Crusoe CEO: Why Everyone Gets GPU Depreciation & AI Energy Costs Wrong

The infrastructure to support AI wasn't going to be centralized. It was going to be distributed where energy was low-cost and abundant. Energy prices actually come down. It's actually the opposite of the narrative that's being told. So what is a legitimate concern?

People are very emotional about data centers. When you look at what data centers are doing, I think the facts are all on our side. There's really like three products that we ultimately sell to customers where we're making money: data centers, GPUs, tokens. You know, the GPU is actually the most valuable thing in the entire data center. Most moats are illusion. Most moats don't exist.

I'm pumped to be here, man.

Yeah.

It's a great question. We've actually ingrained mountaineering into the culture of Crusoe. And we have a core set of company values. One of our core values is actually to think like a mountaineer. And so what does that mean? You know, mountaineering is a practice where things are subject to change and you need to be prepared for change. And so with it, it really embeds this sense of like cultural resilience that I think is really, really important for a business like ours that, you know, is doing both stuff both in the digital world as well as the physical world. But you know, to further the analogy here, you know, for mountaineering, you know, you come up with a plan to basically set out from your camp, go summit the peak, you know, come back safely and you know, you take your photo and you know, your glory is written in the history books. But you have to be prepared for things to go wrong, right? You have to be prepared for the weather to change. You have to be prepared for gear to fail. You have to be prepared for, you know, your partner to get sick or an avalanche to occur. And you know, so you know, the idea is really having, you know, here's a plan A and then I have a plan B, a plan C, a plan D. You know, I have plans all the way down in case different things go wrong. You know, additionally, like you have to have a sense of willingness to endure difficult challenges and difficult times. You know, this notion of enduring the expedition. The expedition could be long and hard and arduous and you know, you have to really dig deep at those moments of pain to work through those challenges.

So that's a different aspect of thinking like a mountaineer. And then you know, another big aspect is thinking with a safety culture, having this notion of safety as a huge priority, which is important when you're doing things that, you know, at the scale we're doing from a construction standpoint. Thinking with the safety mindset of, you know, safety first, second and third. You know, there's this saying in mountaineering that getting up to the top is optional. Getting down is mandatory. So we try to practice that safety culture here at Crusoe. And really, you know, try to avoid any sort of, you know, physical injuries or challenges with digital attacks on the platform as well.

It's tragic just kind of, you know, seeing when people take unnecessary risk because, you know, actually spending a lot of time within high altitude mountaineering culture and communities, you do see oftentimes people taking unnecessary, very careless risks. And it's actually quite shocking and, you know, oftentimes tragedy is avoidable. It is preventable. And you know, I think having that mindset of thinking like a mountaineer and planning ahead is what enables mountaineers to do it for a long period of time.

I think maybe one of the most difficult moments of change for me was I had in my life, I always had sort of the next thing planned before I left the previous thing. So what I mean by that is when I left high school, right, I knew exactly where I was going to college before, you know, the end of high school. When I left college, I knew exactly where I was going to work and, you know, what I was going to do. And when I left my first job, I knew I was, you know, going to work in going to grad school and I had a job lined up even after that. When I left to go climb Mount Everest in 2018, I did not have a plan after that. I did, you know, I had been had a reasonable amount of financial success in my life to that point, but I didn't have the thing planned. And actually having that void of that sort of emptiness really was both challenging and empowering for me because it felt like I could do anything. It was sort of like a blank slate. I had thoughts of like maybe I'm just going to invest my own money and just kind of like enjoy the rest of my life and not, you know, not do much. But I could also start a company. And if I were going to start a company, what are the things, what are the most challenging things that I could, you know, think about tackling that are interesting and, you know, that I would be excited about pursuing on a daily basis. And that was kind of like the emergence of Crusoe sort of came through that chapter of difficulty and challenge of just nothingness, right? Just having an empty, empty slate.

There's a bunch of like psychological aspects to it. I mean, you could look at like Maslow's hierarchy of needs, you know, just like it is empowering to know that like if I take risk and this all goes to zero, like I'm still going to be fine. I'm going to be able to, you know, cover my rent and, you know, feed my family and, you know, cover these base needs that, you know, I may have in my life.

