🎙 We held the third Lecture Club call on Stanford MS&E 435: Economics of the AI Supercycle.
Lecture 3 — Chase Lochmiller (CEO Crusoe). Topic — AI data centers as physical infrastructure: land, energy, chips, cooling, and where growth hits its limits. The call was led by Alexander — thanks for the presentation.
Some thoughts from the discussion:
• Alexander (lead) — The main framing of the lecture: the output of the language model = "digital labor". For the first time in human history, humanity is scaling labor through capital rather than demographics. Stargate plans 1.2 GW (more than Estonia consumes). And also: the shortage of electricians in the USA is 500K by 2030, with salaries ranging from $120-260K per year. AI is the first major energy buyer that physically moves closer to the source.
• Ira — Devil's advocate: the main shock of the lecture — 4000 m³ of water for cooling one building. And the main career takeaway: if AI takes your job — go learn to be a plumber, not an ML engineer.
• Stepan — Software now = ~$0.5T (about 0.5% of global GDP). The global payroll fund = ~$55T. If AI replaces even part of the salaries — the TAM for "software 3.0" grows by 100×. Meanwhile: China has already geographically separated inference/training (the coast for response, Tibet for training), and Russia will follow suit (Moscow vs Siberia).
• Vasily — Y Combinator has been publishing RFS for startups that train electricians and plumbers for data centers for a year now. "Sun/wind" is mostly marketing: what’s really needed is a gas turbine (which can ramp up quickly) or nuclear. A data center is a living organism, engineers need to be nearby, so "building in the desert" won’t work.
• Pavel — The real bottleneck is not chips, but the grid and transformers. Gemini checked Crusoe's composition in Texas: 60% gas, 20% wind, 12% coal. So "green AI" is mostly PR.
• Mikhail — The queue for transformers is 160 weeks (vs 50 in 2021). Crusoe is buying used transformers from closed thermal power plants — "vacuuming" the market. Of the 16 GW announced in 140 projects, only 5 GW is being built, half will be canceled or postponed. The current capacity was built before the AI boom (3 years of approvals + 4 years of construction = 7 years lag).
• Dima — Whitewashing "cheap energy": households will pay the bill for the AI boom. Ohm's law — losses during transmission, and the Northeast US (New York, Boston, Washington) will see an increase in utility bills due to the load on the grid.
• Anya — The internal infrastructure of data centers will be completely updated every ~4 years. The next investment frontier — grid infrastructure companies (Schneider, Eaton).
• Alexander (closing) — AI "reincarnates" nuclear energy. After Three Mile Island (1979) and Chernobyl (1986), the West closed dozens of nuclear power plants — now AI demand is reversing the trend. Rosatom is in a good position to enter this market.
📺 Recording of the call: https://www.youtube.com/watch?v=pLXnUAEc1wI
