Crusoe CEO：为什么所有人都把 GPU 折旧和 AI 能源成本算错了

The infrastructure to support AI did not have to be centralized. It was meant to be distributed where energy was cheap and plentiful. Energy prices are actually falling. Everything is not at all as they say.

So, what is a valid concern? We're currently in a race for data centers, compute power, and energy, and there's no more appropriate guest to join me today than Chase Lochmiller of Crusoe Energy. They provide several different layers of this very important business. He sits next to me after they raised $3.9 billion in a Series F round, valuing the company at $30.9 billion.

People are very emotional about data centers. When you look at what data centers are doing, I think the facts are on our side. There are actually 3 products that we ultimately sell to customers and make money from: data centers, GPUs, and tokens. The GPU is actually the most valuable thing in the entire data center. Most advantages are an illusion. Most of the benefits simply don't exist.

1. How Mountaineering Shaped Chase as a Founder

Ready to start? Chase, I'm so excited about this, man. I already told you, I've been following you like crazy. I've even heard feedback from performance coaches. It was an incredible experience to conduct this research, so thank you very much for doing this with me.

I'm thrilled to be here, my friend.

2. Why Not All Inference Providers Are Created Equal

Now I'll start with something completely strange, something I didn't expect at all: mountaineering. I heard that you are an experienced and skilled climber, and I don't often see people like that in this chair. What have you learned from mountaineering that you think makes you a better entrepreneur?

That's a great question. We actually introduced mountaineering into Crusoe's culture, and we have a core set of company values. One of our core values is to think like a climber.

What does this mean? Mountaineering is a practice where everything is constantly changing, and you have to be ready for change. It really instills a sense of cultural resilience, which I think is very important for a business like ours that operates in both the digital and physical worlds.

To expand on this analogy, in mountaineering you make a plan: leave camp, conquer the peak, return safely, take a photo, and your glory will be forever inscribed in the history books. But you have to be prepared for things not to go according to plan, right? You must be prepared for changing weather, equipment failure, your partner getting sick, or an avalanche happening. The idea is to have a plan A, and then a plan B, a plan C, and a plan D—to have a whole series of plans in case something goes wrong.

You must also have the willingness to endure difficult trials and hard times. This is the concept of endurance during an expedition. An expedition can be long, difficult, and exhausting, and you have to put in your best effort in the moments of pain to overcome these challenges. This is another aspect of a climber's mindset.

Another important aspect is the culture of safety, the idea that safety is the highest priority. This is important when you work on the scale we do, from a construction perspective, thinking in terms of “safety first.” There's a saying in mountaineering: climbing to the top is optional, but descending is mandatory. We try to instill that kind of safety culture here at Crusoe and really try to avoid any physical injuries or challenges related to cyberattacks on the platform.

I have enormous respect for climbers. This is one of those times when I almost can't fathom all the complexity and psychological resilience required to deal with a change in plans, the weather, a partner's illness, and endurance, as you said.

I met an entrepreneur yesterday who had climbed Everest, and he showed me photos of the bodies of those who had died along the way. I thought, “This is just terrible. This is tragic.” When you see people taking unnecessary risks, because having spent a lot of time in the high-altitude mountaineering environment, you often see people taking unnecessary and very reckless risks. It's actually quite shocking, and often the tragedy could have been avoided.

I think it's the climber's mindset and planning ahead that allow climbers to do this for a long time. Can I ask you about resistance to change? What change in your life was the most difficult for you?

I think one of the most difficult moments of change for me was that, in my life, I always planned the next step before I even left the previous one. When I graduated from high school, I knew exactly which college I would go to before I graduated. When I graduated from college, I knew exactly where I would work and what I would do. When I left my first job, I knew I was going to graduate school, and I already had my next job in mind.

When I set out to conquer Everest in 2018, I had no plan for the future. I had some financial success in life at that point, but I didn't have a clear plan. This emptiness, this kind of vacuum, was both a challenge and an inspiration for me, because I felt like I could do anything. It was like a blank slate.

I thought, “Maybe I'll just invest my own money, enjoy life, and not do anything special.” But I could also start a company. If I did, what were the most challenging tasks I could set for myself that would be interesting and exciting to me every day?

3. Why Having Money Can Make Founders Take Bigger Risks

In fact, it was like the birth of Crusoe. It came through that stage of hardship and trials, through emptiness, through just having a clean, blank slate.

My job is to find signals that point to a potentially great entrepreneur, and I've noticed that wealthier entrepreneurs are often better because they don't worry about protection against losses and just see opportunities. Do you think starting Crusoe with some money and success already changed your vision for how to build it?

There are many psychological aspects here. You can think of Maslow's pyramid of needs, and it gives you confidence: knowing that even if I take a risk and everything burns down, I'll still be okay. I'll be able to pay the rent, feed my family, and cover the basic needs that I might have.

An interesting point for me: I never bargain on salaries. I'll be honest, I make a lot of money on the podcast, which is great, but I don't haggle over salaries. Life is too short. I want people to be paid as much as they want. If I get $5,000 or $10,000 less from a few people, I don't care.

This is specific to me. When I was at university, I was convinced that I would become a theoretical physicist and probably a professor, and I devoted my time to research in basic science. I did research in physics, both at MIT and at Los Alamos National Laboratory.

