Apoorv Agrawal · 00:28 "This is larger than the space program, larger than the Manhattan Project — second only to the U.S. defense budget."
Building a 1-gigawatt AI data center costs $60 billion — the equivalent of two Manhattan Projects. The five hyperscalers are now pouring a combined $650 billion into AI data center construction. At one site in Abilene, Texas (population 120,000), 9,000 construction workers are on the job every day.
The guest is Chase Lochmiller — founding CEO of Crusoe, the company actually building Project Stargate (the joint OpenAI / Oracle effort). As an aside, he's also a mountaineer who has climbed five of the Seven Summits, including Everest. The host is Apoorv Agrawal — Stanford GSB instructor and partner at Altimeter Capital (the venture capitalist behind major OpenAI investments).
The central metaphor is Chase's "Electrons to Tokens": feed electricity in the front of an AI factory, get tokens (units of AI output text) out the back. A single building of this kind costs $9 billion equivalent.
For 200 years, the only way humanity could expand its labor force was to raise birth rates — have a child, send them to school, raise them to working age. Twenty years per worker. For the first time in history, you can now add "digital labor" instantly by building a single data center. That's the real reason $1 trillion in capex is moving around the world in parallel.
Key Observations
The real bottleneck isn't GPUs — it's electricians and plumbers (24:30)
Per Chase, GPUs were the bottleneck four years ago — but today's biggest constraint is construction labor. The labor cost for building out 1 gigawatt is $4.7 billion. Gas turbine prices have tripled, from $1M to $3M per megawatt. "You can buy chips, but you can't buy plumbers." That constraint physically determines the pace of AI's evolution.
Electricity prices in Abilene were negative (10:30)
West Texas has strong wind and sun, and government production tax credits had driven over-investment in renewables — but transmission capacity couldn't keep up, leaving electricity prices in negative territory because there were no buyers. That's where Crusoe built one of the world's largest AI computing campuses. The result: 1 gigawatt — Denver-sized municipal power consumption — concentrated into a single company's chips.
H100 prices have risen three years after launch (33:00)
Normally a chip loses value when a new generation ships. But entering the agentic era, demand exploded and the H100 spot price has now exceeded its launch price. Standard public-company depreciation runs five years, but Chase's on-the-ground sense is "it's likely usable for 6+ years, possibly much longer." The IT industry's conventional wisdom — "when a new generation ships, the old generation is worthless" — has collapsed in the AI era.
Video Outline
- (00:00) Guest introduction — Chase Lochmiller
- (01:30) Data centers as the physical manifestation of AI adoption
- (05:00) "Electrons to Tokens" — reading AI through the lens of economics
- (06:20) For the first time in history, you can expand the labor force digitally
- (10:00) Why Abilene, Texas — the land of negative electricity prices
- (15:30) Project Stargate — 8 buildings for Oracle + OpenAI, with Microsoft expansion
- (18:00) Cost breakdown per megawatt — $20M (building) + $40M (IT) = $60M
- (24:00) The real bottleneck — labor and skilled tradespeople
- (28:00) Quad, TX — "behind the meter" design with wind, solar, and batteries
- (33:00) H100 price reversal — the depreciation assumption for chips has collapsed
- (40:00) Student Q&A
Sources
Class #3 | MS&E 435: Economics of the AI Supercycle, Stanford University Spring '26, Apoorv Agrawal