Chase Lochmiller, CEO of Crusoe, in Conversation With Claudine Emeott | Salesforce Ventures

AI infrastructure is this entirely new category of infrastructure. And you know, with this new category, there's actually a tremendous amount of new things that are having to be built for sort of the first time at certain scales.

Thanks so much for having me. So cozy here by the fire.

I think people under appreciate the importance of infrastructure in the quant finance space. And you know, there's been a lot of breakthrough innovations that take place there. It's all very private because, you know, people are always scared about like other competitors, you know, getting a hold of their alpha. But you know, everything from, you know, the network and telcom infrastructure, you know, the high-frequency trading space was famous for, you know, seeking out fastest path, you know, fiber connectivity between Chicago and New York and then building microwave towers to transmit signal faster. And then on the actual predictive analytic side and the predictive modeling side, you know, the quant finance space is really about trying to find prices that in the future will be different and being able to be accurate about those predictions that they're making in a wholly automated and quantitative basis. And that led to tremendous investments in machine learning and artificial intelligence research and a lot of time series modeling and forecasting predictions. And you know, historically the quant finance space did a lot of feature engineering. So, like I had a background in physics. A lot of these features sort of were representative of economic relationships between security A and security B or, you know, the index and you know, the components that make up the index. And then, you know, over the course of time as deep learning became popularized and you sort of had this magic moment with, you know, the there was this big Google paper around like being able to identify cats, you know, from convolutional neural networks in videos. And that was like a seminal moment for the space.

It was a breakthrough. Yeah, yeah, yeah. And you know, with this whole AlexNet paper, you know, that really led to, you know, the quant finance space actually adopting a lot of AI and machine learning techniques that were using deep learning. We started to sort of experiment with things like recurrent neural networks and some of the time series variations on using deep learning. And in that process, you know, I started, you know, training a lot of models instead of on CPUs and classical machine learning techniques. So, we shifted workloads to using GPUs and training these, you know, these big deep learning models. And you know, that experience for me like really made me appreciate the importance of the computing infrastructure, the networking stack, the whole storage ecosystem. And simultaneously I was kind of like working on some side projects, but you know, that were in NLP and in computer vision. And I really had this high conviction that AI and intelligence broadly was going to be embedded in every aspect of the economy in the future. I was like, there's nothing that won't benefit from more, you know, AI embedded in that application and technology to drive greater efficiencies and here we are. Here we are.

Probably 2015.

Yeah, it was a while ago.

Well before. Yeah, I mean it was before the transformer. I mean the transformer was in 2017, but yeah, I mean I think the quant finance space is like very much constantly behind the scenes like in their secretive kind of way always kind of staying at the forefront of infrastructure and you know, it may not seem obvious to people, you know, outside of my circle, but you know, it felt obvious to me that like AI, you know, was going to become an important aspect and was influenced by that research I was doing in the quant finance field.

Yeah, I guess I'll start the answer by sort of stating that AI infrastructure is this entirely new category of infrastructure. And you know, with this new category there's actually a tremendous amount of new things that are having to be built for sort of the first time at certain scales. And so it sort of looks like historical things like yeah, sure there are data centers involved, the category of data center is an entirely new category of data center to power these large-scale AI workloads because of the density, because of the power consumption, because of the cooling architecture, because of the underlying networking fabric that ties these GPUs together. All of that has to be rethought from first principles. And so Crusoe as a business is a vertically integrated AI infrastructure business. And we are focused on both design development of data centers, which includes, you know, a lot of the powering of the data centers themselves, manufacturing modular electrical components to assemble these data centers faster, you know, the cooling architectures, etc. And then we also have a Neo Cloud business that's focused on, you know, standing up these very large clusters of GPUs to enable AI innovators to do their life's work. And you know, that basically spans from, you know, hardware to software. And from data center development to, you know, serving clusters of GPUs or actually serving tokens to folks that are, you know, running large-scale AI applications.

Yeah, so you know, first on the vertical integration piece, I think the reason that's really important is that because so many things are happening for the first time, if you're vertically integrated between, you know, the software, the hardware, and the construction, and the development, you can actually you're in control of your own destiny. You can, you know, the thing that needs to be built, you have the know-how to build it.

Down to electrical switchgear, exactly. It's very, you know, individual components like are creating bottlenecks in certain ways to, you know, scaling AI and scaling intelligence. And you know, power is one of the key bottlenecks. And this is one of the original thesis of Crusoe was that at scale, you know, energy was going to become the bottleneck for all this stuff. And we felt like taking an energy first approach to building computer infrastructure was always like core to our philosophy. And you know, historically the data center market has had these hubs, right? These like markets like Northern Virginia where you have a lot of the data center infrastructure that's been built over the last four decades to support the internet. And we didn't want to be in the, you know, business of like let's build the next data center in Northern Virginia. We said, 'Look, Northern Virginia's actually quite power constrained. Why don't we think about, you know, for AI, it's less about being, you know, it's less about being co-located with the other data centers in Northern Virginia and have the latency, you know, to those other data centers. It's more location agnostic. So, let's think about where we can build that infrastructure where we can access low-cost, clean, abundant energy.' And that led us to markets like Abilene, Texas where, you know, there's a lot of land, it's easier to build, it's one of the best places in the United States to build renewables, there's abundant, you know, solar, wind, and also gas to basically, you know, power any workload that, you know, we need to run out there.

I think we just like had this like recognition that moving data was a lot easier than moving power.

That's the crux of the business. It's like go to where you can access low-cost, abundant power and where you can actually build things, and moving the data in and out of there is actually, you know, that's a very much easier to solve problem than trying to move around gigawatts of power to, you know, very congested or constrained markets where it's very difficult to build.

But I may be biased.