Is AI a bubble or the biggest boom of our lifetimes? The fate of one company, CoreWeave, may tell us everything we need to know. We'll be back with the company's founders right after this.
Welcome to Big Technology Podcast, a show for coolheaded and nuanced conversation of the tech world and beyond. We have a great show for you today because in studio with us are the founders of CoreWeave. CoreWeave CEO Michael Intrator is here with us. Michael, welcome.
Thank you very much. Great to be here.
And CoreWeave's Chief Strategy Officer Brian Venturo is also here. Brian, great to see you.
You are both running one of the most fascinating companies in the AI boom. Everyone has used you effectively as a Rorschach test to read in their beliefs or insecurities about what's going to happen in this AI moment. Some people think that you're the poster child for the AI bubble. Others think that you're perfectly positioned to take advantage of the boom in building that is occurring as demand goes through the roof. A couple stats about you. As of today, the company is worth $42 billion after an IPO earlier this year. You've built eight new data centers across the US in the third quarter alone. And the latest reported numbers have you in possession of something like 250,000 of Nvidia's GPUs, which are the chips that companies use to run AI models and grow them or train them as they like to say. Let's just start off with this because it's been a heck of a ride for you over the past couple years. What has it been like being on the front lines of this AI buildout? Talk a little bit, help people feel the speed at which it's boomed and what it's taken to do something like build eight data centers in a quarter.
All right. So, let's start with that. It's been exhausting.
Yeah. You hit it dead on. It has been incredibly exciting. It has been an unbelievable year. We just IPOed really eight months ago, and it feels like it's been two lifetimes. The company is moving at incredible speed. We are building a massive percentage of the global AI infrastructure that's required to allow artificial intelligence to be what it is. And when I say massive, it's a meaningful percentage.
What's your estimate about the percentage?
That's tough. We think of ourselves as providing enough of the compute that we have the ability to be relevant in the debate of how AI is going to be built and how it's going to run into the future. We don't know what the numbers are. There are lots of different providers of technology. They're being used and there's no real good way to put your fingers on the data, but meaningful. And that's an exciting place to be. Honestly, we talk about this in the company all the time. It's a privilege to come into work and focus your energy, your creativity every day on building a component of artificial intelligence, which is the issue of our time in many ways. And we get to really sit there every day and pit ourselves against those issues. I have a ball with it.
I'm taking a shot of this. Before we move on, I think that's really around the practical side of it. When you're a company growing as fast as we have, where we had maybe 100 employees three years ago, now we have 2500 employees or so, there's an emotional side of this too. Since the IPO, we've been under this spotlight in the world of 'What are they doing? How are they doing it? Are they executing?' Internally, we always set the highest bar for how fast can we do something, how high of a quality can we do it at. As this industry has expanded so rapidly, things happen. You have weather that impacts construction on a project. You have a truck that hits a bridge. All of these random exogenous or idiosyncratic things happen in a supply chain and then it comes back to us and the world is like, 'Wow, you failed.' Inside the company, from a culture perspective, it's been so important for us to manage that. We're doing something at a scale no one's ever done before at a speed no one's ever seen before. Of course, things are going to go wrong, but take perspective. See how much we've done. For our employees, if you're moving at a million miles an hour and you hit a speed bump, it's okay. It doesn't change the trajectory of what you're doing. It just provides the battle scar so it doesn't happen next time.
Yeah. Now, I can imagine it's a rough and tumble world trying to build this with very demanding customers, very important technology that you're deploying and the speed is crazy. It is interesting looking at your founding story. You really started working on providing infrastructure for crypto, was it like Ethereum mining or something like that, and then pivoted in a very smart way to this AI moment, establishing a relationship with Nvidia that's proven to be very useful and helpful for you and probably for Nvidia as well. And now you're again in hyperdrive building data centers. The data centers are, if I have it right, largely licensed or the capacity is rented out mostly to the tech giants. The core customer is Microsoft, something like two-thirds of the demand according to your public filings, but there are others as well.
We actually spoke to customer concentration in our last earnings. There is no customer that represents more than 30% of our backlog. We've done an incredible job. It's been a focus of the company, everything from sales all the way through the build cycle, to really begin to broaden the reach with which our solution touches artificial intelligence. Microsoft is an important customer and a large, creditworthy, and formidable part of the AI ecosystem at large, but we've done a really good job bringing on other wonderful clients, wonderful customers that are going to continue to use our solution as they build their products and deliver them to market.
