Chase Lochmiller (Crusoe) on Powering the AI Revolution (EP.506)

Happy to be back. Yeah, it's been a little while.

Yeah. So, Crusoe was started about six years ago. Our core mission as a business is aligning the future of computing with the future of the climate. So we build large-scale digital infrastructure and computing infrastructure to tackle the most energy-intensive computing problems. Chief among those is Bitcoin mining and digital currency mining. We've built a fairly large business primarily powered off of waste methane that's produced from oil and gas production where currently it's being wasted or burned off as an associated byproduct of oil production. And we're able to sort of collocate data centers on site with that waste gas production. Instead of it being flared or vented into the atmosphere, creating a big greenhouse gas emission footprint, we're able to actually capture it on site and put it to some beneficial use to produce power and power that digital infrastructure. We've also sort of been very focused on energy-intensive computing applications outside of the crypto ecosystem as well and have been very focused on building an accelerated computing cloud platform that's focused on AI and other high-performance computing workloads. So that's been a very exciting area of growth for the business over the last 24 months as things have evolved pretty rapidly in the AI landscape.

I think part of our mission, this whole notion of aligning the future of computing with the future of the climate, we really try to go as far upstream as possible to the source of energy production. And when you look at sort of the demand profile for computing as an entire asset class and consumer of power, it's one of the fastest growing power consumers in the world. Today, global data centers consume somewhere around 2% of global power consumption. That is rapidly going to be rising above 10% over the next decade with AI playing a major role in that and things like digital currencies playing a big role as well. From that perspective, really our philosophy is that it's not going to happen on accident in terms of like how that infrastructure gets powered. If you just kind of leave things to play out as they may, could result in a very negative outcome from a mission standpoint and a global sustainability perspective. Our goal is to really be intentional and thoughtful in terms of how we partner with upstream energy producers to try to drive down both the environmental impact and the cost of this future digital infrastructure.

So talking through some of the issues that we can help resolve as a computing provider, I think when you look at the energy landscape, there's tremendous waste out there. Our first big business which was focused on mitigating methane emissions from flaring. Flaring is this global problem that persists in the oil field where you have about 14 billion cubic feet a day of gas that is being burned off and completely wasted. Putting that in perspective, you could power about 65 gigawatts worth of power which is call it six times the total power footprint of the entire Bitcoin network. So it's a really really meaningful amount of waste energy that's out there causing harm and not benefiting anyone in a positive capacity. And by sort of partnering directly with people that you know are engaged in flaring and wasting that gas, we can provide a value proposition in that we can bring a revenue stream to an otherwise wasted asset as well as drive down the overall emissions footprint and drive down the cost for computing infrastructure that has energy as its largest operating expense.

Another big area of like partnership for us is actually in the renewable energy space. When you look at the way renewable energy gets produced and consumed, the best places to build renewable energy assets are not necessarily in the same places where people are actually demanding and consuming power and not necessarily producing power at the right time. The solar farm produces power when the sun is shining. The wind farm produces power when the wind is blowing. And where the sun is shining, where the wind is blowing most consistently, aren't necessarily in the exact spots where people are demanding power. There's an interesting market dynamic that's sort of played out here where renewable energy producers actually underwrite their projects against two different revenue streams. So, there's the obvious one of power sales. They're in the business of producing power and selling that power. But maybe the less obvious one to most people is actually they underwrite the projects against production tax credits where they actually get paid these production tax credits if they produce renewable energy power and then sell it. It doesn't have any constraint around what price they sell it or who they sell it to but so long as they sell it. So this has resulted in kind of a weird market dynamic where people are building wind farms in places where it's very consistently windy and where they're going to produce wind power sort of the highest percentage of the time independent of if there's actually a buyer there. And so you look at a market like West Texas where the wind is very consistent, plentiful, but there's not a lot in West Texas to actually demand that power. And so people have built these large 100 plus megawatt wind farms. These wind farms are actually generating power very very consistently, but the rate at which they sell that power is often negative. So they're actually paying someone to basically get rid of the power so that they can collect that production tax credit. Sort of created this weird market inefficiency. We can help solve that problem by actually partnering directly with those upstream wind energy and solar producers by collocating digital infrastructure on site and again bringing a market to an otherwise stranded energy resource. So it's honestly not too dissimilar from our flare mitigation business. It's sort of a different problem, different segment of the energy market, but it's very much aligned with our core value proposition of wanting to deliver climate-aligned large-scale computing infrastructure that can also be cost effective.

