The AI power buildout's hidden constraint

The American grid is about to find out what it looks like when one industry starts eating the entire growth budget. The popular worry is that the country is about to overbuild power for an AI cycle that may not arrive. The more plausible failure mode, in the view of McKinsey's latest modeling, is the opposite one: the lights do not stay on because the wires, plants, and permits were not ready in time.

A power plant now takes close to a decade to clear siting, permitting, and interconnection. The demand side is being locked in by a small group of hyperscalers (Microsoft, Google, Amazon, Meta) whose capital spending grew roughly 50 percent a year from 2022 through 2026 while margins held. That is a buyer cohort that is contracted, not speculative. The risk of underbuild rises as the buyer becomes bigger and more patient than the grid it depends on.

The falsifier is plain. If hyperscaler capex rolls back, or if AI execution stalls the way the same survey says it sometimes does, the thesis breaks. Until then, the public conversation worth having is not whether America is building too much power, but whether it is building it fast enough. On the current trajectory, the answer is no.