Hi everybody. Welcome back to New York City. We're here at the New York Stock Exchange overlooking the trading floor. My name is Dave Vellante and this is the NYSE Wired + theCUBE CXO Series. We're here with Chase Lochmiller, who's the co-founder and CEO of Crusoe. Chase, good to see you. Thanks for coming on theCUBE.
Thanks for having me, Dave.
So when did you start the company and why? What was the founding premise?
Yeah, so I started Crusoe in 2018, so about six and a half years ago, actually with a friend of mine from high school. And the premise was really to help accelerate computing-led technologies and computing-led innovations like AI in a capacity that could be sustainably powered. So really at this convergence of AI and energy in terms of how do we unlock all of these incredible innovative potentials without having to bear large-scale environmental consequences.
Did you have early visibility on gen AI? Was it a gut feel?
So maybe taking a step back, I had spent the early part of my career as a quant portfolio manager. And in that role, we had adopted a lot of deep learning techniques, and I went to Stanford for grad school and focused on AI in grad school. And I'd seen all of this incredible innovation taking place and all of these massive breakthroughs in terms of performance for computer vision and natural language processing that was all being driven by these next-generation non-linear modeling techniques, these multi-layer neural networks, and computing was a big driver of making this possible. The two things that really made this possible were we finally had enough data to train these types of models that had been around for decades, and we finally had enough computing power and we could parallelize it in a way that actually made it computationally feasible to converge on something that worked. And so at that point, I really had a high conviction that AI was going to be transformative for every aspect of the human experience. And we had this seminal moment in the breakthrough with ChatGPT being launched in November, 2022. I probably would've predicted that it would've happened sooner personally.
Yeah, I think anyone building in and around the space, when you see the performance gains we're seeing from a lot of deep learning techniques and as people were iterating on them, it really felt like magic, right? You had this superhuman capability oftentimes in a narrowly scoped application, but you could extrapolate in your mind how we would get there to superhuman performance across numerous applications. And that's kind of what we're seeing today with a bunch of breakthroughs.
So when you saw ChatGPT, it was like, 'Okay, finally.' Is that right and is that fair?
Sure. I had my own sense of awe the first time I used it and I was like, 'Wow, we're there. This is happening. We're this very close to passes the Turing test,' and at this point it does pass the Turing test, and this is the seminal moment and really the breakthrough in AI that we've been waiting for.
And the reason I ask that, not to put you on the spot, but I have a friend who's deep into AI like you were at the time, and he's been at it for 20 years, whether it's AI and autonomous driving and really, really understands it well, and I had dinner with him right after ChatGPT was launched. So he says, 'You ever hear of ELIZA?' And I'm like, 'No, I've never heard of ELIZA.' Well, it was actually an IBM mainframe program that you could interact with in natural language back in, I think the '60s or early '70s. And people were like, 'Wow.' And they were using it to help people with mental disorders just feel like they could speak to someone and have an interaction. And so now it's like fast-forward, wow, we're here in a way that actually has the computing power. It can be scaled broadly and I guess inexpensively enough, but the problem is we've got to cool all this stuff and we've got to make it sustainable.
And that was your founding premise. And why is that good business?
Well, look, I kind of consider myself to be a libertarian environmentalist. I care deeply about the planet, but we have to make these things work on a standalone and economic basis. And no matter how you look at it, we're going to need to bring on net new generation to power AI. The energy demands are just so massive. But what leaves me to be very, very optimistic at this moment in time is we're finally at this point where cleaner energy generation is marginally cheaper than fossil-based generation. So you're seeing the lowest cost sources of energy are wind, solar, geothermal in certain locations, hydro in certain locations, and nuclear, if we could get out of our own way from a regulatory compliance standpoint. But that's one of the things that leaves me very, very optimistic about being able to power AI in a sustainably based capacity. And to me, it was really important to approach it this way because if we had massive breakthroughs in terms of realizing AGI, but it accelerated a climate crisis to the extent that we made 95% of the species on earth go extinct, did we win or did we lose, right? We really want a very abundant future that's rich in technology, but also rich in nature as well.
So you're in the technology community, you're a builder. I've been in this business for a long, long time, and it seems like the narrative around... not seems like. The narrative around tech has shifted. When I was a kid, it was very positive attitudes toward tech. Grew up in Massachusetts, and 128 was the sort pre-Silicon Valley. And we almost put it on a pedestal, sort of glorify tech. And today it's, oh, big tech lash and social media's bad. And granted, there's plenty of things that are negative about big tech, but what often gets lost in the conversation is all the good that tech has done. I bring that up because, of course, the media goes right after the energy consumption, whether it was same with crypto, we heard that as a big negative. Now we're hearing that with AI. I bring it up because I was listening to a report around Eric Schmidt this morning. I don't know if you heard this, where his premise was it's a mistake to try to constrict AI based upon the energy argument. Let it go. AI will help us figure out a way to reduce the energy consumption, maybe by taking carbon out of the atmosphere, maybe it'll come up with some way of doing that. Do you have any thoughts on that?
And I'd love to hear your perspective.