Okay. And I definitely want to get into customer concentration in a little bit. That's a good preface to what we'll touch on and already some new data to me. Good to hear that. But I wanted to again get into what it takes to build these data centers. You're assembling them with incredible speed. So I just want to hear a little bit about on the ground, what does it take to put together these data centers?
Historically, let's say two years ago, we were able to go out and buy capacity or lease capacity that was much further through the development cycle. They were basically the shell already existed. It was a fit-out construction process, which means going in and installing the last pieces of the cooling infrastructure, cabinets, conveyance for all the cabling, all the hundreds of miles of cabling we have in these things. But it's shifted over the past year. Now we're doing much more bespoke in-house design to make sure that we're meeting the needs of what our customers' deployment is going to be. It's everything now from how is the cooling and electrical distribution designed, how are we ensuring electrical redundancy and reliability, how are we cooling the air-cooled side of these things? Because you have liquid cooling, but there's still a component of it that has to be cooled with air.
Can we pause on that? These chips run extremely hot, right? Cooling. People talk about cooling for those people who are coming to this for the first time. Being able to run an AI data center, you got to be able to cool the chips if you want to be successful.
This is one of the things that I think the market misunderstands. Everybody believes that there's some differentiation in the plumbing of the liquid-cooled data center. That's not where the differentiation lies. It's all the same pipe and valves and fittings. Everyone's using the same things there. The differentiation comes after you turn it on and how you control those systems. That's what we've done incredibly well as a company. We've very consciously not spoken about it externally for the past couple years because it is our secret sauce: how we provision, validate, and manage those data centers all the way from the power cooling infrastructure up through the GPUs, the servers. And it's why the most valuable companies in the world, the biggest AI labs, actually use us to run their most critical training jobs.
It's a herculean task, right?
It's important to understand that when you're thinking about the ecosystem and the different neoclouds that populate the...
The worst term ever. I hate it. Think of it as in the common vernacular, everybody knows who AWS is, Amazon, Microsoft, Google. Those are the hyperscalers. You can throw Oracle in there if you'd like. But then there's a class of providers that can deliver this infrastructure. We are the leader among that. It's important to understand that if you took all of the other neoclouds and added their GPU fleets up, we would still be a multiple of all of them combined in terms of the number of GPUs that are up and running and delivered to clients. A large multiple.
When Brian is talking about things that the market is struggling to understand, it's important to understand that what differentiates us, what allows us to be as successful as we have been, is that the software suite we have built allows us to take the commodity GPU and deliver a decommoditized premium service that allows people to extract as much value from this infrastructure as possibly can be extracted. That's really what CoreWeave is doing. It's why when Brian says the leading companies in the world and the leading labs in the world are relying upon us to deliver our service, it's because the product they ultimately receive is the product that will allow them the greatest probability of being successful at using the GPUs to deliver the products that their company is building.
So, just to put it in plain English, always helpful for me. When a company like Microsoft will work with you on building infrastructure for artificial intelligence, you've built some proprietary pieces of the puzzle, like your cooling system, like the software that runs the data center, and that allows them to get more out of the chips than they would have typically.
Yeah. The nuance here is that when you build one of these data centers and it has 3,000 miles of fiber optic cabling and it has a million optics that connect into the switches, these things all fail. When they fail, the way that training jobs are run today is if one component fails or one component limits the performance, the balance of the training run is going to be governed by the worst performing component. Our entire job is to build the automation, the predictive analytics, the machine learning models around saying, 'Okay, we're seeing a problem here. How do we gracefully handle these things so it has the least impact on our customers' jobs?' That's the core secret sauce. We have the world's largest dataset of how these things run, how they fail, and we've built all the recovery mechanisms and the software intelligence to help our customers run these things.
Is the demand that you're getting from your customers mostly training the AI models? Because that's what a lot of the infrastructure has been used for: building scaling these models, throwing more compute at them, throwing more data, making the models bigger, and the idea is that the models get better. So are you seeing most of your demand in the training side of things or has it gone to inference where companies are actually using the models and deploying them into production?
