Build the AI Supermajor

Chase Lochmiller is the co-founder and CEO of Crusoe, the vertically integrated AI infrastructure company that builds data centers, rents GPUs and sells tokens. He studied physics at MIT, did research at Los Alamos National Laboratory, and was recruited into quantitative finance before making enough money to step away. He climbed Mount Everest in 2018 with no next job lined up, and Crusoe came out of that blank slate, starting with Bitcoin mining on stranded energy. The company recently raised a $3.9 billion Series F at a $30.9 billion valuation and is approaching 2,000 employees. These signals are from Chase Lochmiller's interview with Harry Stebbings on 20VC.

1. The Planned Pivot. Lochmiller says building an AI platform was Crusoe's goal from day one, years before it was known for anything besides Bitcoin mining. His premise was that compute and data would bottleneck AI, and energy would bottleneck compute. Bitcoin was simply the best way to monetize cheap stranded energy at the time, while a small team built the cloud product on the side, launching to paying customers in early 2022. ChatGPT's launch on November 30, 2022 shifted his probabilities, and he moved resources toward AI data center design and energy sourcing.
2. Bitcoin's Data Center Lesson. Bitcoin miners started on laptops, moved to colocation facilities with five nines of reliability, then realized they could run at one nine. Those "chicken coop" sites cut about 98% of total data center cost. Lochmiller expected a similar break in AI as Nvidia's roadmap pushed chips from 150 or 200 watts toward 600 watts each. Because inference time is dominated by compute inside the building, AI also didn't need to sit in Northern Virginia; it could go wherever energy is cheap and abundant.
3. Vertical Integration Beats Lead Times. For the first two Abilene, Texas buildings, a little over 200 megawatts, Crusoe promised delivery in one year when the next closest of 34 bidders quoted 2.5 years. Vendors quoted 100 weeks for a medium-voltage power distribution center. Crusoe's in-house electrical manufacturing team built one in 28 weeks. Lochmiller says the margin from making it yourself is modest; the payoff is on-time delivery and a clear view of true costs, which he compares to Elon Musk's "idiot index."
4. Plugs Are the Shortage. The supply constraint shows up as a lack of places to plug in GPUs and turn them on. Lochmiller calls the supply chain a game of whack-a-mole where the bottleneck keeps moving. Today energy and skilled labor top the list: there are only so many electricians, welders and plumbers in the US, and they have to be in the place where the power is. That is one reason Crusoe prefers manufacturing infrastructure in factories to running everything as a giant construction project.
5. The Water Myth. Lochmiller calls the claim that data centers drain local water "just wrong." Each of the first eight Abilene buildings is budgeted at about 140 megawatts and uses roughly as much water a year as 10 single-family homes, mostly for staff bathrooms and landscaping. GPUs are liquid cooled through a closed loop that sends water to chillers outside. He says the facts on water, energy prices and jobs are all on the industry's side, even as people react emotionally.
6. Data Centers Lower Power Bills. Harry Stebbings pressed on rising energy prices, and Lochmiller says the data shows the opposite in markets with big builds. Data centers bring new generation capacity, and spreading more megawatts over the same transmission and distribution lines lowers costs for everyone. In Abilene, Crusoe expects to account for more than a third of Taylor County's tax revenue and more than double the tax receipts going to schools. His honest list of downsides is traffic, construction dust and noise.
7. The Political Lightning Rod. Data centers have become a midterm election issue, and Lochmiller never wanted to be "the main character" of a policy debate. He thinks people fear for their jobs, and data centers are the physical form of AI. Stebbings partly blames AI leaders who spent years predicting all jobs would vanish. Lochmiller's counter is that the buildout is driving a boom in US blue-collar work, with factories and field crews hiring skilled trades.
8. Three Products, One Hedge. Crusoe sells data centers, GPUs and tokens, and Lochmiller wants to be an "AI supermajor" in the mold of Exxon or Chevron. Exxon famously doesn't hedge oil because it is integrated: when crude falls, wellhead margins shrink and refining and plastics margins grow. Lochmiller expects AI margins to move the same way across electrical, data centers, chips and services. Right now managed GPU clusters earn the best margins because supply is so short.
9. A Portfolio of Contracts. Crusoe treats the GPU hour as a traded commodity and mixes contract types to manage risk. Five-year take-or-pay rentals to credit-quality customers pay back within the term. Shorter contracts earn higher margins but carry renewal risk. Managed inference and serverless fine-tuning are shorter still and the highest margin of all.
10. Depreciation Runs Longer. Crusoe depreciates GPUs over six years, the industry standard, and Lochmiller expects services to stretch that further. When Crusoe bought Hoppers in 2023, lenders doubted the chips would be worth anything after year three. Three years later, Hopper rental rates are higher than when the chips were new. He says people underestimate how inventive developers are at turning compute into useful services, and older, slower chips can still sell cheaper intelligence.
11. Keep the GPU Busy. The GPU is the most expensive thing in the data center, so an idle one is "just money burning." Lochmiller treats dollars per token as a useful but imperfect measure, alongside throughput, time to first token and time to last token. Much of the edge in inference comes from managing the KV cache across GPU memory, system DRAM, NVMe drives and object storage. He also sees enterprises spending more on closed frontier models while generating more tokens on open source.
12. Most Moats Are Illusions. Asked what he changed his mind on in the past 12 months, Lochmiller says "most moats don't exist," and most are temporary while model capabilities improve this fast. What lasts is moving quickly and adapting. He applies the same speed logic to customers: startups are being asked to commit to compute for 2028, so Crusoe is building small, modular, factory-made data centers to cut time to token. He protects his own time the same way, taking red-eyes to make bedtime and staying unreachable from 7 to 8 a.m. while he makes his kids breakfast.

Watch the full video at https://www.youtube.com/watch?v=Ko7nU1Img40.