All right. Welcome everyone. Tolo Business School's Distinguished speaker series with Mike Intrator, CEO and co-founder of CoreWeave in Commerce with DMCI. Early in his tenure, DMCI focused on two global trends that he determined Columbia Business School must lead on, and those are AI and climate change for the simple reason that those two trends will affect every business everywhere in the many decades ahead. Today's guest, Mike Intrator, is uniquely positioned to discuss both of those topics. I first met Mike over 20 years ago, and we were both working at the intersection of finance and climate change. Mike at Columbia University graduate worked part-time at Columbia Business School while completing his degree, and he wrote his senior thesis on using nuclear power to address climate change. He also helped create, after graduation, one of the first climate investment funds in the world, which finance mitigation projects in China and other countries where his skills and infrastructure development are pretty useful today. Mike co-founded CoreWeave in 2017. The company went public in March and the largest AI related listing. Since the IPO, the stock has tripled, and today its market cap is $66 billion. Throughout his career, Mike has clearly been ahead of the curve on both climate change and AI. Please join me in welcoming Mike Intrator to Columbia Business School. Thank you.
Thank you, Bruce. Thank you all for coming. Thank you, Mike, for coming. So we'll talk about CoreWeave, we'll talk a little bit about AI, we'll talk about climate and energy. But I want to start by asking you to tell us a little bit about your company. You've built a company that now has positioned itself as a huge provider of AI infrastructure, competing with some of these hyperscalers. You are the new kid on the block that now is like a huge company in its own right. You can tell us a little bit about that journey.
Sure. First of all, Bruce, thank you for the introduction. That was great. As Bruce said, I've known Bruce for a long time. We spent a lot of time in the early days of our careers working through some really complicated and interesting and important questions around climate change. It has been truly my privilege. I think of it as a privilege to have spent and will continue to spend my career working on what I believe to be the two issues of our time, which is climate change and artificial intelligence. I think they're seismic in nature and have the ability to really bend the trajectory of the world. So it has been a privilege to work on these two things. I feel very fortunate that I've managed to put myself with the help of an amazing group of people in a position to work at a company and to build a company like CoreWeave that has done some truly unbelievable things. It's been a lot of fun. I'll talk a little bit about the company now. CoreWeave, the way we think about it is it's the AI hyperscaler, it's the force multiplier for people working on artificial intelligence. What we have built is the first AI native cloud. When we started the company, we really started the company in crypto. We were running Hudson Bridge Asset Management, which was a hedge fund that was a systematic algorithmic fundamentally driven natural gas fund. Basically all those big words mean is we were scraping the internet for big data around gas, natural gas flows in the US pipe system, and looking at consumption and weather patterns to drive signals for buy and sell. So that's a little bit of the tech piece. But when you build an algorithmic system, one of the things that happens is you build it and then it's built and you run it and test new things and watch how it works and make sure it doesn't blow up the world. You are basically monitoring the system, which gives you a lot of time to think about other things. We started to think about crypto. Crypto is fascinating. We're curious and nerdy. So we got interested in crypto and then got interested in the security layer in particular. We looked at Bitcoin and thought it was a terrible idea from a security layer perspective because there was some brilliant engineer in China that had built the ASIC machine, and I just assumed they built it and they're probably better at running it than I am. So I didn't really see that I brought a lot to the table there. But the GPUs, those were really interesting because GPUs you can do a lot of different things with. We knew that from the start. The lens that I look at the world with was built in the early days, grown up with Bruce and Jack and all these other guys that focused on this stuff. My lens was incredibly quantitative in nature, hyper-rational and focused on the concept of optionality and flexibility. Flexibility is where things get interesting. The GPU had all of this ability to reorient itself, and we had no idea what we were doing when we started the business. That was the beginning of Atlantic Crypto, which ultimately became CoreWeave that everybody knows. The mission of the company is to rethink the way that the cloud needs to work for a new generation of computing. It stands to reason that when three of you will come out of this auditorium, hook up, build a company, and become the third largest consumer of computational power in the world in six months, the requirements of the cloud need to be different. The cloud went through a very fundamental change that people to varying degrees understand or don't, which was moving from sequential computing to parallelized computing. It was the genesis of what we went about building. We built a company that does three things. The first was to build a company that had a beautiful, elegant technical solution to a problem. That problem was how do you serve compute to the world that is going to require size and scale of parallelized computing that defies human imagination. That was the first piece. It's very hard to say that you have an advantage over Amazon, Microsoft, and Google. There is an audacity, a hubris, perhaps insanity of trying to take on that part of the market. But we did. We think about we, it's like myself and Brian and Brandon. Everything is about risk and exposure. You had the confluence of this new type of cloud computing with this new world changing technology that required an entirely different approach to infrastructure and an entirely different approach to finance. It was only because of the confluence of those three or four enormous moving pieces happening simultaneously that allowed for the dislocation that occurred. It couldn't be done otherwise. We saw the abject dominance of the hyperscalers within the cloud space through the CPU era. There are reasons for that. Once they are able to achieve a systemically lower cost of capital, we have seen those cloud computing giants crush competition based on that one sole advantage. It has nothing to do with the quality, but that only works in a world that moves incrementally. In a world that moves in step functions, that advantage doesn't hold. If we think about it through history, there are all these times where things just change. When things change in a large enough way, there is the opportunity for disruption. We went from sequential computing, think of that as the cloud that supports data lakes, websites, and your mom keeps her photos on it. It is a beautiful, fantastic solution to that problem. But when you move from that environment to parallelized computing, you've changed some of the fundamental assumptions in what the cloud has to do and how it does it. It turns out that the things that made the legacy cloud great at what it did actually are negatives in serving the new type of computing. The hypervisor, which is the piece of software that does an incredible job of the last mile of computing, if something breaks when you're using a CPU based cloud, you never know it because it just goes out and grabs a new CPU, slaps it in there, and you would never even know that anything broke. But that doesn't work in the CoreWeave world.
