The US government is preparing to cut power to the country’s largest data centers, and AI is the reason. PJM Interconnection — the grid operator that keeps the lights on for 65 million Americans across 13 states — announced plans this week to force data centers consuming 50 megawatts or more to switch to backup generators whenever the grid runs short. The mandatory program doesn’t formally start until June 2027. The emergency authority to do it right now already exists. The Department of Energy has used it three times in 2026 alone.
This isn’t a fringe regulatory concern. It’s a structural failure playing out in real time.
The Incident That Exposed the Problem #
On July 22, a single transmission line fault in Ashburn, Virginia — ground zero for US data center density, with 209 completed facilities and 43 more under construction in the county — triggered a chain reaction. Data centers in the area detected the fault and automatically switched to backup power, yanking more than 3 gigawatts off the grid simultaneously. Sensors tracked voltage anomalies stretching from Washington DC to Chicago across 1.4 million monitoring points. It took 10 minutes to stabilize — orders of magnitude slower than the millisecond corrections grid operators normally execute for routine disturbances. Residents as far away as the DC suburbs reported flickering lights.
This was not planned. No human issued a curtailment order. Data centers simply did what they’re designed to do — protect their uptime. The grid absorbed the consequence.
The Numbers Behind the Crisis #
The Ashburn incident was a symptom of something deeper. In December 2025, PJM ran its capacity auction to line up enough generating resources for the coming years. It came up 6,625 megawatts short. Data centers now account for 94% of all new load growth on the grid, while new generation additions have stalled — just 525 megawatts added in the last auction against data center demand that doubles that figure every single year through 2032.
The pricing signal has been brutal. PJM capacity prices have gone from $28.92 per megawatt-day in 2024 to $333.44 in 2026 — an 11-fold increase in two years. Wholesale power prices in the region rose 75.5% year-over-year in Q1 2026. Gartner projects global data center electricity consumption growing 26% in 2026 alone. That cost has to go somewhere, and it’s going to ratepayers: residents in PJM territory are expected to see electricity bills rise around $70 per month by 2028.
Developers Have Already Felt This #
This isn’t hypothetical. On May 7, 2026, a thermal event at a Northern Virginia data center triggered one of the largest AWS outages in recent memory. More than 150 cloud services went down. EC2 instances and EBS volumes in availability zone us1-az4 were impaired. Coinbase went down. FanDuel went down. The root cause: overheating. Not a software bug, not a misconfiguration — a data center running at thermal limits, pushed there by power-hungry AI workloads and the physics of cramming that much compute into a finite space.
As grid stress increases, the pressure on cooling systems rises with it. The thermal margin that separates an operational data center from a May-7-style event keeps shrinking.
What Gets Cut, and Who’s Exposed #
Starting June 2027, PJM’s curtailment program targets facilities consuming 50 megawatts or more. That threshold is designed to protect smaller data centers and colocation tenants while catching the hyperscalers. AWS US-East-1, Azure East US, and Google Cloud us-east4 all sit in PJM territory. Northern Virginia is the heartland. Under the program, these facilities will receive between 30 minutes and a few days’ notice to switch to generators. They’ll be compensated. Your SLA probably won’t cover the gap.
Cloud SLAs are written to cover infrastructure failures within the provider’s control. Grid-forced curtailments are force majeure events. The fact that AWS and Azure both maintain extensive on-site generator capacity doesn’t mean seamless continuity — the May outage proved that power transitions aren’t instant or clean.
What Developers Should Actually Do #
The obvious answer — multi-region architecture — has been best practice for years. What’s changed is the urgency and geography. US-East-1 concentration risk is documented, real, and growing. If your production workloads are in Northern Virginia with no failover, that’s worth revisiting before June 2027.
Longer term, the power math is changing. AI training and inference workloads place sustained, high-density power demands on the grid that no previous era of cloud computing did. Some teams are already shifting compute-intensive training jobs to regions with better power availability, or scheduling them for off-peak hours. That used to be an optimization. It’s becoming a survival strategy.
Virginia is running on 10,500 diesel generators right now. The grid operator is buying time with emergency orders. The curtailments begin in less than a year.