3,900MW gone in a minute: how AI training's power jitter is destabilizing the grid
FinanceYF5 · x · 2026-10-15
- On July 22, 2026, a minor fault in Virginia knocked 3,900 MW of data center load off the grid in one minute — the largest such event in PJM history.
- Citing a Microsoft/OpenAI/NVIDIA paper (arXiv 2508.14318), the issue isn't AI's total power draw but training's volatility: tens of thousands of GPUs ramping in the same second, each swing worth tens to hundreds of MW.
- The near-disaster came from facility self-protection: three voltage dips in a minute triggered cascading backup power switchover, scaling load from 1,500 → 1,800 → 3,900 MW.
- The author also previews analysis of four mitigation layers, who profits, and China's position in the compute-power race.
More from Infra
- NVIDIA to showcase neural rendering, physical AI and robotics at SIGGRAPH Asia 2026 — tomaszbednarz · 2026-10-15
- Google's TPUv7 Posts 1.67x Better Perf per Dollar Than GB300 NVL72 in SemiAnalysis InferenceX — scaling01 · 2026-10-15
- Microsoft's MAI-Image-2.6 and Flash land on Vercel AI Gateway with one API key — lmoroney · 2026-10-15
- Inworld opens up its consumer-AI inference API, citing tens of trillions of tokens a month — thetripathi58 · 2026-10-15
- When will 16GB machines run local LLMs on par with Opus-class models? — bit_freak · 2026-10-15
- vLLM P/D prototype hits 91.24% prefix cache hit rate on 12× GB300 — vllm_project · 2026-10-15