Energy, Power and the Physical Limits of AI

Artificial intelligence runs on electricity, and by 2025 the electricity had become harder to get than the chips. A single frontier training cluster now draws as much power as a mid-sized city: xAI's Colossus in Memphis, Crusoe's Stargate campus for OpenAI and Oracle in Abilene, Texas (~1.2 GW planned, plus a ~900 MW expansion leased by Microsoft), and Meta's Hyperion in Louisiana are each measured in hundreds of megawatts to gigawatts. The International Energy Agency estimated that data centres used ~415 TWh in 2024, about 1.5% of world electricity, and projected ~945 TWh by 2030, with AI the main driver (Energy and AI, April 2025). In the United States, Lawrence Berkeley National Laboratory put data centres at ~176 TWh, or 4.4% of national electricity, in 2023, rising to 6.7–12% by 2028 (December 2024). The four largest US cloud and platform companies guided to roughly $720–750 billion of combined capital spending for 2026 (company guidance, July–October 2026, reported), much of it for buildings and the power to run them. The result has been an unprecedented scramble: restarting shut nuclear plants (Three Mile Island Unit 1 for Microsoft, Duane Arnold for Google), 20-year contracts with existing reactors (Amazon_AWS with Talen at Susquehanna, Meta with Constellation and Vistra), bets on small modular reactors that will not deliver before about 2030, gas turbines on site, sold-out turbine factories, record capacity prices in the largest US grid market (PJM), large-load queues of hundreds of gigawatts in Texas, and political fights over who pays for the grid. This section explains how much power AI uses and why, what a data centre's power chain looks like, the economics of power in an AI campus (power is a small share of cost but the binding constraint on time), the supply options (grid, gas, nuclear, renewables plus storage), the grid bottlenecks (interconnection, transmission, transformers, turbines), water and local politics, efficiency gains and the Jevons argument, and the physical limits that power shares with chips, memory and packaging.