One Fallen Power Line Yanked 3.1 GW Off the Grid in 30 Seconds. The AI Data Center Buildout Is Making It Worse.

One Fallen Power Line Yanked 3.1 GW Off the Grid in 30 Seconds. The AI Data Center Buildout Is Making It Worse.

When a power line fell outside Washington, DC, this week, data centers across the territory of PJM Interconnection, the largest grid operator in the United States, did what they were built to do: they protected themselves. In roughly 30 seconds, 3.1 gigawatts of demand vanished from the grid, sending voltage spikes from Northern Virginia to Chicago 1.

The grid held. No blackouts. Lights flickered across the region, and it took PJM more than ten minutes to restore balance. But the incident exposed a structural flaw at the heart of the AI buildout: the same facilities racing to add compute capacity are wired to abandon the grid at the first sign of trouble, and their synchronized exit makes every disturbance worse for the 67 million customers still connected to it 1.

How one downed wire cascaded

PJM Interconnection manages the grid from New Jersey to Illinois. Northern Virginia, within PJM's territory, holds the highest concentration of data centers in the world 1.

When the power line fell, it triggered a voltage dip. Data centers sensed the fluctuation and switched to backup power, nearly simultaneously. The electrical grid requires supply and demand to stay in near-perfect balance; when the data centers disconnected, their demand disappeared, and surplus electricity surged through the system. At its peak, PJM's grid carried 3.49 gigawatts of excess electricity. The disconnected load represented about 3% of total demand on PJM at the time 1.

Ali Zain Banatwala, senior market models specialist at the Independent Electricity System Operator, told TechCrunch that the data centers "decided to disconnect within a few seconds of each other" when the voltage dip reached them 1.

A few percent of total demand may sound marginal. On a grid that must balance supply and demand second by second, it is not.

Twice as bad, and accelerating

This week's disconnection was twice as large as a similar event in 2024, when 60 data centers simultaneously disconnected and pulled 1.5 gigawatts from PJM's grid 1.

The trajectory is what alarms grid operators. Data centers accounted for about 6% of PJM's load in 2024, according to Synapse Energy Economics. By 2040, they are projected to consume 24% 1. If today's data center load, at roughly 6% of the grid, can produce a 3.1-gigawatt demand shock, the same self-protective behavior at 24% could multiply that disruption roughly fourfold.

The behavior nobody designed the grid around

Every individual data center is acting rationally. Its servers need stable power, and backup systems exist to guarantee it. The problem emerges when facilities clustered in the same region run the same protection logic, so a single disturbance triggers a synchronized exit. Each departure removes more demand, widening the voltage swing, which can trigger additional disconnections in a cascading feedback loop.

"We need to figure a way for these loads that are located next to each other to sequentially either disconnect or reconnect," Banatwala said. An orderly process would let grid operators build procedures in advance rather than react to a sudden hole in demand 1.

Ricardo de Azevedo, CTO at ON.Energy, a startup developing campus-scale power systems, called the region "the canary in the coal mine." His company has built an uninterruptible power supply that buffers an entire data center campus behind batteries and power conversion equipment, so the grid sees one consistent, well-behaved load instead of the volatile demand of thousands of individual servers. ON.Energy is currently installing 3 gigawatts of the systems across four campuses 1.

Mandating resilience, at whose expense

Grid operators are starting to force the issue. ERCOT, which operates the Texas grid, is set to require large loads like data centers to ride through disturbances rather than disconnect, according to de Azevedo 1.

That mandate crystallizes the tension at the center of the AI buildout. Hyperscalers designed their facilities and negotiated their power contracts on the assumption that they can disconnect whenever supply wavers. Ride-through requirements flip that assumption: the largest electricity consumers would be obligated to absorb the grid's disturbances instead of fleeing them. The cost of the batteries, power conversion equipment, and engineering needed to buffer a campus against voltage swings does not disappear. It lands somewhere, whether in cloud pricing, in utility rate adjustments for everyone else on the grid, or in the political fight over whether private compute infrastructure should be subsidized as a public good.

The firms spending the most to build AI capacity are now the grid's biggest customers and its fastest-growing source of instability.