20VC with Harry StebbingsCrusoe CEO: Why Everyone Gets GPU Depreciation & AI Energy Costs Wrong
20VC with Harry Stebbings
Introducing Crusoe and its business 0:00
Chase Lochmiller runs Crusoe Energy, a company that recently raised $3.9 billion in a Series F round at a $30.9 billion valuation. He explains that Crusoe sells three things that actually make money: data centers, GPUs, and tokens. He makes a point of saying the GPU is the single most valuable asset inside a data center, and he argues that most of what people call a competitive moat in this industry is an illusion that does not really exist.
Thinking like a mountaineer 1:06
Lochmiller is an experienced mountaineer, and Crusoe has built this into a core company value called thinking like a mountaineer. He explains that mountaineering trains you to expect change, so you always carry a plan B, C, and D in case weather turns, gear fails, or a partner gets sick. It also teaches endurance through long, arduous stretches, and it instills a safety-first culture, summed up by the saying that reaching the summit is optional but getting down is mandatory. He recalls meeting people who took careless risks in high-altitude climbing and says tragedy in that world is usually avoidable.
Leaving Everest with no plan 6:31
Lochmiller says the hardest change in his life was leaving to climb Mount Everest in 2018 without any next step lined up, unlike every prior transition in his life. That emptiness was both unsettling and freeing, and it eventually led him to start Crusoe. He also describes an earlier pivot, leaving a path toward theoretical physics research at MIT and Los Alamos National Lab because the pace of scientific discovery felt too slow, which pushed him into quantitative finance to build financial security before founding a company.
From Bitcoin mining to AI platform 11:01
Crusoe's founding idea was always to build an AI platform, since Lochmiller saw compute and data as the coming bottlenecks for AI, with energy as the bottleneck behind compute. Early on, most resources went into monetizing wasted energy through Bitcoin mining, which was the best monetization engine available at the time, while the AI platform was developed quietly on the side. Crusoe launched its cloud platform in early 2022 after studying what AI researchers at MIT and Stanford actually needed. The release of ChatGPT on November 30, 2022, shifted his view of the world, convincing him that demand for AI compute infrastructure was about to explode, and he draws a parallel to how Bitcoin mining data centers evolved from hobbyist rigs to stripped-down, low-redundancy facilities that cut costs by around 98 percent.
Rethinking Data Center Design for AI 15:00
The speaker explains that his company anticipated AI chips would grow steadily more power hungry, moving from 150 to 200 watts per chip toward a future of 600 watts per chip, and that this shift would change what a data center needed to look like. He also argued that AI computing did not need to sit in the traditional centralized hubs, like Northern Virginia, which has long handled most internet and web traffic. Because so much of the time to serve a neural network is spent on compute inside the data center rather than on the journey to reach it, AI infrastructure could be distributed to wherever energy was cheap and abundant, rather than clustered in a few hub cities.
Vertical Integration Solves Supply Bottlenecks 17:01
The real supply constraint in AI right now is not chips themselves but the lack of physical places to plug GPUs in and run them. Building large data centers involves a chain of shifting bottlenecks, something the speaker compares to a game of whack-a-mole. When building the first two buildings in Abilene, roughly 200 megawatts of capacity, he committed to a one-year timeline while the next fastest bid among 34 competing developers was two and a half years. A critical bottleneck was the medium voltage power distribution center, a component with a 100-week lead time from outside vendors. By manufacturing it internally through the company's own vertically integrated electrical manufacturing team, they cut that to 28 weeks. Being vertically integrated, he says, gives both flexibility to clear bottlenecks and clear visibility into the true end-to-end cost of building an AI factory, an idea he likens to Elon Musk's concept of the "idiot index," comparing raw material cost to finished product cost. The goal in Abilene was to build a full 1 gigawatt scale computer, with every GPU and core designed to run on one cohesive RDMA fabric across the whole campus. Looking ahead to 2026, the focus shifts toward scaling inference and utilization of models rather than building ever bigger training clusters, meaning smaller clusters and faster "time to token" become what matters most.
Energy, Labor, Policy, and Public Misconceptions 22:02
Energy availability and skilled labor, including electricians, welders, plumbers, and construction workers, are named as the two biggest constraints on building AI infrastructure in the United States. He disputes the idea that U.S. power is significantly more expensive than China's, while acknowledging China's labor costs are lower. On regulation, he argues red tape is not really a problem but something to be navigated, since big infrastructure investments need thoughtful policy rather than an unregulated free-for-all. He stresses that policy concerns are hyper-local: people care whether data centers create jobs, raise energy costs, consume water, or pollute, and whether they act as good community stewards. He pushes back hard on the claim that data centers use enormous amounts of water, noting that a 140-megawatt building in Abilene uses about as much water annually as 10 single-family homes, thanks to a closed-loop cooling design. He also argues that, contrary to popular belief, data center investment tends to lower local energy prices over time by spurring new generation capacity, rather than raising them. Legitimate downsides of construction include traffic, dust, and noise, offset by major local economic benefits in Abilene, including a surge in restaurant and hotel business and tax revenue that will account for more than a third of Taylor County's receipts, more than doubling funding to local schools. Asked what share of currently planned data centers will actually get built, he suggests around 50 percent, pointing to risks like failed permits, land acquisition falling through, and delays in utility interconnection agreements.
