I'm Ina Fried, chief technology correspondent at Axios. Thank you so much for joining us. We've almost made it through Davos week. This is my last interview of 14, but we saved the best for last. Please welcome CoreWeave CEO Michael Intrator.
Little logistics. Normally I have my stuff better, but 14th interview out of 14. It's your last interview, too.
It is. And this is it. So, whatever's left in my voice, you're getting it.
I was all for a champagne toast because I thought it might make things more fun, but you have a few more meetings.
They're going to get what they get, too. Yeah. You know, not everyone has heard of CoreWeave, but you guys have become massive in just a few years as kind of the place people go when they need more AI capacity. You've also become one of the biggest signals of just how intense this AI infrastructure buildout has become. When you look at your book of business, because you guys are obviously trying to decide how much to build when, how confident are you that this demand that we're seeing is durable, not just front-end loaded?
Yeah. So, it's a good question and one of the things that's important is CoreWeave is building an incredibly capital-intensive business, right? We build everything from the ground up, you know, some of the data centers that we use, call it like 15-20% of the actual physical data centers, all the way up through the computing, through the software layers that allow us to manage and extract as much value as you can from the NVIDIA chipset. And so when you think about building a business that is that capital-intensive as a small company, as an emerging company dealing with the largest competitors in the world, who are also sometimes great clients, which is an interesting dynamic, the way that we have gone about doing it is we are building our infrastructure in response to signed contracts with long-term demand. And so, by way of background, the three founders really come from a risk management perspective. And when we enter into these gargantuan contracts, we have an eye towards can this be underwritten by lenders, and that is the only way that we are willing to enter into a contract. So, for example, to make all that gobbledygook make sense is I'm not buying GPUs until Microsoft or Meta or Nvidia or one of the large trading companies comes and signs a contract with me. And so I'm leaning into the creditworthiness of the counterparty in order to insulate my company from what we believe is an up-to-the-right scenario, but certainly could encounter some air pockets. And when you're building a business and you're using long-term contracts, you insulate yourself from that. And so that's how we think about the creditworthiness of the company.
And that makes total sense. At the same time, I'm sure you are also looking, as everyone else is. What are the early indicators that maybe demand is slowing and what would you watch? What single metric would you say folks in here should be watching to say things might be overheated? Because one of them to be honest to me was when everyone started signing deals where they were becoming a customer and investing in the company. Like as a business journalist that sounds sort of circular, but you're assuring me that these deals are all solid and going to go forward. What would you look to as an indicator that things are slowing?
So, one of the things that's really interesting about CoreWeave is because of the position that we sit in, and I'll use a, you know, we're Switzerland in some ways, right? Like a lot of different really large important buyers use us. So, we're getting information flow about the demand signals from the entire industry. Another place that's a really good place to look is the growth of inference. And I've always said, you know, inference is the monetization of the investment in AI. And as long as you see this incredible violent ramp in inference occurring, it's telling you that there are people that are using the products in order to extract value.
Because training is the risk capital. Inference is the business humming for your customers, not for you.
Yeah, I was going to say that like training is risk capital in the space. For us, the risk capital is the long pole in how you go ahead and build and deliver this. So, we really have two long poles that we focus on. The first one is the investment in the software, right? We need to build the best solution to deliver infrastructure to the world. And so we invest in our human resources, in our software, in our ability to deliver the infrastructure. And the second one is the physical data centers themselves because that has a long lead time associated with it. And so that's where our risk capital goes. It doesn't go into the GPUs. GPUs, which are the overwhelming majority of the cost of compute, that's not a risk capital destination for us.
And I also, you know, I do want to turn to the upside because there is an AI boom happening. I don't want to just focus on what if things turn down. But I know for your investors and for people watching the space that is a thing to watch. What would overbuild look like if it did show up? Would it be lower utilization, pricing pressure, shorter contracts, like what are some of the signs? And then we'll turn to the fun stuff.
Look, it's important to understand what we can understand from the market and what we have to extrapolate, right? And so what do we know about the market right now? The market is pinned, right? The collective global capacity to deliver infrastructure cannot keep up with demand. And there's been a lot of discussion about the powered shell as being the limiting factor. There's been discussions about how energy is going to become the limiting factor.
Even electricians, labor. So let's go broader than that. Electricians are quite dear. But labor in general.
Labor in general. But also like look, you know, I mean, look what's gone on within, you know, some of the more sophisticated memory components. Like, you look at those charts and they're like earth-shattering.
By the way, if you can't get memory to upgrade your phone or computer, you got it all.
Yeah. So, and that's constrained by the fabs, right? Like, and so like there's the market itself. In a normal functioning market, you can look at supply and demand and you can model it out and you can kind of come to an understanding of where things are. And in a market that's pinned, you're able to follow it up to the capacity of the market to deliver and you can get a bunch of smart analysts, you can get a bunch of AI and kind of work out where the choke points are going to be. But really at the end of the day you are building for the demand signal and that leaves us in a position of trying to accelerate our ability to deliver this infrastructure because the largest, most sophisticated buyers, the people that are touching a global universe of clients are telling you we need more.