This is specific. When I was in university, I mean, I was convinced I was going to become a theoretical physicist and likely a professor, you know, and I was spending time doing research in basic science, right? So I was doing research in physics, you know, both at MIT and then at Los Alamos National Lab. And at some point, I like, so early on in that journey, I was like pretty committed to living this monastic life of like not really earning a bunch of money but just really being having this life of discovery where I get to go discover the mysteries of the universe and, you know, that would be my life's work and pursuit. At some point I realized it's a very slow-moving industry. You know, this industry of discovery, you know, it may be changing today because of AI but at least at that point in my life it was a very slow-moving industry and I felt like I wanted to do something very fast-paced and I felt like I had sort of lost my life's purpose. Like I was like, okay, I don't feel like I have a purpose anymore. So absent having a purpose, I was like, man, I guess I should get rich. Like that's maybe the next best thing to like having a monastic calling that you're devoted to. And, you know, I was looking around for different ways to make money. And, you know, this headhunter found me and recruited me into this field of quantitative finance. And it was like, hey, you're really good at solving math problems. If you come work in this industry, they just pay you a lot of money to, you know, solve math problems. And I said, this sounds great. Sign me up. But I think once that need was met of like, okay, I have enough money. Okay, what next? And I think to your original question of like, you know, starting from this position of strength of, you know, being an entrepreneur and, you know, actually having had some financial success, it's empowering to take bigger swings, empowering to take bigger risks, which can ultimately lead to bigger outcomes.

Yeah.

You know, it's interesting investors that you spoke to validated this claim, but, you know, from the outside, it does appear to be a pivot. Internally when we started the company, the goal was always to build an AI platform. That was what I had sort of spent my career working on with AI, machine learning, you know, algorithms, and, you know, had been a big user and early adopter of deep learning for, you know, certain tasks. And I felt like it was this meta science that ultimately was going to, you know, transform every industry and, you know, at that scale compute becomes the bottleneck. Compute and data become the two bottlenecks and, you know, what is the bottleneck for compute? Energy. So that was kind of the premise of starting the company. Now, we obviously weren't dedicating all of our resources towards building an AI platform day one. We were dedicating actually the majority of our resources towards monetizing waste energy with Bitcoin mining. That was kind of the philosophy of starting the company and, you know, finding out, finding and seeking out these low-cost abundant energy resources that we could monetize with compute. Bitcoin was the best monetization engine we had at the time. But we were building this AI platform from very, very early days. So when you think about at what point do I shift the resource allocation to this other thing that maybe I'm working on in the side, I think about the world in a very Bayesian approach, right? I don't think the future is deterministic. I think it's very probabilistic and I sort of have a view of what the future's going to look like. It's somewhat, it's more certain if you go out just a little bit of time. If you go out, you know, a further amount of time, you know, it gets, it's a little bit hazier, but I still have a view of sort of what things are going to unfold. And what was happening with this AI platform was it was sort of a project that Chase had devoted some time and resources and engineering towards internally and we were committed to kind of, you know, exploring this pathway. But it was only, you know, we had launched the Crusoe cloud platform in early 2022 with our first set of paying customers. You know, before that we were working with a lot of researchers at MIT and Stanford and, you know, different groups trying to understand what's actually valuable for AI researchers like from an infrastructure platform. And so we finally launched in early 2022, but the moment that changed everything was obviously November 30th, 2022, the launch of ChatGPT, right? This was like the tweet heard around the world, the, you know, the moment that just shifted everything in terms of like, okay, like AI is here and, you know, there are really interesting, you know, use cases that are going to explode from sort of this moment. And that was kind of like a moment where my view of the world shifted quite a bit.

I just felt there was going to be far more demand for AI compute infrastructure solutions and, you know, we were very well positioned by actually having launched that platform at that point.

We had sort of watched this evolution of the data center to support Bitcoin, right? And if you look at the history of Bitcoin mining, it's pretty fascinating, right? It started with like hobbyists on their laptops in this sort of decentralized way and then people started to use GPUs and then people started to lease data center capacity from, you know, more traditional colocation facilities that have like five nines reliability. People started to develop ASICs and then at some point people realized, you know what, we don't need five nines reliability to mine Bitcoin. We can do it with like three nines. Oh wait, no, what about one nine? Like how much does that reduce the actual data center cost? And they were able to eliminate something like 98% of the total data center cost with like these like chicken coop type data centers, you know, very, very low cost, no redundancy, but, you know, ultimately brought down the payback period for these investments in Bitcoin mining. And, you know, we felt always that something similar was going to play out in AI. And, you know, we're long-term partners of Nvidia and we'd been working with Nvidia and I was looking with them at the roadmap, you know, many years ago and I was like, okay, our current generation of chip is, you know, 150 watts or 200 watts and, you know, the next generation is going to be like 300 watts and then maybe, you know, there's going to be a future generation that's maybe 600 watts per chip. And I was like, okay, well, you know, if as the power density goes up, it's going to change what the data center looks like to actually support these things. And we also had a perspective on the compute workload, the workload itself for AI. It didn't need to be in these centralized facilities. I think so long people have been anchored on these like centralized hubs for data center computing infrastructure. So, Northern Virginia is like the hub that runs most of the internet. Most web applications run out of Northern Virginia. But it didn't have to be that way with AI because, you know, so much of the time to serve a neural network is actually the compute that's happening in the data center, not the time to get to the data center. And so that for me I felt like opened up new geographies. And we felt like the infrastructure to support AI wasn't going to be centralized. It was going to be distributed where energy was low-cost and abundant. And so when we think about things that we did differently, it was like investing more heavily in how to design and develop the data center to support AI and investing heavily in how do we find and source energy resources to support the scaling of AI.

