It's not the size of AI's load. It's the pulsing.

The load AI data centers place on the grid is not steady. It pulses. Training runs synchronize entire GPU fleets, and synchronized machines draw synchronized power. The backup gear data centers have always relied on was never engineered to ride a metronome. The market still prices that gear as if it were.

Uptime Institute analysts frame the failure mode plainly: running equipment at pulse load is like over-revving an engine. Components rated for steady highway operation are asked to redline. Batteries cycle faster, generators log more hours, compressors short-cycle. The depreciation clock on backup assets quietly compresses.

The mechanical anchor is Shannon Miller, founder and president of Mainspring Energy Inc., who told Bloomberg's Naureen S. Malik — syndicated by Business Standard — that, in her estimate, a gigawatt-scale campus is equivalent to a city the size of Boston and that roughly half its load can flicker on and off every few seconds. That is the load profile Uptime's mechanism operates on: not a higher mountain, but a faster heartbeat.

This is a bill-of-materials story, not an energy-shortage story. The hundred-billion-dollar AI build-out, already under lender and investor scrutiny, assumes backup gear depreciates on multi-year cycles. The pulse-fatigue frame says those cycles are shortening, turning a capex line into something closer to opex.

The falsifier is clean. If refresh intervals on backup gear hold steady through 2027, the frame fails. Until operators publish those intervals, the safe assumption is that the clock is faster than procurement.