Loading
Loading
Grid, electricity, nuclear, and energy economics for AI compute. The physical foundation beneath every GPU cluster.
Compute and energy meet here. The same chip delivers very different tokens-per-watt depending on rack design and cooling.
Ratio of total facility power to IT load. 1.0 is theoretical perfection. State-of-the-art liquid-cooled AI sites run 1.1; legacy air-cooled ones run 1.5+.
Modern AI racks pull 120-200 kW. The same building can host 5x fewer racks at AI density than at cloud density. Power, not floor space, is the binding constraint.
Lifetime cost per MWh across capex, opex, capacity factor, and lifespan. The number that decides which source actually gets built.
Average output divided by nameplate. Nuclear runs 90%+. Solar 25%. The difference reshapes the buildout mix more than headline cost does.
P = IV, I²R, Ohm's Law — three simple equations set the hard ceiling on how much compute fits in a building.
Grid interconnection can take years. Solid-oxide fuel cells sidestep it — megawatts of clean on-site power in months, no substation, no combustion. The bridge while the grid catches up.
Compute scales on an 18-month cycle. Power scales on a 10-year cycle. The frontier model is whatever fits in the gap.
A frontier training run runs at 1 GW continuous. By 2030, 5–10 GW per cluster. The actual numbers behind the headlines.
Rack power, cooling loops, on-site generation. The physical anatomy of where 5 GW gets used.
Grid, behind-the-meter gas, fuel cells, SMRs. The buildout mix that actually shows up by 2030.
Who pays when utilities build new transmission for AI loads? The rate-case fight is the political story of 2027.
Transmission is the slowest layer of any modern economy. The bottleneck inside the bottleneck is wires, not watts.
The honest accounting: water, power, noise, taxes, jobs, property values, and trust. Steelman the objection, then check the data.
AI training loads swing too fast for iron-core gear.