AI is running out of Power

Summary

The video argues that after chips, power is the next major AI bottleneck, especially firm grid capacity. It explains why AI datacenter developers are moving to behind-the-meter gas generation despite higher cost, because early power can be worth billions. It contrasts ERCOT and PJM market outcomes and says the AI supply chain is expanding into turbines, transformers, fuel cells and skilled labor.

• SemiAnalysis forecasts US AI datacenter power demand rising from about 3 GW in 2023 to 28+ GW by 2026 and 84 GW by 2030.
• Grid interconnection queues are clogged with speculative requests, and available grid headroom is approaching zero before turning negative by 2027.
• xAI's Colossus used rented truck-mounted turbines and engines, making behind-the-meter generation the template for the industry.
• Behind-the-meter power is not cheaper than grid power; it is a timing arbitrage because six months of AI cluster revenue can be worth about a billion dollars.
• Labor shortages in turbine manufacturing and specialized casting are a hard physical limit on the buildout.
• PJM capacity market prices rose 9.3x due to simulation and forecasting issues, while ERCOT's energy-only market kept household prices roughly stable.
• The AI supply chain is expanding from chips to gas turbines, fuel cells, engines, transformers, substations, switchgear and rare earths.