Where it's actually manifesting is there are not places to plug in GPUs. So that's ultimately the supply constraint there. There's just not places where you can plug in GPUs and turn them on and run AI workloads.

You know, the supply chain to support large-scale AI data centers is sort of like a game of whack-a-mole. At different points, different bottlenecks come into play and I think that's actually one of the critical aspects of being vertically integrated is that we're able to navigate a lot of those critical bottlenecks and situations. I'll give you an example. When we set out to build the first two buildings in Abilene, the first, you know, it was, you know, a little over 200 megawatts of compute capacity. We had sort of committed to doing this in one year and the next closest bid was two and a half years. There were 34 other data center developers and, you know, but speed was of the essence. So I said, sure, we can do it in a year. One of the key bottlenecks was what's called a power distribution center. It's where, you know, you receive medium voltage power and then you, so at around 34.5 kV, and then you distribute it to low voltage transformers that ultimately go into the data center and power the racks. When we went out and canvassed the market for vendors that were supplying these medium voltage power distribution centers, the lead time for this one component was 100 weeks. And I said, well, I don't have 100 weeks. I committed to doing this in a year. So I went to our internal team and we had consciously vertically integrated electrical manufacturing. We said, how quickly could we make this ourselves? The team did a bunch of work, sourced components to make these things and we were able to do it in 28 weeks. And so being able to navigate those types of key bottlenecks by actually having the resources internally gives you ultimately a lot more flexibility to unblock these key bottlenecks. Number one. And number two, it gives you a perspective on like what is the true cost of a lot of these different, you know, things that are being built. You know, Elon has this very like interesting framing of this. He calls it the idiot index in terms of, you know, what is the cost of raw materials and then what is the end product relative to that cost of, you know, just the raw materials going in. And, you know, we try, by being vertically integrated, we have a perspective on like the end-to-end cost of everything going into building and operating an AI factory.

Electrical manufacturing is not like an ultra-high margin business. Like today, it's actually spreads have gone up quite a bit. Margins have gone up quite a bit just because they're so short in supply, but it's not where like the critical alpha of it. It's not the main reason we're doing it is like stacking margin. It's more availability and getting out availability, being able to deliver on time. And then also it gives us a critical, it gives us a critical innovation platform by being able to, you know, start from first principles and say, okay, we can bring in raw materials and we can build anything we want in the world. What is the thing that we want to build to ultimately serve the workload? And so when you look at the campus in Abilene, the inspiration for that was like, okay, we want to build a one gigawatt scale computer where you can interconnect a giant cohesive cluster of GPUs that can operate one cohesive workload across the entire campus and what does it look like to actually build that thing from the ground up. And so that's how, you know, the cores are designed. That's how the GPUs are designed. That's, you know, all of it is designed to run on one cohesive RDMA fabric.

When we think about scaling of inference, you know, 2026 is really about the scaling of utilization of AI infrastructure or utilization of LLMs and foundational models for useful tasks. You know, it's been this era of, you know, scaling of agents and tokens, token utilization. And, you know, during that moment of scaling infrastructure as people are finding utility in these AI models, it's driving a lot more demand for inference. It's the era of the agent, the era of inference. And for that, you don't need a gigawatt scale computer to serve these models. You can do it with a much smaller cluster. And, you know, what matters is actually how quickly can I go from not having anything to being able to deliver tokens. Like, so what is my time to token?

You know, it varies. I would say energy is definitely a key constraint. So power that's available for compute is a key bottleneck. Alongside that, labor is a very key bottleneck. There are a finite amount of skilled trade workers, electricians, welders, plumbers, construction workers in the United States. And then when you overlap trying to find it, you know, if I have access to energy in this unique place, can I get the labor to that place to actually bring the AI factory to life? You know, that is a big challenge.

Power is not significantly more expensive.

No.

There are subsidies going in place, but, you know, in the United States, you know, we are globally competitive with power pricing.