At the beginning of that journey, I was quite committed to this kind of monastic life, where you don't make a lot of money but just live for the sake of discoveries, to unravel the mysteries of the universe. This was to become the work of my whole life.

At a certain point, I realized that this is a very slow industry—the discovery industry. Maybe it's changing now because of AI, but at least at that stage of my life, it was a very slow industry, and I felt like I wanted to do something very dynamic.

I felt like I had somehow lost the meaning of my life. I was like, “Okay, I don't feel like I have a purpose anymore.” In the absence of a goal, I thought, “Dude, I should probably get rich.” This is perhaps the best alternative to a monastic vocation to which you devote yourself completely.

I was looking for different ways to make money, and a headhunter found me and recruited me into the field of quantitative finance. He said, “Hey, you're really good at solving math problems. If you come to work in this industry, you'll just be paid a lot of money to solve math problems.” I said, “That sounds great. Sign me up.”

But I think once that need was met, it was like, “Okay, I have enough money. What next?” Answering your initial question about starting from a position of strength, as an entrepreneur who has already achieved some financial success, this allows you to take bolder steps and greater risks, which can ultimately lead to greater results.

4. How Crusoe Went From Bitcoin Mining to AI Infrastructure

One last thing before we get to what I planned to talk about. This is just a continuation of what you said about the endurance element in mountaineering. Perseverance is great in some ways, but it can also make you keep doing things you probably shouldn't. Sometimes the right moment comes to change course rather than keep going.

You went through very big changes in business, at least that's how it looked from the outside. A lot of your investors have said that you always had a prediction that this would lead to where we are today, but it seems like it was like Bitcoin: phase 1 and then phase 2, AI.

I don't want to really get into this—maybe it's weird—but I want to ask your advice. What advice do you give to others about when it's time to persevere, to keep going and insist on your own path, and when to change tactics and take a different path?

It's interesting that the investors you spoke to confirmed that statement, but from the outside, it really looks like a change of course. Internally, when we founded the company, the goal was always to create an AI platform. This is something I've worked on throughout my career, with AI and machine-learning algorithms.

I was an active user and one of the first to start applying deep learning to certain tasks. I felt that this was the metascience that would ultimately transform every industry. At this scale, computing power becomes a bottleneck. Computation and data become the 2 main constraints.

Do you know what the bottleneck for computing is? Energy. That was a prerequisite for starting the company. Of course, we didn't dedicate all of our resources to building an AI platform from day one. We actually dedicated most of our resources to monetizing excess energy through Bitcoin mining.

That was the philosophy behind the company's founding: to find inexpensive and accessible energy resources that we could monetize through computing. Bitcoin was the best monetization tool we had at the time, but we had been building this AI platform from the very early days.

When you think about when to shift the allocation of resources in favor of this other cause that we might be working on in parallel, I look at the world through a Bayesian approach. I don't think the future is deterministic. I think it's very likely, and I have a certain view of what the future will look like. It's more defined if you look at a short period of time; if you look at a longer period of time, it gets a little hazier, but I still have an idea of how things will play out.

As for this AI platform, it was a project that I had dedicated a certain amount of time, resources, and engineering effort to within the company, and we were determined to explore that path. We launched the Crusoe Cloud platform in early 2022 with our first group of paying customers. Before that, we were collaborating with a lot of researchers at MIT and Stanford, as well as other groups, trying to understand what was really valuable to AI researchers from an infrastructure-platform perspective.

5. The ChatGPT Moment That Changed Crusoe’s Strategy

We eventually launched in early 2022, but the moment that changed everything was obviously November 30, 2022—the launch of ChatGPT. It was like a message that went out to the whole world. That was the moment that changed everything: we realized that AI was here, and that there were going to be really interesting use cases that would explode after this event. That's when my perspective on the world changed dramatically.

What exactly?

I just felt that the demand for infrastructure solutions for AI computing would be much greater, and we were in a very advantageous position because we had already launched our platform at that time.

What did you do differently right after that?

Since the advent of ChatGPT, we have watched the evolution of data centers to support Bitcoin. If you look at the history of Bitcoin mining, it's pretty fascinating, isn't it? It all started with enthusiasts on their laptops in a decentralized way. Then people started using GPUs, and later, renting data-center capacity from traditional colocation facilities with the reliability of “five nines.”

People started developing ASIC miners, and at some point they realized that Bitcoin mining didn't require 100% reliability. They could get by with “three nines,” or maybe even one nine. How much does this reduce actual data-center costs? They were able to reduce about 98% of their total data-center costs by using “chicken coops” instead of data centers. They were very low-cost and had no frills, but ultimately they shortened the payback period of the Bitcoin-mining investment.

We always felt that something similar would happen in the field of AI. We are long-term partners of NVIDIA and have worked with them, and many years ago I looked at their roadmap, where it was stated: “Our current generation of chips is 150 or 200 watts, the next generation will be about 300 watts, and maybe in the future there will be 600-watt chips.” I thought, if power density increases, that's going to change the look of the data centers needed to support these systems.

We also had a vision for the compute workload. The workload for AI itself didn't necessarily have to be in these centralized facilities. People have been tied to these centralized hubs for data-center infrastructure for so long. Northern Virginia is the hub where most of the internet runs, and most web applications are run from Northern Virginia.