Yeah, it's a good question. So we've tried to actually build out a portfolio of different computing applications that have different reliability constraints. And so our cloud business today is primarily we consider to be mission critical. We try to have a very very high runtime whether it's running on flare gas power or on-site renewables. We've had to sort of engineer some solutions that can deliver five nines of power reliability. Now, Bitcoin mining is obviously a lot more flexible. So, when you sort of pair this together with cloud computing, you end up with like sort of this stack of interruptibility. And ultimately, we're going to be adding some preemptable instances on our cloud platform that will be able to manage to lower reliability standards, but probably still you'd want to target higher reliability than something like Bitcoin mining. And that's one of the beauties of Bitcoin mining, frankly, is it's like the most flexible, interruptible, curtailable computing workload in the world. And it's sort of like a power sponge. It can soak up any excess power and any excess kilowatt hours that are being produced and it can turn waste energy resources into the hardest soundest digital money that's ever existed. I think what's cool is when you look at it from a campus perspective, you can actually orchestrate the way power gets produced and utilized on site to actually manage multiple different computing workloads in a way that actually benefits all of them and actually drives down the cost for all of them. Because in the case of your cloud computing workloads, you're typically having to engineer backup power generation resources and other things to manage a high degree of uptime. If you can now monetize those backup generation resources with Bitcoin mining, it actually makes the overall cost of the whole campus go down quite significantly.

Yeah, that's right. I think we've tried to align economic incentives with environmental impact outcomes. The trade-off that we've really made here, any business I feel like is, you know, you're just making a set of different trade-offs. But the trade-off we've made here is that we're actually willing to build data centers in places that people historically haven't, which means maybe some challenges on building out the networking architecture and footprint. Maybe it means having a little bit more latency to get to the location. But because of that, we've actually targeted workloads and applications that are more tolerant of a bit more latency. And that's really what's led us to Bitcoin mining is a great use case. It's a globally decentralized network. Having hundreds of milliseconds of latency is actually tolerant within Bitcoin. You know, you can actually calculate what your cost is per extra millisecond of latency, but it affords you quite a bit of flexibility there. In the AI landscape, when you're looking at training large language models or running these big training workloads, those are very latency agnostic as well. Even on the inference side, if you're feeding tens or hundreds or thousands of tokens into a big transformer-based model, the inference time actually way exceeds any sort of additional latency that's required to get to a data center that might be in a place like West Texas where it's like you're adding 50 milliseconds of latency compared to a data center that's in Northern Virginia. But ultimately that doesn't matter because the inference time is 2 seconds. It becomes kind of a rounding error and that characteristic of AI has actually opened up new possibilities of how the data center world architects the infrastructure needed to support its growth.

Yeah, there's a lot of interesting hardware solutions that are coming out. This boom has unfolded. It's created a very attractive prize for, you know, many and I think many chipmakers are sort of watching Nvidia's stock go up and up and up as their data center sales unit is, you know, knocking it out of the park. You know, it is a very fast-moving landscape. When you look at the H200, which is the next generation from the H100, ultimately it's a very similar architecture to the H100. It's the H100 with a little bit more HBM high bandwidth memory. What we're seeing from a benchmarking standpoint is it performs slightly better on certain inference workloads, but it's fairly similar performance on most training workloads. So, we are excited about that. The next generation of NVIDIA architecture we're also very excited about as well. You know, we're sort of expecting towards the end of the year and into early next year and we're also excited about kind of other things that are happening in the silicon, you know, the AI silicon ecosystem between AMD recently launching the MI300X which is sort of their competing product to the H100 which we've seen especially from a price performance standpoint be able to get some gains on the H100, especially on inference workloads. That chip actually has a lot more high bandwidth memory compared to the H100. It's 192 gigs of HBM. We think it's sort of an interesting landscape in terms of, you know, how things are going to play out. Intel's obviously not sitting idle as well. Been sort of investing in what's called the Gaudi platform. Constantly sort of watching that and excited to kind of see what they build. And you know there's a whole suite of independent startups that are focused on this as well between SambaNova, Cerebras, Graphcore, Groq that are all sort of having their own solutions that may be competitive in the overall landscape. But you know ultimately Nvidia is still sort of the king here and really the dominant player in the overall AI chip space.

So, you know, I think the numbers I saw were that Nvidia was going to produce 500,000 H100s in 2023 or that was the estimate and they were going to roughly produce around 1.5 million here in 2024. So, there was a mad scramble to get H100s in 2023 when they were very high in demand, but I don't feel like it's like eased up at all because a lot of the things that are working are actually demanding and requiring a lot more capacity and then you have a lot of big players that have made major moves in terms of building out large clusters to try to compete in the AGI race. So Facebook announced last week that they're going to be building out a 600,000 H100 or no H100 equivalent cluster. I think of which was about 350,000 were actually H100s. But when you look at that, I mean that's a staggering capital investment. I mean it's like $20 billion capex investment in terms of building out this cluster and that's really just coming from one company. You know, there's a lot of different other folks that are out there competing in this landscape. And then the applications that are working driving a lot of adoption between ChatGPT and many of these other chat bots, things like Character AI, they're just demanding tremendous amounts of resources to support the inference workload. So from our perspective and our seat, the supply crunch hasn't loosened at all. In fact, maybe it's got worse.