Can I stop you for a second? Because I think it would be interesting for you to visualize for the crowd when you build out a new data center that in some sense can be configured to work as one computer. Tell this crowd, how big is that computer in physical size, in needs? Because these things, I've seen some of them, they're daunting. You actually specialize in that field. So tell us a little bit about that.
So it's like the death star, but in a good way. Initially it was a single death star. Now what's happening is the entire space is configuring so that you can bring in multiple facilities of this enormous scale and have them work in concert. It started out with a cluster of a few thousand GPUs, think of that as a couple computing racks. Then it turned into football fields, and then it turned into multiple football fields. It is as wild as it sounds. The scale of these things is amazing.
What Mike is describing is a football field or four football fields that are connected when they're actually working, so from the point of view of the user, they're one computer. You have hundreds of thousands of GPUs that are themselves parallel machines, each doing hundreds of thousands of parallel calculations simultaneously, all working on the same problem. That is incredible. To go back to your crypto origins, crypto mining was in some sense a GPU exercise. You found yourself with a much more pure play infrastructure that was GPU based rather than the hyperscalers that may have had 10% GPU, 90% CPUs. You mentioned that you had different architecture in your data center that when you made it 10, a hundred, a thousand times bigger, that architecture was actually much more efficient for the pure play AI.
The software that we built allows you to understand what's going on in this computer that takes up multiple football fields in a way that the hyperscalers just couldn't. With hyperscalers, you'd have a problem and you'd have to send them a what just happened, and two weeks later they'd come back with it. The way we built it, you would get real time feedback to your dashboard and know what broke or why it broke, or if it was going to break, can we coordinate it off before it breaks? The problem is when you run a data center as a computer, an AI factory as Jensen will speak to it, when one GPU breaks or one connection flaps, the entire data center goes down. As you guys have all been watching the numbers, the cost of these things is astronomical. Losing an hour on one of these machines is losing an hour on a $40 billion piece of machinery. That's expensive.
Do you think that this sort of mode that you had initially is a defensible mode when you think about the next three to five years and the hyperscalers and what they're doing?
