One Downed Power Line Exposed the Grid's Problem in the Age of AI Data Centers

This week a transmission line snapped not far from Washington. Normally the grid recovers from something like that within seconds, but this time it took more than ten minutes. The cause: more than three gigawatts of data center load that stopped drawing power almost simultaneously.

According to Ting Labs, a startup rolling out a network of IoT sensors that plug straight into household outlets, the event caused a voltage spike across the entire territory of grid operator PJM — from Northern Virginia to Chicago. It never escalated into a full-blown blackout, but lights flickered across the region. Northern Virginia, home to the world's largest concentration of data centers, sits squarely within PJM's footprint — the country's biggest grid operator, serving 67 million customers.

How a Small Glitch Became a Big One

The grid runs in a state of near-perfect balance: supply and demand have to match. When the line went down, the data centers detected the voltage fluctuation and switched to backup power, shedding roughly 3.1 gigawatts of load from the grid in about 30 seconds. The grid partially recovered, but then more facilities dropped off. At its peak, PJM was left with an extra 3.49 gigawatts, and stabilizing it took another eleven minutes. The data centers that went offline accounted for about 3% of total demand.

The problem is that these sites make their decisions in fractions of a second, and they do it without coordination: sensing the voltage dip, they all disconnected almost at once. “This is the canary in the coal mine,” says ON.Energy CTO Ricardo de Azevedo, noting that incidents like this involving large loads are happening more and more often. The event echoes a similar case from two years ago, but this time the scale was twice as large: in 2024, 60 data centers shed 1.5 gigawatts in unison.

What They Propose to Do

Experts see two paths. The first is to teach neighboring sites to disconnect and reconnect in turns, so operators have time to build procedures around it. The second is to design data centers so that they ride through disturbances rather than turning away from them.

The startup ON.Energy has developed an uninterruptible power supply for an entire data center campus — not just for the servers, but for the chillers and other equipment as well. In essence, the company tucks the site behind a bank of batteries with advanced power electronics: the grid “sees” a single smooth, predictable load instead of the jagged profile of individual nodes. Such a system can absorb fluctuations, charging when there's a surplus of energy and giving it back during a dip, and it adapts to the grid within milliseconds. The company is currently installing a total of 3 gigawatts of these systems across four campuses.

Regulators have gotten involved too: Texas grid operator ERCOT intends to require large loads to “ride through” disturbances. There isn't much time, though. Whereas in 2024 data centers accounted for about 6% of PJM's load, by 2040 their share is projected to climb to 24%. Without a solution to the problem, the scale of these incidents will only grow.

Source: TechCrunch