It's a great question and it talks to the split or the delineation of where the market's been for the last three years and where it's going. Our customer base for the last three years has primarily been the largest AI labs and enterprises that are building the capabilities of AI. It's now shifted from the people building those capabilities to the people that want to use those capabilities to change business outcomes. This is where all the enterprise adoption is coming from. One of my favorite services out there is Lovable. You go to Lovable, you can build any app you want. There's a chatbot that helps you go through it. We're finally starting to see people chain together these capabilities to build real products that solve problems. Our business for the last three years has really been around the creation of those capabilities and has very quickly shifted to include not just the creation of them but the deployment of them and use in business practices. One of the things I didn't expect was that what looked like training two years ago is how inference was going to look today. You're still dependent upon highly connected storage. Your backend networks become critical to this because the models are so large. There's really no difference between training infrastructure we deploy to build those capabilities and what our customers are ultimately using to serve them.
So has inference overtaken training for you?
We serve a tremendous amount of inference. I actually don't know the answer to that. Six months ago I would have said it was two-thirds training and one-third inference. It's probably close to 50/50 now. There's also some of our big customers that they go from using a campus for training, they'll launch a new product, they'll have to spill over for inference. A lot of this is very dynamic and it's been built to be so.
This may provide a segue to some of the other subjects that you'll ultimately get to in this podcast. For me, watching inference, understanding that inference is the monetization of the investment in artificial intelligence, is one of the most exciting trends that exists within AI. We have a front row seat across the entire cross-section of almost every large important lab that's building this stuff. Watching them increasingly move from one-third inference climbing towards 50%, and at times it's even over 50% of the fleet being used for inference, is just an amazing indication of the scale of the demand to use artificial intelligence to serve customer inquiry. And that means everything.
All right, one more question about this. Why does CoreWeave need to exist? Why wouldn't these big companies like Microsoft just build their own data centers? Why are they licensing it from a third party?
It's a great question. There was a void in this market. The biggest clouds in the world today are built off the cash engines of peripheral businesses. Google's built on search. Amazon's built on retail. Microsoft was built on enterprise software. We came pretty much out of nowhere. The moment in time for us to be able to get ourselves into this position was driven by crypto. You mentioned earlier that we came out of Ethereum mining. We were able to leverage the revenue from Ethereum mining to go out and build and deploy additional scale so that when crypto went away, we had the infrastructure in place and we hopefully had enough clients that we became at escape velocity. We recognized that compute was going to be valuable. We didn't necessarily know at the time what it was going to be valuable for. I don't think Mike and I ever had this idea of there's going to be this hundreds of billions of dollars a year in capex for AI. But we had the thesis that compute is going to be incredibly valuable and we wanted to own a lot of it. We looked at that compute resource as an option. We said, 'Okay, what are the best things that we can do with this?' That's how we've always approached different business problems: what is our asset, how do we monetize it the most effectively, what's the most valuable way to use this?
I'm going to jump in here on this, but I want to go back to something that we kind of talked through as we started this. We've built a software stack from the ground up to optimize for the use cases associated with parallelized computing. We do it better than anyone else. The reason we exist is because we deliver a fantastic product that is highly in demand and incredibly differentiated. We serve the largest players, but we also serve a ton of other AI companies that are building applications where they have the choice to go and use us or to go and use one of the hyperscalers. Many, many of them choose to use our solution because it allows them to more effectively deliver compute. One of the things that's really lost on this is that there's not an understanding of how fundamental the change from cloud 1.0 into cloud 2.0 is as you moved from sequential computing into parallelized computing. When you made that leap from hosting websites and data links into driving parallelized computing for artificial intelligence, it stands to reason that a fundamental change in how compute is used will also require a fundamental change in how you build the cloud to serve it. We took advantage of that transition to build best-in-class solutions. And that's why we exist.
So, I've heard an argument made that basically the big tech companies, to build these data centers, they have to forecast demand out years in advance. It's a massive capital commitment. They're not sure whether it will pay off. And CoreWeave is useful to them because you're taking the risk and then they will be able to use your capacity and sort of rent it out as opposed to having to make these big investments on their own. It's their ass if things go wrong.