I think of the moat as I think of most things in this space, it evolves. When we were launching the company, the moat is we have a better piece of cloud. I believe fundamentally that all of the neo clouds and all of the hyperscalers will ultimately wind up building the cloud that they use to serve AI in an architecture that is similar to the architecture we built. To give you an idea of how unique the architecture is, our relationship with Jensen began when we went over there and sat in their spaceship office. They brought in junior level people, and they just kept bringing in more and more senior level people. We topped out at Jay Purry, who is the COO, and Ian Buck, who is one of the guys who kind of wrote Cuda, the CTO. We went home and we were just like, that was weird. A month goes by, we didn't hear anything, and then we get a call from our representative, and he says, Jensen wants you on the phone in 15 minutes. That's a perfectly Jensen thing. No time to prepare. You're just standing there. It's like your nightmare of waking up naked in the middle of a classroom. He doesn't want any packaging. 15 minutes we get on the phone with him, and it was supposed to be a half hour call, and he spent 45 minutes bearing down on Brian on the architecture. Why are you building it this way? How does that benefit my clients? How is that going to make the infrastructure work better? Why didn't you go this way? Why didn't you do that? It was unbelievable. His mastery of the concepts was incredible. Then he just turned to me, and I did not say a word in the technical discussion. There was no reason to expose myself to that. He just turned to me and for 45 minutes just ripped into me. How are you gonna raise the money? Why do you think you can finance this? How do you manage credit for startups? How do you do this? How do you do that? Are you gonna be able to expand? How are you gonna get the data centers to give you capacity when you have no credit? That was two and a half years ago. I suffer from time blindness, so things have been moving pretty fast. The way that the call ended, there were four or five of us actually on the call, the sales guy who I'm pretty sure was hiding under his desk the entire time, myself and Brian, one of the three founders, Jay Purry, and Jensen. At the end of that hour and a half call, he just turned to Jay and said, tell Vishal, we're gonna invest in these guys. These guys are building the cloud the way I need it to be built. This is the vision of how it's supposed to work. The technical differentiation was enormous. That was the first place we started. The physical ability to deliver infrastructure, this stuff is super physical. You gotta build a data center, you gotta deal with the cooling, you gotta deal with an untold amount of power moving through these systems. To build an organization that's capable of doing that at velocity and at scale. Those are two concepts that I hold the company to at all times. Scale decommoditizes compute. Getting another GPU doesn't help anybody. Building a cluster that's large enough to train the most important models in the world is no longer a commodity, that is a solution. The third area that we focus on is finance. We have been the tip of the spear thinking about how you go about building a company like this. How do you access the capital markets? I always joke, when we started this business, you go out to Silicon Valley and you say, I'm gonna do this, this, and this. First they look at you like you're crazy because who takes on the hyperscalers? Then they look at you like, even if it was a good idea, we've been investing in software, and that's not really capital intensive, and that's how we've made the last trillion dollars. So we're not investing in you. That was disheartening. But to build infrastructure, whether it's the railroads, the power grid, power generation, or fuel exploration, at the end of the day, it doesn't happen in the equity markets. We understood that. Our project finance DNA understood that. We looked at the way we plugged into both markets. West coast, let's go change the world. Finally got some traction there. And East Coast capital, which I always say has only one rule, which is, do not lose my fucking money. That's it. That's the only rule that they play by. That is the scientific term. There may be more eloquent ways, but there's certainly not more accurate ways. On the finance side, we have been pushing that forward. Even through the IPO, which we IPO'd three days before liberation day into the tariffs under the direct advice from the investment bankers, you cannot do this, you cannot do this, you cannot do this. To which I responded, that's cool. Why don't you guys step out, you got 20 minutes. Come back in with a better answer. Because we're going public tomorrow. I believed fundamentally that the bankers worrying about their franchise, worrying about the capital, none of that mattered. What mattered was us becoming a public company, because this company is the definition of what needs to be in the public market. It needs to be able to access the most inexpensive, largest, deepest pools of capital to build infrastructure on a planetary scale. They didn't understand that.
Let me ask you a question about that. We are now finishing a three to five year period of massive build out. You and I were chatting a little bit before we started, we're talking about training models, inference, and how these things are actually changing. What do you see over the next three to five years?
The most important trend, and I would argue maybe the most important trend right now in the world, is the adoption of inference. Inference is the monetization of artificial intelligence. It's the translation of all of this investment into economic return. It is efficiency, it's deflationary. It is where I go to understand that the investments that are flowing into this market are not insane. When you see companies using the models to solve real economic problems that will generate massive productivity gains, that's why this works. If you don't see inference, that's bad. If you do see inference and you see it accelerating and broadening in its use cases, that is the primary indication of where artificial intelligence will be over the next five to seven years.
Also becoming more cost and energy efficient.
Yes, which is important. The strides that have been made, we think about sustainability in two ways. Sustainability is near and dear to my heart. It's where I started my career. I thought it was important then, I think it's important now. We are seeing unbelievable improvements in the efficiency from the technology from a computational power perspective. It is faster, it's more energy efficient. The cooling has moved from air to liquid, which has caused a 60% decrease in energy. The strides that are being made are phenomenal. Yet the total energy being consumed is obviously growing. We think about that in as holistic a way as possible because this infrastructure, the data centers that we're building, the energy that's being provided to them, that's not a one term or two term presidential problem. This is making investments in infrastructure that will survive for 20, 40 years. You need to really think about that from an environmental perspective, a sustainability perspective, and a financial viability perspective. As the issues of sustainability become more important over time or less important over time, that will impact the viability of different pieces of the very expensive infrastructure building. We just announced in the UK another billion and a half pound investment into data centers up in Scotland, which is a hundred percent renewable. We are definitely incredibly focused on it to the extent that the grids can handle it.
Can you tell us a little bit about your geographic play?
