Skip to content
Sign in
Artwork for Exponential Learning
Exponential Learning · Thursday · 19 min

Racking Up the Power Shortfall: America's 33-Gigawatt AI Problem

A deep dive into this week's power update on the US data center power shortfall. The updated model raises 2026-28 incremental US data center power demand to 97 GW, with a gross shortfall of 57 GW and a net shortfall of 33 GW (34% of chip demand) after time-to-power solutions — roughly six New York Cities of missing power, widening to 72 GW by 2029. We cover why rack-scale NVL72 architecture blew up the forecast, the efficiency paradox (cost per TFLOP down 94% from Hopper to Rubin Ultra, operator margins heading toward 90%, yet total power demand rises), time-to-power economics for powered shell providers, behind-the-meter generation, and the report's beneficiaries across powered shells, power equipment, BTM developers, Equinix, and SpaceX's turbine foundry angle.

0:00-19:31
Play in

transcript

No transcript — this publisher did not publish one.

show notes

A deep dive into this week's power update on the US data center power shortfall. The updated model raises 2026-28 incremental US data center power demand to 97 GW, with a gross shortfall of 57 GW and a net shortfall of 33 GW (34% of chip demand) after time-to-power solutions — roughly six New York Cities of missing power, widening to 72 GW by 2029. We cover why rack-scale NVL72 architecture blew up the forecast, the efficiency paradox (cost per TFLOP down 94% from Hopper to Rubin Ultra, operator margins heading toward 90%, yet total power demand rises), time-to-power economics for powered shell providers, behind-the-meter generation, and the report's beneficiaries across powered shells, power equipment, BTM developers, Equinix, and SpaceX's turbine foundry angle.

more episodes

All episodes