With AI, this doesn't necessarily have to be the case because a significant portion of a neural network's serving time is spent on computations that occur in the data center, not on the time it takes to access it. For me, it seemed that this opened up new geographies, and we felt that the infrastructure to support AI would not be centralized. It would be distributed wherever energy was cheap and available.

When we think about what we did differently, it was investing more in the design and development of data centers to support AI, as well as making significant investments in finding and attracting energy resources to scale AI.

I want to break this down a little bit and get back to the structure that I had planned. I promise, this is great. But if we start with data centers, then we have Crusoe Cloud, managed inference, and energy flows through all of that, so I'll break that down as well.

Regarding data centers, we constantly hear about supply constraints today. How limited in supply are we really?

6. The Real Bottleneck in AI Infrastructure Today

This is manifested in the fact that there are simply no places where you can connect GPUs. This is ultimately a supply constraint. There are no places where you can plug in GPUs, turn them on, and run AI workloads.

The supply chain to support large AI data centers is like a game of whack-a-mole. Different bottlenecks arise at different times, and I think that's one of the critical aspects of vertical integration: we're able to overcome many of these bottlenecks.

7. How Crusoe Built in 1 Year What Others Said Would Take 2.5 Years

I'll give you an example. When we started building the first 2 buildings in Abilene, it was a little over 200 MW of computing power. We committed to doing this in 1 year, while the next-closest offer was 2 and a half years. There were 34 other data-center developers there, and speed was critical. So I said, “Sure, we can do it in a year.”

One of the main obstacles was the electrical distribution equipment. This is where the medium-voltage electricity comes in, at about 34-34.5 kV, and then is distributed to the low-voltage transformers that ultimately power the data center and the equipment racks.

When we surveyed the market for suppliers providing this medium-voltage distribution centers, the wait time for the component was 100 weeks. I said, “I don't have 100 weeks. I committed to doing it in a year.” So I reached out to our internal team, and we deliberately vertically integrated electrical manufacturing.

We asked, “How quickly could we do this ourselves?” The team did a great job finding the components to make these systems, and we were able to do it in 28 weeks. The ability to overcome such key obstacles by having your own resources first gives you much more flexibility to address them. Secondly, it gives you an idea of the true cost of many of the things that are being built.

Elon has a very interesting perspective on this. He calls it the “idiot index”—the ratio of the cost of raw materials to the cost of the final product derived from those raw materials. Through vertical integration, we're trying to understand the full cost of everything that goes into building and operating an AI factory.

How much additional margin do you squeeze out by having this in-house production process? What is the “idiot index” gap?

Electrical manufacturing is not a business with ultra-high margins. Today, spreads have increased quite a bit, and the margin has increased significantly simply because of the shortage of supply. But that's not where the critical advantage lies, and it's not the main reason why we do it.

It's more about margin accumulation. It's about availability, ensuring high availability, and being able to deliver everything on time. It also gives us a critically important platform for innovation: being able to start from first principles and say, “We can take the raw materials and build anything we want in the world.”

What is the thing that we want to build to ultimately serve the workload? When you look at the Abilene campus, the inspiration was, “We want to build a gigawatt-scale computer where you can put together a huge, cohesive cluster of GPUs that can run 1 cohesive workload across the entire campus.”

What does the process of creating something like this from scratch look like? That's how the cores are designed. This is how the GPUs are designed. It's all designed to work on 1 coherent RDMA fabric.

When we think about scaling inference, 2026 is really about scaling the use of AI infrastructure, or using LLMs and foundation models for useful tasks. It is the era of scaling agents and using tokens. At this moment of scaling infrastructure, people are finding utility in these AI models. This stimulates much greater demand for inference.

8. Energy & Labor: The Biggest Constraints on AI Data Centers

This is the era of agents, the era of inference. You don't need a gigawatt-scale computer to run these models. You can do this with a much smaller cluster. What's really important is how quickly I can go from having nothing to being able to generate tokens. What is my time to first token?

I like the analogy you gave, that it's like a game of whack-a-mole. What is the biggest supply constraint we have today?

It varies. I would say that energy is definitely a key constraint, so the power available for computing is the main bottleneck.

Along with this, labor is a very important bottleneck. In the United States, there is a limited supply of skilled workers: electricians, welders, plumbers, and construction workers. When you try to put that together, if I have access to energy in this unique location, can I get the workforce there to actually bring the AI factory to life? This is a big challenge.

With all due respect, if energy and labor are the 2 biggest constraints or challenges, then the UK is in a complete mess. Our energy prices are 4 times higher than anywhere else, and we have no labor. When people are there, they simply don't work. You have both in the US, but they're significantly more expensive than in China, to be honest.

Energy is not much more expensive.

Really? No. Don't you think that the Chinese government is subsidizing data centers with significantly cheaper energy to run them? I'm asking, not asserting.

There are subsidies, but in the United States, we are globally competitive in energy prices.

Of course. So, when we look to the future, then—

I agree with the workforce. Labor in China is much cheaper than in the United States.

What about policy and regulation? Everyone I talk to says, “Oh, it's bureaucracy. It's politics and regulations that prevent us in the US from building.” Is this as much of a problem as I'm told?

