A 4,000-megawatt wake-up call
In July, a nearly 4,000 MW data-center load in Northern Virginia tripped offline during a transmission disturbance — roughly two nuclear plants' worth of demand vanishing from the grid in seconds. In May, NERC had already raised its highest-level (Level 3) reliability alert, telling operators to treat AI campuses as grid actors, not passive customers. In July, PJM — the country's largest grid, spanning 13 states — proposed cutting power to data centers during shortages. In mid-October, FERC rejected that plan, ruling it may violate the Federal Power Act and handing the mess to states and utilities.
The signal isn't subtle: power, not silicon, is becoming AI's hard ceiling.
The bottleneck is moving
For two years I watched boards obsess over GPUs and talent. That era is ending. PJM alone forecasts 70 GW of new large-load demand by 2038. Grid operators can't interconnect that fast, so they're doing what utilities always do under constraint — pricing risk and rationing.
Expect three shifts:
- Interruptibility becomes standard. New large-load customers will be asked to shed demand on short notice. If your contract doesn't specify curtailment terms today, your state utility will write them for you later.
- Location becomes a power arbitrage. Colocation decisions will be driven by where generation is surplus and curtailment risk is priced low, not just where fiber is cheap.
- Efficiency becomes capacity. Smaller on-device models, off-peak training, and inference that can flex down aren't cost niceties anymore — they're the difference between running and waiting.
Recompute the denominator
Every AI decision now has a kilowatt denominator. Audit where inference runs and whether it can shed load without breaking the business. Put power availability and curtailment terms into RFPs and SLAs now. The grid won't wait for your roadmap — it's already writing its own.