I wouldn't classify it as a problem. I would classify it as something that needs to be navigated. And, you know, when we think about policies, we want good policies to be in place. We don't want just the wild west, everybody out doing whatever they want, building whatever they want with no care for anybody else in the world. You know, these are big investments that are being made. They have to be done thoughtfully. And so we are supporters of the right policies being in place, but it does cause friction in terms of developing things faster. That's actually one reason we also like the idea of manufacturing a lot more of this infrastructure as opposed to making everything a giant construction project.

Well, I mean, the thing about policy is it's hyper local. Like all politics are local. And what people care about is they care about like are there jobs being created? Is this going to drive up my energy costs? Is this going to take all of my water? Is this going to pollute the air so that, you know, my kids suffer from some disease as a result? Like, you know, that they care about. They care about are the data centers being good stewards and good citizens of the community and are they actually creating value for the community. And then, you know, what are other ways that the data centers are sort of giving back. And so, you know, this comes in the form of things like, you know, tax revenues and things like that. So I think one of the challenges for AI data centers broadly right now is that I think there's a lot of misinformation out there about what the impact of a data center is. The number of people that have told me that data centers use all the water in the world is crazy and, you know, it's also just false. So when you look at all of these modern AI factory designs, they use pretty close to zero water. Like one of our giant buildings in Abilene, Texas that consume, you know, it's a call it 140 megawatts of power is like what's budgeted for each of those buildings on the first eight buildings. They use about the same water annually as about 10 single family homes. The primary water usage there is like the staff using the bathroom, washing their hands, us watering the plants on the property. It's very diminutive water usage and there is water in the building. We are using water to cool the GPUs. But the critical aspect here is we've actually designed a closed loop architecture. So the cold water is entering into the racks and then it's going out to a chiller outside and we're sort of exhausting the heat.

The water requirements argument is just wrong. Yes.

I think this is an important aspect for the data center industry to get right and to tell effectively. We like to look at data for these things. When you look at markets where data centers have made investments and have built big data centers, typically energy prices for communities have come down. Because it catalyzes more investment in energy generation technology, in energy generation capacity and you end up with more megawatts being advertised over the same transmission and distribution infrastructure and as a result people's energy costs end up coming down. When you look at some of these big builds that need a lot of net new generation capacity, we are very supportive of, you know, the data center industry helping to bring online new power production in order to support those energy requirements.

The data has shown that energy prices actually come down. It's actually the opposite of the narrative that's being told. So what is a legitimate concern?

Anytime there is a big construction project, it does cause other aspects, right? There's benefits. There's people with money in town that, you know, are spending and, you know, if you talk to any local business owner in Abilene, times have never been better, right? The amount of money being spent at restaurants and coffee shops and hotels and all the local services is, you know, is more than it's ever been in the history of Abilene. The challenge is with that comes traffic, right? With it comes some dust from construction and some, you know, just noise that, you know, is involved with a very large construction project. But, you know, I think that will go away and there will be long-term permanent jobs in place that, you know, is continuing to support the local community. And there's tremendous tax revenue, right? If you look at the tax revenue that's coming from this, we're going to account for more than a third of the tax revenue in Taylor County and Abilene, the town. And this is transformative for local services like the police, the fire department, the roads, the schools, for the school system. We're more than doubling the tax receipts that are going to the schools.

50% seems reasonable. It's one of these things, you know, it comes back to thinking.

You know, it's one of these things where, you know, kind of coming back to this notion of thinking like a mountaineer when you're going through these planning processes. A lot of things can go wrong, right?

Getting, you know, permits, getting entitlements, getting land, get, you know, if there's, you know, you're trying to acquire a certain piece of land and, you know, the person you're trying to buy it from ultimately doesn't want to sell. Getting an, you know, getting a large load interconnection agreement done with the local utility. Getting an air permit if you're bringing online new generation or

Yeah. And they'll just give you a green light on day one.

It's a benefit to the data center developers if you're viewing it through a purely Machiavellian lens. Yes.

I've never wanted to be the main character of a political policy debate. And, you know, data centers have come front and center to the upcoming midterms in the United States and they've become this very polarizing topic. You know why that is? I think they are kind of this, I think there is concern around, you know, this era of AI and I think people are legitimately concerned around like, am I going to have a job in the future? Am I going to have, you know, am I going to be able to provide useful work to the world that, you know, I'm going to be able to get compensated for. So I do think that that is a concern on people's minds and data centers are sort of this physical manifestation of AI, right? And I think putting all of that to the side, like the biggest aspect is, you know, it is an emot-, it's a current, I think there people are very emotional about data centers. I think when you look at the facts of what data centers are doing from job creation, from driving down energy costs, from being

water neutral, from generating long-term tax revenue that is transformative for the communities that we're investing in. I think the facts are all on our side. So, you know, slightly blame with respect to Dario and a lot of other leaders who've said for the last few years, 'We're going to replace all jobs. We're going to replace all jobs,' and then surprisingly they hate you.