That is a narrative. I don't think that actually tracks with the reality of the situation. The reality is the large hyperscalers are building as fast as they can. Google went out and released a press release where they're building $50 billion worth of infrastructure while they're still buying from everyone else they can. Microsoft is building internally and they're buying from lots of other players. I feel like that argument is model fitting. Somebody's got a preconceived notion of what this is going to look like and now they're reconstructing the facts on the ground to fit that model so that they can say, 'Look, I'm right.' But the reality is I look at it very differently. I look at the way that we built our competitive advantage over the hyperscalers and other neoclouds. We understood that this type of computing was going to be important and we built the infrastructure and the software to be able to serve it when the demand emerged, and we did it in a very risk-managed way. When I look at the future, when I think about the investments that go into building an AI factory and I think about how much money is being put into the data center versus how much money is being put into the compute that goes inside of the data center, I think about the data centers as being basically an option on being able to provide and be relevant for the delivery of compute into the future. We take our risk dollars as a company and we invest in the long poles. The long poles are really twofold: one is building the best software in the world, and the second one is having access to the data center capacity to be able to deliver compute when a wave of demand hits this market that requires you to deliver it. You can't just wake up and say, 'Hey, I want to deliver a gigawatt worth of infrastructure.' What you have to do is you have to start years in advance building that gigawatt of infrastructure so that you're in a position that when your customers say, 'Hey, I just produced a new way of using AI that's going to require a gigawatt worth of infrastructure,' you're able to serve it. We're going to have a tremendous portfolio of infrastructure that is going to be able to be deployed into the future. We're really excited about that. We think it's a wonderful way to go about building our business.
And that's the question about the bet. You're betting that AI is going to continue to be adopted at a wild rate.
That's not entirely accurate. What we are doing is we are making the majority of our investments by taking long-term contracts from creditworthy entities, using those contracts as a way of raising money to build the infrastructure where the demand and the credit and the capital has already been secured. Let's say 85% of our exposure is to deliver compute to investment grade or AI labs or other large consumers of compute. The other 15% is our exposure to long-term contracts to be able to do that exact thing in the future. That's the way I look at it. I think it's a much better way to think about how we're taking on risk, how we're dealing with leverage, and how we're positioning ourselves. If the market continues to grow, we're in a great position. If the market stabilizes in and around this, we're fine. If the market contracts, there's some new technology, then we will be left with some portion of that 15% that may have to wait for a few years before the market grows back into it. We are fine with that. People have talked about how the founders of this company look at the world with a different lens because we don't come from Silicon Valley. We come from the commodity space. We come from Wall Street. We think about option value. When we think about compute, we think about what is the option value associated with it. When we think about the data centers, we think about what is the option value to be able to build to be relevant in the future. That's the way we go about allocating our risks and securing the contracts that we have in place right now.
To speak to one thing here, you talked about if the market contracts. I think that we would love for that because it presents tremendous opportunity for us. You're in a position where there's going to be distressed assets. There's going to be consolidation possibilities. That's when opportunity really comes in. There's a lot of times where we sit there and say, 'Okay, we're looking for M&A, we're looking to invest in things, but the valuations don't make sense.' For Mike and I, we've made our careers on waiting for those opportunities and saying, 'These are the things that I want to buy when things don't necessarily go right for them.' That's really what excites us. One of our other founders last week, he got on the phone with me. He's like, 'I love this, Brian.' I'm like, 'What, Brian?' He's like, 'This is the one where you start like this. You're so focused on where are the opportunities? How do I go take things over?' I say it to some people every once in a while: I feel like when there's headwinds in the market, it's actually easier to do this job than when the tailwinds are blowing at 1,000 miles an hour.
But can I ask how have you set up the company to make sure that you're not the distressed asset when the contract...
Look at our construction of our customer contract portfolio. Last year, everybody talked about how customer concentration and exposure to Microsoft was a bad thing, but they have a better balance sheet than the US government. I'm not worried about them performing in their long-term obligations to us. That's basically the best possible position we can be in. We've been super thoughtful about the way that we choose which customers to work with and how we manage the credit exposure so that we're certain that the investments we make will be paid back. If you look at the people that are providing us the debt to do those projects, like Blackstone, they're some of the most sophisticated people in the world. For their underwriting committees to come in and say, 'Yes, I want to do this and I want to scale it up as aggressively as possible,' you're telling me you're going to pit some financial analyst against John Gray? I'm going to go with John Gray.
