The Grid Wasn't Ready: How a Single Downed Power Line Exposed AI Data Centers' Hidden Threat

On an otherwise ordinary day this week, a power line fell outside Washington, D.C. Under normal circumstances, the electrical grid would shrug off such an event in seconds. Instead, operators at PJM Interconnection, the nation's largest grid manager, stretching from New Jersey to Illinois, spent more than 10 minutes fighting to stabilize a system suddenly flooded with excess electricity.

The culprit wasn't the downed line itself. It was what happened next: more than 3 gigawatts of AI data center load vanished from the grid in roughly 30 seconds.

A Cascade in Northern Virginia

Northern Virginia hosts the densest concentration of data centers on Earth. When the power line failed, voltage fluctuations rippled across PJM's network, from Northern Virginia to Chicago. Sensors deployed by Ting Labs, a startup monitoring electrical outlets in homes across the region, captured the disturbance. Lights flickered. Voltage spiked. And though no blackout occurred, the message was unmistakable.

"It’s the canary in the coal mine," said Ricardo de Azevedo, CTO at ON.Energy, in an interview with TechCrunch. Events involving massive, synchronized loads like data centers, he warned, are "happening more and more."

The mechanics of the failure are telling. When data centers sensed the voltage dip caused by the initial line failure, they did exactly what they were designed to do: they switched to backup power, disconnecting from the grid to protect their servers. But with dozens of facilities making that same decision within seconds of each other, the cure became worse than the disease.

Each disconnection removed demand from a grid already struggling to rebalance. The result was a supply surge at its peak, an extra 3.49 gigawatts of electricity with nowhere to go. The disconnected load represented roughly 3% of total PJM demand at the time. That may sound minor, but the electrical grid is a precision instrument. Supply and demand must remain in near-perfect equilibrium; even small mismatches cause voltage to sag or spike. When those fluctuations grow large enough, they trigger cascading failsafes, forcing more facilities offline and deepening the crisis.

"We need to figure a way for these loads that are located next to each other to sequentially either disconnect or reconnect," said Ali Zain Banatwala, senior market models specialist at the Independent Electricity System Operator. Without an orderly process, grid operators cannot develop robust procedures to manage these events in advance.

Déjà Vu, But Bigger

This wasn't PJM's first warning. In 2024, a similar incident saw 60 data centers simultaneously disconnect, yanking 1.5 gigawatts from the grid. This week's event was more than twice as severe. The trend line is sobering: in 2024, data centers accounted for about 6% of PJM's load. By 2040, projections suggest they will consume 24%.

If facilities continue to treat grid disturbances as a signal to flee rather than absorb, the consequences will escalate from flickering lights to something far more serious.

The Case for "Riding Through"

The solution, according to a growing chorus of engineers and grid operators, isn't to prevent data centers from protecting themselves; it's to change how they protect themselves.

One approach gaining traction is the concept of "ride-through" capability. Rather than disconnecting at the first sign of trouble, advanced data center designs can absorb fluctuations and continue operating as stable grid citizens.

ON.Energy, for example, has developed an uninterruptible power supply architecture scaled to entire data center campuses. The system places a sophisticated battery bank and power conversion layer between the facility and the grid. The grid sees one steady, predictable load rather than the volatile demand spikes of individual servers and chillers. When the grid surges, the system charges its batteries. When it dips, the batteries dispatch power to keep servers running. The response time is measured in milliseconds.

The company is currently installing 3 gigawatts of this technology across four data center campuses coincidentally, roughly the same amount of load that dropped off PJM's grid this week.

Regulators are also moving. ERCOT, which manages Texas's independent grid, is preparing to require large loads like data centers to demonstrate ride-through capability rather than simply disconnecting during disruptions.

The Infrastructure Reality Check

The AI boom has been measured in model parameters, training compute, and capital investment. But the limiting factor may prove to be something far more analog: the stability of the electrical grid that powers it.

Data centers are no longer passive consumers of electricity. At gigawatt scale, clustered in tight geographic regions, they have become active participants in grid dynamics capable of amplifying small disturbances into regional crises. The question is no longer whether data centers will affect grid stability, but whether they will be engineered to help stabilize it or inadvertently destabilize it.

This week's incident was a warning shot. With data center load projected to quadruple its share of PJM's demand by 2040, the window for building smarter, grid-friendly facilities is narrowing. The next downed power line may not end with flickering lights.