20VC: The Future of Datacentres: What You Need to Know | Why Everyone Gets GPU Depreciation and AI's Energy Costs Wrong | Who Really Makes Money From AI & Why Most Moats Don't Exist with Chase Lochmiller

Crusoe's CEO says speed to power wins the AI build-out

Crusoe is betting that AI infrastructure is a vertically integrated energy business, from power and buildings through GPUs to tokens. Power and build speed are scarce, so whoever delivers capacity fastest captures margin as it moves between layers. Lochmiller argues the economics hold up, while he and host Harry Stebbings clash over public backlash.

Takeaways

• Power and skilled labour limit AI build-out. The binding constraint is places to plug in GPUs. Crusoe committed to building Abilene's first two buildings, over 200 megawatts, in one year against a rival bid of two and a half. It brought electrical manufacturing in house, cutting a 100-week lead time to 28.
• Inference does not need gigawatt campuses. Serving models works on much smaller clusters, where time to token matters most. Crusoe began by chasing cheap, abundant energy, and Lochmiller expects infrastructure to spread to such sites. Bitcoin miners cut about 98 percent of data centre cost by accepting low reliability.
• Crusoe aims to be an AI supermajor. Like Exxon, it is vertically integrated from data centres through GPUs to tokens, so margins shift between layers. Managed GPU clusters earn the most today because of shortage. Contracts mix five-year take-or-pay rentals, shorter higher-margin deals and managed services.
• GPU depreciation fears are overstated. Crusoe depreciates chips over six years, the industry standard, and expects managed services to extend that. Hopper rental prices are higher now than when the chips were new. GPU hours trade like a commodity, and compute futures are emerging.
• Moats are an illusion in AI. Crusoe competes and collaborates with others across the stack, and Lochmiller praises Fireworks. Open-source models generate more tokens while closed frontier models attract more spending. He expects enterprises to use a mix, with domain-specific efforts from Harvey and Cognition.
• Demand outlook rests on scaling laws. Lochmiller says the future is not very unknown because scaling laws keep proving out. Customers are asked to forecast demand further out as build timelines lengthen. Gavin Baker's view that the lowest-cost producer of intelligence wins points to dollars per token and KV cache management.

Where host and guest disagree

• Community impact. Lochmiller says modern designs use close to zero water, host-market energy prices have fallen, and Crusoe supplies over a third of Taylor County tax revenue. Harry pushes on emotional public reaction and midterm politics.
• Project risk. Harry cites a claim that half of planned data centres will not be built. Lochmiller says permits, entitlements and land are what stop projects, and the facts favour the industry.