we can help you build your company. We can help you deliver your service. We can help you be successful because of the infrastructure that you are going to be able to run your products on." That has been an engine for us.
Absolutely. In terms of roadmap, anything you can share on where customers are pulling you next?
We've always talked about the fact that we are client-led, right? Being client-led means that you spend time with your important clients, and you talk to them about what they're trying to accomplish, and you let them kind of help you build the product. A lot of the storage solutions that we have were built specifically to help folks that had a very specific problem. When you look at the way that our roadmap has led us internationally, right, to ensure that we are both resilient from a capacity factor, but also able to serve different markets with low latency infrastructure. All of those things are really manifestations of our clients and what they need, and that really does inform the roadmap.
I mean, sitting at the center of the flywheel, having the conversations with all of these incredible entrepreneurs or existing companies that are integrating artificial intelligence into their workflows. That is really, really fascinating, right? Just absolutely fascinating to see how that works.
Maybe the perfect case study is actually someone like a Caterpillar, which you were talking about earlier. Give us a sense for the flavor of the enterprise customer conversations that you're having today, where customers will say, "Look, we had previously used one of three hyperscalers.
Yep.
And we're now incrementally moving net new workloads or even ramped up workloads over to CoreWeave." What does that look like in practice?
What it looks like is when we bring a customer like that over, we try to wrap our arms around them and support them through the process, right? It really depends on who the customer is and what they're trying to build and what they're trying. We try to make it so that all of their resources are being focused on integrating AI to make their company better, faster, stronger, however you want to talk about it. What we really are trying to do is simplify the process of ensuring that they have the most performant infrastructure to be able to train internal models, right? That they're able to serve their inference calls effectively. All of those pieces have to come together for us to be viewed as a viable alternative to one of the hyperscalers.
The breadth of the services, it's not endless, right? We really do say, "Hey, this is our lane, and our lane is to be able to serve all the components that are required to drive artificial intelligence." We don't wander too far from it, right? The world doesn't need another solution for X, Y, and Z. What it needs is a best solution for the AI workflow, and that's where we're focusing our energy.
Mike, you have already touched on the diversity of your customer base. Do you have a view or how do you think about the mix between frontier models and open source, open weights? Does that have a second derivative impact either on the health of your customer base or the health of your business?
We are in a position when we think about closed source versus open source of believing that this is not going to be a binary breakdown. There are going to be use cases for open source. There are going to be use cases for closed source. As a supplier of the cloud that is required to run these use cases, regardless of whether it is open or closed, we are sort of indifferent. We want to make sure that our infrastructure is able to serve both sides of that fence equally well, equally as efficiently. That is sort of how we have approached the problem. I think there are other people that have to spend more time deciding which models are appropriate for which workflows within their organization.
But for us, it is the idea that we will provide the computing power that you need in order to be successful regardless of how you choose to allocate the infrastructure in terms of whether it is open source or closed source model. We are sort of a cop-out, but we are sort of indifferent on that. It is just more demand for compute is better for our business.
The other part of the strategy that I wanted to ask you about is the relationship with NVIDIA on selling two pieces of software. Maybe bring us up to speed on some of your ability to take proprietary CoreWeave software and sell it outside of the immediate CoreWeave stack.
Yeah. That was pretty cool.
It's very cool.
Yeah. It was. It really was. NVIDIA designated our software stack as a reference architecture, and that's super exciting for us because it really is a recognition of the quality of the solutions that we have built, right? That was great. We had some really excited engineers that have put their blood, sweat, and tears into building the infrastructure. It was a wonderful recognition that they thought the quality of what we had built was high enough to be designated a reference architecture. That was really great. What it means in reality is that the ability to provide third parties access to our software solution on their infrastructure allows us to get some leverage on other balance sheets. That's a great thing, right? We can go into companies that want to own their own infrastructure, and we can provide them a software layer.
It's called Omni, is what we call it. It also allows us to build infrastructure in jurisdictions that we maybe are not comfortable owning or delivering an asset-heavy solution to, for whatever reason. Once again, that's a great way to make use of other entities' ability to take the risk around the physical infrastructure while still being able to generate returns and to expand our TAM for the products that we bring to market. That's been a very exciting part of our portfolio. We closed a deal on that a couple of months ago, and there's a whole bunch coming down the pipe that really look like that for various reasons why we would use that model instead.
I want to spend a couple of minutes here on capital structure, because the appetite of the debt markets to own Neocloud businesses ebbs and flows over time. Tell us a little bit about your plans for capital raising. How do you think about the optimal balance sheet structure for a company like CoreWeave?
Yeah. Look, we took a strategy around building our company that is, in many ways, part and parcel for how you go about or how the world has historically gone about building infrastructure and capital-intensive business. If you're SpaceX and you're valued at $2 trillion and you can sell some equity, you sell some equity and you can go ahead and build whatever you need to build, because $2 trillion is a lot. If you're building organically, like we were, the question is: how do you raise the capital to build at a scale that allows you to be of relevance? The answer to that was clearly the debt markets. I think that the market struggles to understand what we have done in the debt markets.
The largest engine of our borrowing is occurring at an SPV level, where we are providing transparency through to the credit of the offtake. That is why we were able to build structures that allowed the credit market to underwrite what we were doing. Because they were basically looking through and saying, "Okay, it's Microsoft on the other side of this transaction. The money's going to flow into the SPV and then pay back the debt before it goes back to CoreWeave." That was an incredibly effective engine for raising capital at an order of magnitude that has rarely ever been done by a company as new to the market as CoreWeave. That was a wonderful way of enabling us to drive the scale that we needed to be able to serve our clients. We also raised some debt at the parent co.
We raised convertibles up at the parent co. We have a very not a lot of religion around this stuff, right? We look at the space and say, "What is the most effective way, the cheapest way, to raise the best possible capital to be able to execute on our roadmap?" That has been the defining North Star around the capital market structures that we use. We have been the tip of the spear around the innovation deck. We did the first GPU transaction. Just kind of looking back in history, the last debt structure we did was the first time anybody has been willing to take, or any of the lenders have been willing to take renewal risk.
That's an important step forward to being able to access capital markets to be able to take advantage of the short-term contracts, because now the lenders are taking renewal risk. The one before that, we were able to get the venture rated so that we were able to borrow at A- credit. That's unbelievable. For a company like us to be able to borrow at such a low cost of capital was kind of a crowning event within the group that builds these capital structures. It was incredible. It puts our borrowing really at par, or close to at par, with many of the hyperscalers that are out there raising capital, and that allows us to operate more effectively and more competitively.
Hey, Mike, congrats on all the milestones. Please join me in thanking Mike for his time.