How much new AI power demand?
The construction gap
Visualizing the reactors we are not building
Each square is one gigawatt-scale nuclear plant. Amber squares are plants your scenario requires. The tiny green count is large reactors currently under construction in the United States: effectively zero.
Why electricity is the bottleneck
Training and serving frontier AI models requires enormous, always-on electric load. A single hyperscale AI campus can draw as much power as a mid-size city, and demand is compounding: US data centers already consume roughly 4-5 percent of national electricity, with credible forecasts of 9-12 percent by 2030.
Schmidt's framing is blunt arithmetic. Ninety gigawatts of new firm capacity is about the output of ninety large nuclear reactors. The United States commissioned its most recent new reactors, Vogtle 3 and 4, after roughly 15 years and over 30 billion dollars, and currently has essentially no large reactors under construction. Grid interconnection queues stretch four to seven years, and gas turbine backlogs run to 2029.
The result: the constraint on AI progress may shift from algorithms and silicon to substations, transmission lines, and generation. Whoever solves power first, whether through nuclear restarts, small modular reactors, gas, or solar-plus-storage, gains a structural advantage in the AI race.
Assumptions used in this calculator
1 large nuclear plant = 1 GW. Average US home demand = 1.2 kW (about 10,500 kWh per year). One H100-class GPU with cooling and overhead = 1.5 kW. Solar with storage delivering firm power = about 10 acres per firm MW. US total generating capacity = about 1,300 GW. Figures are order-of-magnitude estimates for intuition, not engineering studies.