Yeah. I often describe this as the energy paradox with AI. There's all these gripes about AI's energy usage, and the solution to AI's energy usage is AI. AI is the tool that we've been looking for that's going to be able to drive the scientific breakthroughs that we need to achieve low-cost sustainably powered futures. This is applied across technologies like fusion and advancing fusion developments and a lot of particle physics and plasma physics simulations. And with these very complex geometries, can really only be done with large-scale accelerated computing infrastructure. Things like advanced battery chemistries. We've really only scratched the surface in terms of the material science and the battery chemistry space that's possible and cost-effective, and what do these things mean at scale? There's certain chemistries that have become popular, but the search space is far wider than we've ever really fundamentally explored. We had a big partnership with SES AI recently, it's a publicly traded battery technology company that we're working on supplying the AI compute infrastructure to support the development of this foundational model for inorganic chemistry in something called DFT, diffusion function theory, which is sort of the underlying scientific theory that describes these battery chemistries. And again, the search space is massive. We haven't been able to fully explore with traditional classical techniques, but AI really is the solution there. I'll give you one more example. I met a company recently called Orbital Materials. It was a team that spun out from DeepMind, and what they're focused on is actually using AI to develop these custom-engineered materials for direct air capture systems that are purpose-built for specific thermal loads. So what I mean by that is that one of the key inputs to a direct air capture system is heat. The materials we often use to basically soak up that carbon are things that are oftentimes naturally occurring, that can be carbon sinks, things like limestone. Just because it naturally occurs in nature doesn't mean it's the best and most economic way to soak carbon out from the atmosphere. In the data center world, one of our key exhaust products is heat, but it's low-temperature heat. It's oftentimes not that useful for a lot of different applications. But one of the things that we were talking about is, could we engineer a purpose-built, custom-engineered material that was designed to basically absorb carbon with the thermal load that's the exhaust of our data centers? What about that? You could have a renewably powered data center that's taking the waste heat from that data center to suck up carbon from the atmosphere. You'd have a carbon-negative data center. So to me, Eric Schmidt is spot-on, and I'm so optimistic about AI being the mechanism to invent the solutions needed to power the future sustainably.
Well, these are such novel ideas you're talking about. Of course, they're created by humans. Well, imagine when you put AI in there. But now that doesn't mean that we shouldn't approach AI in an energy-efficient manner. And that's what you're doing. You're scaling AI, you're building energy-efficient data centers and AI ecosystems. Tell us more about Crusoe.
Sure. So Crusoe is a vertically integrated AI infrastructure business. We try to take an energy-first approach where we build data centers in locations that have low-cost, clean and abundant energy. The traditional data center model is build data centers in the networking hubs like Northern Virginia, and then everybody wants to be in that data center corridor. We're building in less traditional data center markets, areas like West Texas where you have very abundant wind and solar power that's oftentimes curtailed or negatively priced. And actually, the biggest problem is they don't have enough load. Other markets like Iceland where there's rich geothermal and hydropower that is more than abundant enough to power the needs of the 400,000 people that live in Iceland, they can turn on a lot more generation. They just have no natural need for it. It's all carbon-free power. We can build large-scale GPU clusters there to train these big large language models and power them in a sustainable basis that's also very cost-effective. So we really have two core products. The first is actually just the data center layer where we will build these large-scale AI data centers. We typically lease them to very big companies like the hyperscalers, and then we have a layer above that that's a managed AI cloud platform, that's a serverless architecture for AI startups and bigger enterprises that are looking to do something in AI, whether it's training, fine-tuning or inference, we can really help support them with serverless compute infrastructure.
Interesting business model because there are companies that are basically alternative to cloud, to hyperscalers that sell from the GPU clouds. You do that, but you also build data centers, which is a different business. It's related, but it's novel in the sense that most companies pick one, but it seems, A, synergistic, but also, as we were talking about last night, it throws off cash, it's lucrative, it's long-term contracts, and so it brings stability. I don't know how the economics compare long-term with the other business, but nonetheless, it sounds like it's pretty viable.
Yeah, look, I think what we're seeing in the data center design space is a completely new paradigm. What a data center was 20 years ago compared to an AI data center today, it's a completely different design from a cooling, from a power density, from a rack density, from how much space it's required. And as Jensen Huang, the CEO of NVIDIA likes to say is that the data center is the new unit of compute. So from my perspective, in order to be at the cutting edge, building world-class AI compute infrastructure, you don't just have to be good at putting a cluster together. I think having a lot of informed expertise around the actual data center itself is a critical component to being successful there. And so we've tried to do both from both the data center side as well as the infrastructure software layer, the high-performance software-defined networking architecture that enables high-performance AI workloads to train in these multi-node or multi-GPU capacities.
Tell me more about your software IP.
Sure. So we've done a lot to create a very high-performance virtualization stack that enables innovators to basically engage with the AI compute infrastructure to drive forward their specific use case, whether it's training a new model, running a big inference workload. But we've tried to engineer things with AI as the use case in mind. So there's a bunch of different elements that really make us special. A lot of low-level kernel optimizations that we've made that make things like startup times way faster for bringing up these big clusters. But the cluster really is what a lot of people are focused on. So the networking stack becomes incredibly important. And building this high-performance software-defined network that we've done at Crusoe that enables multi-tenancy across the network has been a critical aspect because we will oftentimes have a single cluster of GPUs that's in what's called a rail-optimized architecture, which means you can share information directly from one GPU directly into memory with another GPU. You're not sort of bypassing any... It's not like data's going from GPU memory to the CPU or to the core networking, the NIC across the ethernet fabric. You're actually going across an InfiniBand fabric in a high-performance, high-bandwidth capacity to be able to run these multi-node workloads.
So isn't that a function of the large SRAM that NVIDIA has? Are you adding IP on top of that versus say, a chiplet architecture, which is going to be asynchronous to all the XPUs? My understanding is NVIDIA with that large memory, shared memory has direct access for all the GPUs. Are you adding IP on top of that or just optimizing it so long?
Correct. So what you're talking about is actually what's on the chip. So the on-chip HBM, an H100 has 80 gigs of HBM and these models and the data sets are far bigger than that. So you actually need not only multiple GPUs, there's typically eight GPUs on a server, but actually multiple nodes.
