Powering AI is an architecture problem

Details

Ashburn, Virginia hosts the world's largest data center cluster, and it has now experienced two cascading failures from single fault events: a 2024 surge arrester failure dropped 60 facilities and 1,500 MW at once; the July 22, 2026 transmission fault knocked out over 3 GW in seconds. The pattern is not random — it is structural. AI data centers behave differently from traditional industrial loads. An AI campus can swing 70% of its load in milliseconds during a training run, then trip offline at the first upstream fault to protect billions in compute hardware. Each action is rational in isolation. At gigawatt scale, the synchronized behavior creates cascades the grid's protection logic amplifies rather than contains. A 2024 Virginia failure traced directly to protection schemes that count voltage dips and disconnect on the third one — exactly as designed, at the worst possible moment. The article proposes a three-part architectural fix: move power delivery from 480V to medium voltage (13.8kV+); eliminate eco-mode bypass that exposes racks to raw grid transients and sends AI load swings out unfiltered; and rewrite protection logic so data centers function as grid stabilizers rather than dropout loads. A new wave of multi-gigawatt AI campus interconnections is landing on this unchanged architecture. Note: this article is sponsored by ON.energy and published in MIT Technology Review. The proposed solutions align with ON.energy's product positioning, which should be factored into how the recommendations are weighted.