I wouldn't classify this as a problem. I would classify this as something you have to learn to work with. When we think about politics, we want the right norms to be in place. We don't want it to be the “Wild West,” where everyone does what they want and builds what they want without regard for anyone else in the world.

These are huge investments being made. They must be well thought out. Therefore, we support the implementation of the right policy. But it really creates friction when it comes to rapid development. Actually, that's why we like the idea of manufacturing most of this infrastructure instead of turning everything into grandiose construction projects.

I completely understand. If I were to task you with driving policy to help this ecosystem thrive, what would you do? What would you have done differently? You have a magic wand.

The thing is, politics is a purely local issue. As they say, all politics is local politics. People are concerned about whether jobs are being created, whether this will increase their energy costs, and whether it will take all their water. They worry that it will lead to air pollution, which will make their children suffer from some kind of disease.

People are concerned about whether data centers are good stewards, whether they are worthy members of the community, and whether they truly create value for it.

How else can data centers help communities?

This could be, for example, in the form of tax revenues and the like. I think one of the problems for AI data centers in general right now is that there's a lot of misinformation about their real impact.

9. Why Chase Says the Data Center Water Argument Is Wrong

The number of people who have told me that data centers use all the water in the world is just crazy, and that's simply wrong. When you look at all these modern AI factory designs, they use almost zero water.

For example, one of our huge buildings in Abilene, Texas, which might use, say, 140 MW of electricity—that's the budget for each of the first 8 buildings—uses about as much water per year as 10 single-family homes. The main use of water there is by staff: toilets, handwashing, and watering plants on the premises. This is a very, very small amount of water consumption.

There is water in the building. We use water to cool GPUs. But critically, we have developed a closed-loop architecture. Cold water enters the racks, then exits to an external cooler, where we remove the heat. So, the argument about water requirements is simply false.

10. Do Data Centers Actually Raise Energy Prices?

Yes. Okay, but what about rising energy prices?

I believe it is important for the data center industry to approach this issue correctly and communicate it effectively to people. We like to rely on data in such matters.

When you look at markets where there has been investment in building large data centers, energy prices for communities have typically declined. After all, this stimulates more investment in energy-production technologies and generation capacity, and ultimately more megawatts are distributed through the same transmission and distribution infrastructure. As a result, people's energy costs are decreasing.

When you look at some of these large facilities that require significant new generating capacity, we strongly support the data center industry helping to deploy new energy sources to meet those needs. So, Ashish, this is also misinformation. Energy prices will not increase for local cities and towns. The data suggests that energy prices are actually falling. This is the exact opposite of what is being told in this story.

So, what is a valid concern? Any major construction project has its consequences, right?

There are also advantages. People with money come to the city to spend it, and if you talk to any local entrepreneur in Abilene, you'll hear that times have never been better. The amount of money spent in restaurants, coffee shops, hotels, and all local services is greater than at any time in Abilene's history.

The problem is that with that comes traffic. With that comes construction dust and the noise that comes with any very large construction project. But I think this will pass, and instead there will be long-term, permanent jobs that will continue to support the local community.

This is huge tax revenue. If you look at the tax revenue from this, we will provide over a third of the tax revenue in Taylor County and the city of Abilene. It changes local services like police, fire, roads, and schools. Regarding the school system, we are more than doubling the tax revenue that goes to schools.

I completely agree with the domino effect you are observing. I think it's one of the worst and saddest things when we see how many millionaires are leaving the UK. People don't consider the domino effect of their departure in all these different aspects.

11. Why 50% of Planned Data Centers May Never Get Built

We see that many planned data centers are never built. We've seen very prominent people say, “We expect 50% of the planned data centers to never actually be built and operational.” What percentage of planned data centers do you think will actually fail?

50% seems acceptable.

This is one of those things that brings you back to thinking like a climber. As you go through these planning processes, a lot of things can go wrong, right? Give me a decimal. Why do I need this? I know. Let's end this. I mean, Neil told me you would be punctual. 50.

You know, it's one of those things where, going back to this idea of thinking like a climber, as you go through these planning processes, a lot of things can go wrong. What is the main reason why things go wrong, and what prevents this?

Getting permits, rights of way, and land; if you're trying to buy a piece of land and the person you're trying to buy it from doesn't want to sell; getting a large-load interconnection agreement with the local utility; getting an emissions permit if you're putting in new generation—

With all due respect, I don't want to go back to that, and I'm not an agent of the CCP, I promise, but in China, they'll move 3 million people off the dam for you on forklifts, and they'll just give you the green light on day 1.

Yes, that's definitely an advantage. This is an advantage for data center developers if you look at it through a purely Machiavellian lens.

Yes. I am a venture capitalist. I don't give a damn about sentimentality regarding your house. You will be relocated to a better place.

Again, charity is not our forte.

I completely understand. Can I ask you, as we look at these data centers, are they now being used as a political tool in a way that we didn't expect? Does this bother you?

12. Why Data Centers Became Politically Toxic

I never wanted to be the protagonist of political debates. Data centers have become a central topic in the upcoming midterm elections in the United States, and they've become a very controversial topic.

You know why? I think there's a concern around this era of AI, and I think people are quite rightly worried: Will I have a job in the future? Will I be able to do useful work in the world for which I can get paid?