The interesting aspect to all this is that data centers are this massive employer. They're leading to this massive economic boom for the entire blue-collar labor economy in the United States. This massive resurgence and re-industrialization of the United States. We're hiring people, you know, in factories, in the field. And these are, you know, tons of folks that, you know, are skilled trade workers, right? People working with their hands to bring the infrastructure of intelligence to life. And I think the interesting aspect is that people are concerned that AI is going to take jobs, but yet it has so far only created tremendous amounts of jobs and tremendous amounts of economic development. So again, I think it's one of these moments where people are having an emotional reaction to something, but if you look at the facts of what's actually happening, the data is on the side of data centers being actually a very positive investment for the communities.

I would say the price of a GPU hour, it's kind of like a traded commodity. You're even seeing kind of these commodity platforms creating compute as a tradable asset. And I think people are sort of thinking through this lens of like compute is going to be this next, you know, great commodity. So, you know, there's a couple platforms like the Compute Exchange, or, you know, I think there's a handful of others that are like creating these tradable futures products. So, you know, I kind of view it as like that's a product that we're producing and, you know, there is like commodity price fluctuation in it. When I think about payback periods for this, I think, you know, when we are taking debt against it, I think there's different ways we think about like managing that risk and we sort of think about a portfolio of different services and a portfolio of different returns that we're going to get for this.

So what it means, maybe putting it precisely. So we have a range of different deals that we're doing with the compute we're buying. We are renting capacity on a long-term basis, you know, call it five-year contracts to, you know, credit-quality customers that, you know, we expect to pay. Those are going to pay back within, you know, that time frame and they're going to cash flow during that time frame. There are shorter-term contracts that we'll do at higher margins, but they're riskier because at the end of the contract it's like, is there renewals? There are other things that you're going to, you know, earn from that resource. And then finally there's other services that we provide and package. So things like Crusoe managed inference, things like our serverless fine-tuning product that actually helps people — that's more of a developer tool that ultimately enables developers to, you know, fine-tune, you know, some of these open-source models, these leading-edge open-source models, so they can get better performance for maybe some domain-specific task. Or actually as their applications are scaling, they can actually serve the tokens needed to, you know, from an inference basis to, you know, enable that application scaling. Those contracts are typically much shorter-term in nature and they end up being higher margin to Crusoe.

Correct. So, one of the analogies I've used in the past is, you know, when we think about Crusoe and sort of our vertically integrated strategy, there's really like three products that we ultimately sell to customers where we're making money. We can sell data centers, we can sell GPUs, and we can sell tokens. Like those are the three core products that we sell. And why does that matter? You know, if you look at a different market, I think that there's very interesting parallels for, you know, AI infrastructure and AI compute is the energy market. If you look at the oil and gas market and you look at the value chain from, you know, the wellhead to the product, you know, the gas station or, you know, the plastics that are being sold to people, there's a lot of different ways to make money across the value chain, but it typically gets broken down into the upstream business, which is drilling wells and sort of, you know, producing oil. There's the midstream business of like transporting oil and gas to, you know, a refinery. And then there's sort of the downstream business, which is typically refineries, gas stations, basically getting it into these finished products that you're ultimately selling to customers. Well, what Crusoe is focused on doing is what we believe the opportunity is, is to build an AI supermajor, right? So there are these supermajors like Exxon and Chevron in the traditional oil and gas industry that are vertically integrated across upstream, midstream, downstream. And where that becomes actually important is that, you know, Exxon very famously doesn't hedge their oil exposure, and you say, how could that work, you know, during this, you know, very crazy commodity price cycles? Well, you know, they're sort of inherently hedged by being vertically integrated. And I think that's where the margin accrual is going to move around. And I think you're seeing a very similar phenomenon play out in AI infrastructure. But when oil prices come down, Exxon makes less margin at the wellhead, but their margins in the downstream business actually go up. So, you know, because the oil prices are lower, the margins on gasoline, on plastics, on all these finished products, they actually go up. And I think we're seeing a similar phenomenon play out in AI where it's like, you know, our margins are going to move around across electrical, data centers, chips, and services.

I would say today, like managed GPUs, like, you know, managed compute clusters is like incredibly high margin like right this instant. And, you know, the reason for that is just a massive shortage of supply.