AI Data Centers Dropped 3 GW at Once, Nearly Taking Down the PJM Grid

In July 2026, more than 3 gigawatts of data centers in Northern Virginia disconnected almost simultaneously from the PJM Interconnection grid — 3.1 GW of load vanished from the system in 30 seconds. That is twice the size of a similar failure in 2024 and one of the largest such incidents in the operator's territory.

What happened in Virginia

When a disturbance hit the grid, data centers near Washington switched en masse to backup power and stopped drawing electricity from PJM all at once. The trigger was a downed transmission line: the resulting voltage spike made the data centers' protection systems disconnect from the grid at the same time.

• Location — Northern Virginia, in the territory of PJM Interconnection (67 million customers from New Jersey to Illinois)
• More than 3 GW of data centers disconnected almost in sync
• 3.1 GW of load disappeared in roughly 30 seconds
• Peak generation surplus — 3.49 GW
• It took 11 minutes to stabilize the grid

Why this threatens the grid

A sudden load drop is dangerous because power plants keep generating electricity for demand that has already disappeared. The resulting surplus risks frequency spikes and cascading outages. According to PJM, data centers made up about 3% of total grid demand at the time of the failure, the surplus reached 3.49 GW, and it took 11 minutes to return to normal.

The problem is growing along with AI. Data centers accounted for roughly 6% of PJM's load in 2024, and the operator forecasts their share will rise to 24% by 2040. The 2026 incident was twice as large as the comparable 2024 event, when 60 data centers totaling 1.5 GW disconnected from the grid at once.

"This is a canary in the coal mine," said

Ricardo de Azevedo, CTO of ON.Energy; he says such events are "happening more and more often."

How engineers propose to fix it

Engineers propose making data centers "ride through" grid disturbances instead of dumping their load all at once. ON.Energy is installing uninterruptible power supply systems at four data center campuses totaling 3 GW; they smooth out disturbances without cutting load instantly. In Texas, the grid operator ERCOT already requires large consumers to ride through brief disturbances without disconnecting.

The problem is that current protection protocols tell data centers to disconnect at the first voltage deviation. Until rules like Texas's apply everywhere, every large campus remains a potential gigawatt-scale switch for the entire grid — and the more capacity gets built for AI, the higher the cost of a single synchronized drop.

"We need to figure out how to get these loads to disconnect and reconnect sequentially, rather than all at once," said

Ali Zain Banatwala, senior market modeling specialist at the Independent Electricity System Operator.

What this means

AI growth is constrained not only by chips but also by the physics of power grids: gigawatt-scale data centers behave like one giant switch, and without new rules a synchronized load drop can destabilize a grid serving 67 million customers.

Frequently asked questions

What happened to the data centers in Northern Virginia?

In July 2026, more than 3 GW of data centers near Washington disconnected almost simultaneously from the PJM grid, shedding 3.1 GW of load in 30 seconds; it took 11 minutes to stabilize the grid.

Why do data centers disconnect from the grid en masse?

During a voltage spike — for example, from a downed transmission line — data centers' protection systems switch them to backup power. When dozens of facilities do this at once, the grid loses gigawatts of demand within seconds.

How do engineers want to solve the problem?

The proposal is to require data centers to "ride through" brief disturbances without disconnecting. ON.Energy is installing UPS systems totaling 3 GW at four campuses, and Texas's ERCOT already requires large loads to operate in ride-through mode.