Maybe a second on just one of the fundamental building blocks of how we have expanded the way we have and how we use debt, because I think that's one of the misunderstood components of how you build or how we have built this company. It is really important to understand that the way we build the components is we go into the market. Let's use Microsoft because we've used them, but there's lots of other clients you could use and they're totally interchangeable. From the perspective of the structure, it's still the same. We go to them and we say, 'Hey, we've got access to this data center.' They say, 'We need compute.' We say, 'Okay, we're going to sign a contract.' They sign a contract for five years. We structure that contract in a way that we can go back out to the Blackstones of the world and we can borrow money from them to go ahead and build the infrastructure to deliver to Microsoft within the five years of the contracted period with Microsoft. We pay for the infrastructure, we pay for the opex, we pay for the interest, and we earn an enormous margin on the infrastructure. So yes, there is debt. We're not arguing that. We believe fundamentally when you build any type of infrastructure at this scale, debt is the correct way to go about doing it. The examples run through history. Whether you're talking about building a power plant, building a distribution grid for electricity, whether you're talking about the telephone, whether you're talking about the steam engine and railroads, you go throughout history, this is the tool that you use. We didn't invent anything new here. We just took a tried and true method and applied it to the specifics of depreciation associated with this asset, of the obsolescence curve associated with this asset, and made the contours so that it worked in an airtight manner so that guys like John Gray or Blackstone or BlackRock or any of the big lenders could look at it and say, 'I understand how they're going to underwrite this. I understand the risk in this.'
Understand that these guys are going to deliver compute to that balance sheet. They're going to get paid back and when they get paid back, we're going to get paid back. So, let's lend them the money.
And that's lost on the market. They think we're running around with this incredible capacity to take on risk. But that's a really low-risk approach. Matter of fact, it's way more low-risk than saying, 'Hey, we're going to do it on equity because we're saving our equity for the long poles that you've got to invest in.' That's where you want to put your bullets. You want to use the debt markets to deal with a depreciating asset. It's the way it's done. It's the way it's been done throughout history.
Yeah. By the way, it's great that we're able to have this conversation. This is what we want to do on the show is take this complex stuff, talk about what the reactions have been in public, speak with the principles, and actually get the story. So, thank you for talking it through with me. And on that note, let's continue. The argument I think that would be made is not that Microsoft isn't good for the money. The argument would be made that generative AI is still a developing category. It hasn't really shown the ability to turn consistent profit. And so the companies that are investing in a big way in it may one day wake up and say, you know, we can't, we don't really want to do that buildout. OpenAI for instance, let's just use them as an example, they have something like 1.4 trillion committed to spend on infrastructure. I think OpenAI might be the only ones that believe that they'll actually spend that 1.4 trillion and maybe they're investors. So what do you think about that risk that AI is because AI is new and not as predictable as you would have in a different category, you know, finance by debt, that therefore it is riskier even if the credit rating of a company like Microsoft is golden.
A couple things on OpenAI because they are the tip of the spear in many ways for artificial intelligence. They have a franchise that has 800 million monthly users of their product, which is fully one-tenth, one out of every 10 human beings on the planet logs on to OpenAI. Fastest growing tech product in history. I use it all the time. I am addicted to it and I don't even find it in a bad addiction way. It's an amazing product. I won't argue with that. So you've got this product that's out there, and then you have this $1.4 trillion, which I believe has been confirmed by everybody, but OpenAI, who would actually probably have issues with that number in terms of how much they're spending, when they're going to spend it, what are options, all those kind of things. And so, I just think it's a narrative shaping there. There's an incredible amount of people out there that are talking through how this is going to be done, when it's going to be done. And I don't think that they necessarily have all the correct information. That's number one. Number two is that you listen to both Brian and I talk about how we think about credit. We're pretty sophisticated how we think about credit. We've built our entire careers long before we started this company thinking about risk management in credit. OpenAI will be a percentage of our credit exposure just like Microsoft will be a percentage of our credit exposure. And the way that you manage credit against an unbelievable potential company, but a company that may not have the credit rating that is strong enough to support their aspirations or they may have to tone it down, is you just make them a limited percentage of your overarching business and you accept the risk on that while you mitigate the risks using credit from other companies. Companies like Meta that we signed a 14 billion contract with, like Microsoft, I mean just incredible companies. And so you just think of them as how much investment grade exposure am I going to take, how much non-investment grade exposure am I going to take, and what's the correct ratio and how am I going to mitigate that over time. And that's the way we look at it.
And what happens if one of these companies over time wants to walk away? Let's say Meta says, yeah actually artificial intelligence we can develop it much more efficiently, or Microsoft says, yeah AGI is actually a decade away, not three years away.
