Very happy to welcome Chase Lochmiller and my partner David Khan. And Chase is somebody who I've gotten to meet a couple times over the years. And the two things that have stuck out to me and why I'm excited to have him here. One, guy's brilliant. Good starting point. Two, he has a really interesting and different point of view on this market than most of the people in this room because it's anchored more around servers, steel, power, the raw infrastructure even beneath the GPU level that is powering this whole revolution. And our partner David is the perfect person to interview him because for better or for worse, David's simple observation that there's a heck of a lot of money going into the ground and not a lot of application revenue coming back to pay it off yet made him famous and/or notorious for AI's $600 billion question. So I will hand it off to David and Chase. Thank you.
Thanks Pat. So I'm excited to introduce this session. We're going to talk about the AI industrial revolution, underlining the word industrial. We're going to talk about steel, servers, and power. Jensen already talked to us about servers. So we're going to focus on steel and power. I think everybody here knows that there's a race going on between the most powerful companies in the world to build the biggest GPU clusters. We heard about 100,000 GPU clusters. We're now hearing about 300,000 GPU clusters and big tech companies are talking about million GPU clusters. And the person at the center of all of that is Chase. And so I'm excited to interview him and we're going to talk about what's going on behind that. Chase, right now you have I think 5,000 people in Abilene, Texas building a data center.
We do. Yeah. 5,000 people every day on site kind of working around the clock to, you know, accelerate AI infrastructure.
Maybe talk to us a little bit about, we think about, we understand Jensen's servers. We know that they go in a data center. Maybe some people here can visualize a data center or have been to a data center. But what does it take to start from land basically and build one of these AI factories?
Yeah, I think it's important to like one take a step back and just evaluate like what is the AI factory? What is the AI data center of the future and how does it differ from existing data centers? Because, you know, like a data center as a, you know, it's a bit of an overloaded term to be honest. Like calling, you know, a data center from 20 years ago that's serving, you know, a web page the same thing as, you know, a high performance AI data center of today that's, you know, managing a, you know, 100,000 plus GPU cluster, it's like, you know, calling a go-kart and like a, you know, Range Rover like both vehicles, right? They're both vehicles but like you probably wouldn't use them for the same things. So, you know, I think when you look at today's data center infrastructure and what's really required to scale AI, it is a very industrial process, right? The amount of energy that goes into it, you know, I know it's been talked about quite a bit but I think helping put it in context, like, you know, I think Northern Virginia many people would consider to be like the center of the world for data centers, right? That's like where most of the internet happens. You know, we've been building it for decades, right? So, what's like the capacity of Northern Virginia? It's about, you know, JL published a report end of 24. It was like roughly 4 and a half gigawatts of total capacity in Northern Virginia. Like, Crusoe is a 7-year-old startup. You know, we're building big AI infrastructure. We're currently working on about 2 gigawatts of, you know, capacity currently under construction. We have about 20 gigawatts in the pipeline. So you're talking about like, you know, it's just completely transforming both the energy landscape and the data center landscape to make AI a reality because of the amount of infrastructure required to build it. Now, like the bottlenecks of it, you know, they're constantly moving around, right? I mean, I think there was a period of time where it was impossible to get chips, you know, and that was the major bottleneck. I think a lot of that is, you know, eased up as, you know, production both at TSMC and Nvidia have sort of increased to, you know, make chips more available. But, you know, power, steel, data center space, that has become a major bottleneck and these are complex supply chains that, you know, are not easy to accelerate. So, you know, Crusoe's always kind of taken this very first principled vertically integrated approach to things. And so, you know, just to give you an example, we've stood up a lot of manufacturing ourselves domestically here in the US to help accelerate timelines. Like speed is a critical advantage for us. So, you know, we've got quotes for, you know, low voltage switchgear, right? So it's basically like, you know, where all, you know, it's sort of the electrical control room for the data center. And the quotes we were getting were 100 weeks. I was like, that's insane, you know, 100 weeks. I don't have 100 weeks. So we stood up a factory and we started making switchgear ourselves. Now we make it in 22 weeks. And just like finding elements in the supply chain where, you know, we could really help accelerate a lot of this infrastructure has been critical to our success in sort of building these very large scale projects, you know, faster than anybody's ever done it before.
How have you convinced big tech companies to work with you? When I think about Amazon, Google, Microsoft, I mean, these companies are data center companies. And for the last wave, the cloud wave, this was an internal core competency for these companies, right? They all built their own data centers. They have massive teams doing that. And yet you guys, this startup that seven years ago was building little containers on top of flare gas is now a partner to a lot of these companies to build their data centers. How do you make that argument?
Well, I think it's worth appreciating that while big tech is good at a lot of things, they're not inherently designed to move quickly. In fact, like structurally, they're designed to not move quickly, especially when there's like very big purchasing decisions. So, just like the approval process of like, you know, when I need to make a decision to buy something, it's like I just make the decision and then it happens that day. And, you know, if you look at like Microsoft, there's like 19 layers of approval that sort of go into all these processes. So, like, you know, just from the starting line, there's like a massive disadvantage for them to move quickly. Now they've all developed great programs and they're looking at this like, you know, a big data center, you know, 100 megawatt plus, would, you know, 3 to 5 year type programs. You know, our first phase of what we're doing in Abilene, we're hitting 300 days. Our second phase we're actually just over 200 days. So we're like massively accelerating the time to market of this infrastructure which is the major value prop to big tech companies and big AI labs that are interested in this type of scale and purpose-built infrastructure.
There are rumors that you guys are involved in Stargate. I don't know if you're allowed to comment on that, but what do you think about these mega scale data centers, not just at the level of big tech, but also nation states, right? Something Jensen talked about and talks about on his earnings calls is sovereign data centers and sovereign AI strategy. As we go from data centers in Northern Virginia that you and I can go visit to these massive projects, what does it look like? What is the real estate footprint? How should governments think about it? How do you see that evolving?
Yeah, I think there's two big verticals that are driving this. One is that, you know, because of the scale of energy needed for this infrastructure at scale, like, you know, you really do need to take an energy first approach and that's really what Crusoe's done, like bringing the demand for computing infrastructure to areas where we can access low-cost, clean, and abundant energy. That's really what brought us to West Texas. That's what brought us to Abilene, Texas, was that there was a massive overbuild of renewables, particularly wind, where people were getting negatively priced power on the wind power they were generating and selling, something like 30% of the time. And there was actually an abundance of curtailment, which means they could be producing power, but they actually just don't because they don't want to sell it at a negative price. And so, you know, that's sort of the opposite problem that AI has. AI needs a lot of energy. So it was kind of a natural fit to say, well, nobody's ever built a data center in Abilene, Texas, but if we did that, we would have this abundant clean energy to power it. So that's really what drove us to be there. I think, you know, the sovereign AI comment, I think, is a pretty interesting one. Because that's the other dimension that we're seeing a lot of geographic dispersion in the demand for this type of infrastructure. And that really has to do with like data locality. When people think about, you know, there's data that, you know, maybe owned by a nation state or, you know, can't leave the, you know, four walls of the kingdom, like, you know, it has to live there and they want to do something very special with that. So that's really catalyzed a lot of demand for investment in domestic AI strategies and I think people are looking at this from like, you know, geopolitical power shifts as well.
