The Power Problem Behind AI

01The mismatch in one comparison

A hyperscale data center goes from groundbreaking to energized in roughly two years. The transmission that serves it takes five to ten. Everything interesting about AI infrastructure follows from that gap.

• Buildout timelines: silicon (months) vs. buildings (years) vs. grid (decade)
• Why the gap widened: load growth returned after twenty flat years

02From chips to megawatts

The unit math is unforgiving. Take an accelerator drawing on the order of a kilowatt, multiply by tens of thousands per cluster, add cooling and distribution overhead, and a single training campus lands in the hundreds of megawatts.

• A worked example: cluster size → IT load → facility load at realistic PUE
• Why each model generation raises the floor rather than the ceiling
• Inference: smaller per site, but everywhere — a different grid problem

03Power-first site selection

Fiber and land used to choose data center locations. Now interconnection capacity does, and the map of AI infrastructure is being redrawn around available electrons.

• The migration toward ERCOT, the Midwest, and the Southeast
• "Powered land" premiums and the speculative market around substations
• Why some announced campuses will never be energized on schedule

04The workarounds

When the grid cannot deliver, buyers route around it — each option trading cost, carbon, or time.

• Behind-the-meter gas turbines: fast, controversial, increasingly common
• Nuclear PPAs and restarts: long-dated bets on firm power
• Batteries and demand flexibility: shaving the peaks that drive upgrades

05The efficiency counterargument

Every efficiency gain in compute history has been answered by more demand. The question is whether AI breaks that pattern or confirms it.

• Jevons paradox applied to compute: cheaper intelligence → more of it
• What would have to be true for AI power demand to flatten
• Why we treat demand forecasts as scenarios, not predictions