I think this is an issue that worries people. Data centers are a kind of physical embodiment of AI, right? Putting all that aside, the biggest aspect is the emotion. It's the zeitgeist, I think people are very emotional about data centers.

I think if you look at the facts of what data centers do—job creation, energy-cost reduction, water neutrality, and generating long-term tax revenue that transforms the communities we invest in—the facts are on our side.

So, look, regarding Ireland and many other leaders who have said in recent years, “We will replace all the jobs. We will replace all jobs.” And then, surprisingly, they hate you. Oh my God, cake or death?

Cake, please.

The interesting aspect of all this is that data centers are a huge employer. They lead to a massive economic boom for the entire blue-collar industry in the United States, this massive resurgence of reindustrialization in the United States.

We hire people in factories and in the fields. That's a lot of people who are skilled workers, right? People working with their hands to bring the infrastructure of intelligence to life.

I think the interesting aspect is that people are worried that AI will take jobs, but so far it has only created a huge amount of jobs and a huge amount of economic development. Again, I think this is one of those moments where people get emotional about something, but if you look at the facts of what's actually happening, the evidence is on the side that data centers are actually a very positive investment for communities.

13. The Economics of GPUs and AI Compute

By the way, if someone is involved in data center security, I would be happy to invest in this business. Honestly, this is a very good play for private equity, not venture capital.

This is a phenomenal business. Another part of your business is, of course, providing GPUs and compute power. Can you help me understand the economics at this level? When you look at it like I’m a venture capitalist, how long does it take to get payback on the equipment?

I would say that the price per hour of GPU compute is like a trading commodity. You even see these commodity platforms treating compute as a trading asset. I think people are looking at this through the lens of computing being the next big commodity.

There are several platforms, like Compute Exchange and Orn, and I think there are several others that are creating these trading futures products. So I look at it as a product that we create, and there are fluctuations in the market price.

When I think about payback periods and the fact that we take on debt for this, there are different ways to manage this risk. We’re looking at a portfolio of different services and a portfolio of different revenues that we’re going to get from that.

What does this mean, to be more precise? We have a number of different deals that we enter into based on the computing power that we purchase. We lease capacity on a long-term basis, say 5-year contracts, with reliable customers who we expect to pay. They will pay for themselves during this period and provide cash flow along the way.

There are short-term contracts with higher margins, but they’re riskier because, at the end of the contract, the question arises: Will there be an extension? Is there anything else you’re going to make money from using this resource?

Finally, there are other services that we provide and offer in packages. This is, for example, supervised inference from Crusoe, our serverless training product, which helps developers customize some of these advanced open-source models so they can get better performance for specific tasks. Or, as their applications scale, they can serve the tokens needed for inference to enable that scaling. Such contracts are usually much shorter-term, and as a result, they yield higher margins for Crusoe.

Okay, I completely understand you. You want to have a portfolio with different margin metrics to form what you consider to be a healthy margin in combination.

14. Crusoe’s Strategy: Sell Data Centers, GPUs and Tokens

Right. One of the analogies I used earlier is that, when we think about Crusoe and our vertically integrated strategy, there are actually 3 products that we sell to customers or that we make money from. We can sell data centers, we can sell GPUs, and we can sell tokens. These are the 3 main products we sell.

Why is this important? If you look at another market, you’ll see some very interesting parallels with AI infrastructure and AI computing. That’s the energy market.

If you look at the oil and gas market and the value chain from the well to the finished product—like gasoline or plastics that are sold to people—there are many ways to make money in that chain. It’s usually divided into the upstream sector, which involves drilling wells and extracting oil. There’s a transportation sector, or midstream, which is the transportation of oil and gas, for example, to an oil refinery. And there’s the downstream sector, which is usually refineries and gas stations—in other words, the transformation of raw materials into final products that you sell to customers.

15. Building the “Exxon of AI Infrastructure”

So Crusoe focuses on what we think is the possibility of building an AI supermajor.

In the traditional oil and gas industry, there are supermajors like Exxon and Chevron, which are vertically integrated across the upstream, midstream, and downstream sectors. This becomes important because Exxon is known not to hedge its oil risks. You ask, “How can this work during such crazy commodity price cycles?” Well, they’re already hedged to some extent through vertical integration.

I think where the margin is going to come from is going to change, and I think we’re seeing a very similar phenomenon in AI infrastructure. When oil prices fall, Exxon gets lower margins on its wells. But in reality, its margins in the downstream sector are increasing. As oil prices are lower, the margins on gasoline, plastics, and all these finished products are actually increasing.

I think we’re seeing a similar phenomenon in AI, where our margins will fluctuate between electricity, data centers, chips, and services.

Which level is currently the most profitable?

16. Where the Highest Margins Are in AI Infrastructure Today

I would say that today, managed GPUs, meaning managed compute clusters, have incredibly high margins. The reason for this is simply a huge supply shortage.

17. What “Take-or-Pay” Really Means in AI Compute

I completely understand. Can I ask about a phenomenon or concept called “take-or-pay” in this business? Can you explain “take it or pay it” to me and anyone else who doesn’t understand it?