Data Centers Become A Political Flashpoint 30:30
Chase Lochmiller addresses why data centers have become a polarizing issue ahead of the US midterms, tying it to public anxiety about AI taking jobs. He argues the emotional reaction ignores the facts: data centers create jobs, drive down energy costs, are water neutral, and generate long-term tax revenue for host communities. He pushes back on the idea that AI leaders predicting mass job replacement should be surprised by public backlash, pointing out that data centers are currently fueling a blue-collar hiring boom and a broader re-industrialization of the United States, employing skilled trade workers to build what he calls the infrastructure of intelligence.
Selling Data Centers, Chips, And Tokens 34:00
Crusoe's compute business works like a commodity, with GPU hours increasingly traded on exchanges. Lochmiller describes a portfolio approach: five-year contracts with credit-quality customers that pay back steadily, shorter higher-margin but riskier contracts, and developer-facing services like managed inference and serverless fine-tuning that carry the highest margins. He compares Crusoe's vertically integrated strategy to an oil and gas "super major" like Exxon, selling across upstream (data centers), midstream (GPUs), and downstream (tokens), so margin shifts between layers rather than disappearing, similar to how Exxon's refining margins rise when crude prices fall. Right now, he says, managed GPU clusters are the highest-margin layer because supply is so short. Most of Crusoe's GPU rental revenue is "take or pay," meaning customers pay for committed capacity whether or not they use it, though Lochmiller isn't worried about a demand crack because AI's benefits span the whole economy.
Rethinking Chip Depreciation And Timelines 41:32
Crusoe depreciates chips on a standard six-year cycle, but Lochmiller believes the real usable life is longer, since Hopper GPUs bought in 2023 now command higher prices than when new. He says early financiers underestimated how quickly application developers would find new value from existing compute. Managed AI services help extend monetization further by decoupling the service from the underlying chip. On demand forecasting, Crusoe builds projections from direct customer conversations, but the lag between ordering and deploying AI infrastructure keeps growing, pushing Crusoe toward smaller, modular, factory-built data centers to shorten delivery times.
AI Progress and Uncertainty 46:03
The conversation turns to how fast AI has advanced, noting that the Turing test was once seen as an almost impossible milestone, yet it was passed with little fanfare, and that AI is now solving long-standing Millennium Prize problems. Despite this progress, there is real uncertainty about jobs and industries, illustrated by a lawyer who has not written a document herself in five months because she now relies on AI tools. The guest argues the direction of change is certain, even if the exact shape of that change is not.
Measuring the Cost of Intelligence 49:30
Asked about the idea that the lowest cost producer of intelligence wins, the guest says dollars per token is a useful measure but imperfect, since tokens vary in value and efficiency. Performance also depends on throughput, how many tokens per second a system can generate, and latency, meaning time to first token and time to last token. He explains that the GPU is the most expensive asset in a data center, so keeping it constantly busy matters enormously, and that managing the KV cache, a memory system that stores previously computed answers so they do not need to be recalculated, across GPU memory, system memory, and storage is central to delivering fast, efficient inference.
Open Source, Frontier Models, and the Future 54:00
He notes that more money is spent on closed frontier models, but more tokens are actually generated on open source ones, and expects open source to keep growing in importance for data ownership and for companies wanting to build models on their own private data. He expects a mix of open and closed approaches rather than one dominant path, pointing to companies like Cognition and Harvey training on their own data while frontier labs keep expanding general capability. In a closing quickfire round, he describes prioritizing family over work, including protecting an unreachable hour each morning for his kids, and says he believes data centers are a genuine community benefit despite public skepticism. He adds that he no longer believes in lasting competitive moats, seeing most as illusions in a fast-changing technological landscape, and says he is unsure whether Crusoe will be public by 2028.
Capital Needs and Going Public 1:01:31
Building data centers and large scale GPU clusters takes enormous capital, and being a public company offers advantages like access to scaled capital resources. He believes Crusoe is ultimately better off public eventually, though the timing still needs to be right.
An Admired Figure Not on the Board 1:02:00
Asked who he wishes sat on his board, he names Michael Dell, admiring him as a person with strong business instincts, a great family man, and someone he looks up to, even though Dell is unlikely to join.
What 2018 Self Would Find Unbelievable 1:02:31
He says his 2018 self would be stunned that the bold bet on an energy first AI cloud actually worked, with energy becoming the real bottleneck to scaling intelligence at this scale. He recalls once being awed by a friend's company of 150 people, and now Crusoe itself employs nearly 2,000, a number he says he never takes for granted.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