Of course. “Take-or-pay” is essentially when a customer agrees to pay for something regardless of whether they use it or not. For example, in the energy industry, take-or-pay contracts are common. If I reserve 100 MW of capacity, I pay for those 100 MW even if I don’t use them.

One might ask: Are all your revenues based on take-or-pay terms?

GPU rental agreements are usually based on a take-or-pay basis.

Yes, got it. Many people believe that the first sign of problems will be a drop in consumer demand. If this drop in demand spreads further and they have take-or-pay terms, they’ll say, “We don’t need it right now.”

I don’t really look at it that way. I think the thing is that AI is transforming so many sectors of the economy. It’s not just one thing that benefits. This happens in all areas without exception.

Consumer demand for what exactly?

Consumer demand for ChatGPT and core services.

I believe that, to be honest, we’re only at the beginning stages. I’m fascinated by the leaps and bounds of improvement we’re seeing in some of the newer models, fundamentally opening up possibilities for new discoveries that were previously impossible. The capabilities of the models are only growing.

When people think about the usefulness of these models and this infrastructure, this is the most important infrastructure that will be part of the future of all humanity. When you take on that perspective, it’s inspiring to see what’s actually possible.

I couldn’t be more optimistic or more confident about how a weakening in consumer demand for one thing would accelerate the scientific discovery of new materials, new medicines, and new things that would ultimately improve people’s lives.

18. The Real Risk of GPU Depreciation

When you buy chips on a scale like you do, you have to think about the life cycle. What do you think about the risk of chip depreciation?

Today, we depreciate assets using a 6-year depreciation cycle. It’s an industry standard.

One of our philosophies as a business is that we will invest in infrastructure for the long term. We will invest in data centers for the long term. We will invest in energy for the long term. We will invest in chips for the long term.

The more ways we have to turn these long-term investments into cash, the better off our shareholders will be and the better off our business will be. That’s what prompted us to invest in this set of managed AI services.

When you look at this business, where you abstract the actual chip—the computing power—from the service that people get, it opens up new monetization mechanisms that can work much longer. I think there will always be a demand for advanced silicon technology. There will always be a demand for frontier models.

There will also be demand for much cheaper alternatives. The ability to provide intelligence at a lower cost using older-generation chips that may be slower will still have value. Therefore, we believe that abstracting away some of the complexities through a set of services that allow these chips to be monetized for longer will ultimately extend the amortization cycle beyond 6 years.

19. Why Old GPUs May Stay Valuable Far Longer Than Expected

What do you think everyone misunderstands about buying chips and their depreciation that they should know about?

Listen, when we started making really significant purchases of Hoppers—and that was in 2023—we got feedback like, “We don’t even know if it’s going to be valuable after year 3.” Isn’t that right? 3 years have passed, and the prices for using Hoppers have become higher than the rates that were in effect 3 years ago, when they were brand new.

Early financing required a very quick payback, as well as significant collateral for debt service. I think people underestimate the ingenuity of applications and developers in turning computing power into value and economically useful services.

20. How Crusoe Forecasts AI Compute Demand

I was talking to Lee Jacobs, and he was telling me about that risky approach of having to order millions of dollars’ worth of NVIDIA GPUs before it was “cool,” as he put it. My question to you is: How exactly do you approach demand forecasting today? Is it just taking whatever you can get?

No, we build forecasts based on communication with customers. I believe we’re a very customer-oriented organization. We want to provide our customers with complete solutions.

The challenge today is that customers are increasingly being asked to forecast demand over the long term as the time to launch AI infrastructure increases. This creates problems for the entire industry.

One of the challenges Crusoe focuses on is reducing the time to result, or delivery time, of a managed GPU cluster by deploying small, modular AI data centers that provide just-in-time infrastructure, especially for small clusters. We believe this will be transformational for the industry.

This is a very important shift, especially when you look at early-stage companies or startups that are required to book capacity for 2028. That’s an eternity, right? It’s like…

You can get tendonitis by then.

That's right. Things can still change.

Was there a point in your career when the future seemed so unknown?

I actually don't feel like it's that unknown. I think we've made this breakthrough in the overall model development cycles. I believe these scaling laws continue to be confirmed.

It seems to me that we are constantly changing the criteria for what is artificial intelligence and what is AGI. Just think about the Turing test. When I first took computer science classes in high school, the Turing test was considered a critically important concept for true machine intelligence. Then it seemed like an impossible task.

We passed the Turing test, and almost no one even celebrated, right? Just think about the possibilities of the models we have today. And now we're solving the “millennium challenges”, right?

When I was a student at MIT, these “Millennium Challenges” came up, and it was like, "Wow, these are unsolved problems that have real monetary rewards." You could earn $1 million if you solved one of these problems. They remained unresolved for decades, and now we have AI that solves these critical problems.

I agree with you. I didn't mean to sound negative, not at all, but we see these swarms of exiled agents doing terrible things. We see entire industries under threat.

My girlfriend is a lawyer. 6 months ago, she didn't use Legora, and now she hasn't written a single document in 5 months. There is such uncertainty in everything. We have RSI, which could come and change absolutely everything.

I understand the comment that the future of work looks uncertain in some respects. But I also feel that there is really no doubt that AI will become a critical driver of work and productivity that will fundamentally change the way things are.

What's more important to me is that AI precisely develops capabilities that match or surpass human levels in terms of overall productivity. Of course, will this change everything? Yes, this will change everything.

There is uncertainty about the changes themselves, but for me, it is certain that there will be changes. Maybe that's another way of saying it.

21. Does the Lowest-Cost Producer of Intelligence Win?

I understand that perfectly. Since we mentioned that managed inference is a big part of your business, I completely agree with you. Gavin Baker, one of your investors, said that whoever produces intelligence at the lowest cost wins.

When you think about it, what is the unit of measurement? Is it dollars per token or per learning cycle? What is this unit for you?

Dollars per token is a good indicator. Dollars per token is probably a good metric in this context, but it doesn't necessarily mean the efficiency of using tokens. Not all tokens are the same. Not every token has the same unit of value, but I consider it a good indicator.

I think there are different parameters that people optimize for from a usage perspective. For example, throughput, which is the number of tokens per second that you can generate for a particular use case of the model, and time to first token, which is latency. There is time to first token and time to last token—how fast you can deliver it.

A lot of it comes down to a combination of different factors, right? What is the cost of this? Through vertical integration and ownership of the infrastructure, you can accumulate significant margins throughout the process. This gives you more flexibility in exactly how you build it.

Additionally, I believe that utilizing the GPU itself is a critical aspect of extremely efficient token delivery and achieving high-performance metrics such as latency and throughput.

GPUs are generally the most valuable thing in the entire data center, right? This is probably the most expensive element. That's something you have to keep working on all the time. If it's idle, it's like burning money.

If you look at it from that perspective, then, in my opinion, one of the key aspects is memory management, in particular KV-cache management. A KV cache is a key-value cache.

You can think of it this way: when tokens are fed into a large neural network, you can compute the output tokens by running a forward-propagation process and doing all these matrix multiplications. This takes time. You may already know the answer for this input token before entering the data. If you have entered these same tokens before, you know the answers and can find them in this lookup table.

The KV cache can get quite large, so it can go far beyond the amount of memory you have on the HBM chip. The ability to manage this across different GPUs, HBM levels, system memory and DRAM, and also things like NVMe, is critical.

That means any solid-state drive in the system, and then extending that to the object-storage layer of your cluster. Being able to move data very efficiently is a critical aspect of serving inference and supporting GPU workloads, so that you can really make progress on latency and throughput metrics.

What is it that no one sees in this layer of the value stack that you provide, and what should everyone see? What don't we know?

The important aspect is that not all inference service providers are the same. There are great open-source projects like SGLang and vLLM, but there are ways in which they are very general or universal.

When you look at an inference provider like Fireworks, Fireworks is great. They did an incredibly good job.

To what extent are they competing with you and displacing your managed inference business?

I believe there are several levels in the stack where Crusoe competes.

Absolutely. I guess my question is, what do they have that you don't have because of your specialization?

There are certain workloads that Fireworks handles incredibly well. The user experience is something that is unique to each of these inference platforms, and people may like one more than the other.

There is a range of different tools available, and inference service providers offer a range of different features. If I were them, I could do that.

I come back to thinking about the analogy with the oil and gas industry, where, if you look at the value chain, there are different places where you can compete and places where you can collaborate.

When Exxon pumps oil directly from wells, they can compete with another oil and gas company to drill and operate those wells most efficiently. But they can sell to the same customers. If they build a main pipeline to transport oil, they can provide capacity in that pipeline to other people with whom they compete.

Again, I think it comes down to the fact that I see a lot more ways to collaborate with people in the AI infrastructure stack than to compete with them.

22. Open vs Closed Models: Where the Tokens and Dollars Go

I completely understand this, and I agree with you. A lot of the world is competing with you if you look at all 3 levels.

When you look at open and closed models, given your experience from a managed inference perspective, what do you see in terms of the distribution of spend and tokens between open and closed systems?

I really think people spend more money on closed front-end models than open ones, but they generate more tokens on open source than on closed models.

We believe that open source will actually be very important and valuable. I think this is important from a data-sovereignty perspective, when people want to own their own models and their own intelligence. There is a lot more private data that is not yet used and could lead to significant productivity gains.

Do you think we will live in a world where many companies, both mid-sized businesses and large enterprises, will have their own models on their own data? Will that actually absorb a significant portion of the frontier-model business?

I think it's possible, but I also believe that there will always be a demand for frontier technology. Not to say that it doesn't exist, but you look at your Harvey and Ramp, and you were an investor in Maqore. We are investors in Viacom, and both companies support this and certainly spread it as a marketing message. I'm just wondering if that's true.

Regarding the capabilities of the models today, I believe there is huge room for improvement by incorporating more private data repositories. You can do it with advanced closed-source models, or you can do it with open source and own the model yourself.

I think in the future it will be some combination of these things. I don't think any one approach will become dominant.

I know, I know. There's a lot of interesting things happening right now, like companies such as Cognition, which are retraining their own models using a lot of their own data. Companies like Harvey might be trying to take a similar approach, particularly from a legal perspective.

There are a lot of these amazing subject-oriented approaches to models. Frontier Labs is doing this too, right? They're looking at that as well and trying to essentially push the boundaries of knowledge.

23. What Managed Inference Looks Like in Five Years

What will the managed inference market look like in 5 years?

The key aspect is to provide companies with an easy way to access intelligence. It could be their own intelligence, right? It's about deploying these specialized individual models that are very niche for their specific tasks. They can be trained on a lot of their own private data.

The key aspect of managed inference is that it abstracts away a lot of the infrastructure complexity of running this service across many different GPUs and data stores, with different models, in terms of how the requests are routed and which models are used.

24. Quick-fire Round

I’d like to do a blitz survey with you because there are many questions I want to ask, but I also understand that you need to mind your own business. Let’s start with one thing: raising children. I’ve heard from many of your friends and ambassadors that you’re also an incredible dad.

What if you gave me advice on how to succeed at work and still be a great father? Very important. What must be done? What absolutely can’t you do?

It’s really a matter of priorities. Spending time with my children is one of the greatest joys in my life.

I’ll stop you there, but I can’t—you’re raising $3.9 billion.

I mean, of course. It sounds trite, but children are the best thing in the world. For me, it’s like this: I have a very busy schedule and a lot of complicated commitments at work. I’m often on the road, but it all comes down to prioritization.

Often, this means I have to choose worse flights. I often fly at night so I don’t miss the time to put the kids to bed. I’m home almost every weekend, and those weekends I dedicate entirely to coaching the soccer team, helping with swimming, or whatever else my kids are passionate about.

I think it’s important to set aside specific times when people in the company know that I’m completely unavailable. If I’m at home in the Bay Area, everyone knows I’m unavailable from 7:00 to 8:00 in the morning. I just don’t have any meetings during that time because I’m making the kids breakfast, getting them ready for school, and getting them energized for the day. I’m simply not available on the days that I’m home.

What do you think many people disagree with you about, which makes you unpopular?

Perhaps it’s that data centers are a good asset for communities and should be welcomed in the communities where we build them. Again, it comes down to all the benefits that we think we can provide, whether they’re economic, whether they’re related to energy, or whether they come from the investments that we make in communities through schools.

Are you concerned about the income and wealth inequality that we’re seeing?

I’m not concerned about income inequality or wealth inequality because I think there’s something in human nature that goes, “Hey, I have this and the other person has that, so why don’t I have that?” There’s a bit of a “jealous neighbor” aspect here, but I still think it’s natural when you have these huge things that drive economic progress, where everyone actually benefits significantly more than if those investments weren’t there.

The 3 IPOs exceed the combined value of 45 years of IPOs. The concentration of capital is unprecedented.

Yes, but think about all the value that is created for the world.

Of course. 100%.

The market capitalization of a company is one thing, but if you look at the value being created through widespread access to intelligence, which will provide a huge productivity boost for the global economy, this is many times greater than the market capitalization that is being created.

What has changed in your views over the last 12 months?

I guess I’ve changed my mind about what creates sustainable, long-term competitive advantages. I think many of you know that venture capitalists love to ask about the “moat”—in other words, what is your long-term security? What will provide you with stable profits above the market?

I think I’ve changed my mind about most of these “moats” being an illusion. Most moats simply do not exist.

What makes you think that?

I believe that most of them are ephemeral, especially at a time of accelerating technological progress and rapidly improving model capabilities. I think it all comes down to the ability to move quickly and adapt effectively to the changing chessboard in front of us.

Will Crusoe become a public company by the end of 2028?

I’m not sure. The company needs a lot of capital to build data centers and AI factories, deploy large-scale GPU clusters, and build an infrastructure layer of intelligence on which a lot of global economic value will be created. This requires significant capital.

Public company status has many advantages, including access to large-scale capital resources. We really believe that, ultimately, the company will be better off in the public markets. The only question is when it will be appropriate for us.

Who would you like to see on your board of directors?

I would say Michael Dell, and he will not join my board. I just have incredible admiration for Michael as a person. He’s an incredible person. He built an incredible business and has excellent business instincts.

He’s a wonderful family man. I really look up to him a lot.

I talked to Zach before this. Zach’s a great guy. I love Zach.

Some fool.

Oh, my God. Sorry, that was unprompted. I didn’t see that. This is so funny.

I completely understand.

Last question: What would Chase from 2018 find most incredible about the modern Crusoe?

I think Chase from 2018 would be amazed that it all worked. It’s just incredible, because it was grand, bold, and ambitious to build such a cloud platform with artificial intelligence.

The idea of energy prioritization really coming to fruition, and energy becoming a bottleneck for scaling intelligence, would have seemed impossible. The fact that this happened on such a scale, I think, 8 years ago would have seemed simply impossible.

When I think back to that period, there were people who gave me great advice. I had a close friend who was an entrepreneur whose company at the time had about 150 people. When you’re just starting a business, the idea of having 150 employees seems almost unbelievable. I kept thinking, “How can I possibly have 150 people under my command?”

Today, Crusoe employs almost 2,000 people, and even when I say it out loud, it sounds like an insane number of people who have dedicated their most productive career years to Crusoe. They work side by side with us in the trenches every day.

I never take for granted that people invest their time in us. This is an extremely important contribution, because these people have made a bet that this is a cause to which they want to dedicate their years of life.

It was a great pleasure to talk to you. As I said, it has been a real pleasure to follow you and hear your many stories. Thank you so much for joining me. It was fantastic.

Thank you very much for inviting me. This was great.