A single downed power line exposed a growing AI data center problem. Here’s how to fix it.

A power line went down outside of Washington, DC, this week. Typically, the grid will only need a few seconds to recover from such an event. But this one took more than 10 minutes because more than 3 gigawatts of data centers stopped drawing power almost simultaneously.

The event caused the voltage across PJM’s grid to spike from Northern Virginia to Chicago, according to data collected by Ting Labs, a startup that uses a network of IoT sensors outside of people’s power sockets.

The event did not cause a blackout, but it did cause lights to flicker across the region. The incident showed the impact data centers can have on the grid – an effect experts believe will become more common.

Northern Virginia, located in the PJM region, is home to the largest number of data centers in the world.

“It’s the canary in the coal mine,” Ricardo de Azevedo, CTO at ON.Energy, told TechCrunch. These types of events involving large loads such as data centers “happen a lot,” he added.

The event echoes what happened two years ago, also on the PJM grid, and could foreshadow larger events if data centers are not designed to properly handle power supply disruptions. PJM Interconnection manages grids from New Jersey to Illinois and serves 67 million customers, making it the largest grid operator in the United States.

When a power line went down this week, it caused data centers to switch to backup power, and about 3.1 gigawatts of load disappeared in about 30 seconds, according to PJM data. The grid seemed to recover a bit, but after a while more loads dropped. At its peak, the PJM grid had an additional 3.49 gigawatts of electricity on it. It took another 11 minutes before it stabilized. The disconnected data centers represented about 3% of total demand at PJM during that period, according to Reuters.

A few percent may not sound like much, but the power grid needs to operate in a near-equilibrium state, with supply and demand closely matched. If they don’t, the voltages can decrease or increase. The grid and the devices connected to it can tolerate small fluctuations, but if that fluctuation grows too much, it will create failsafes within the grid or within individual utilities, causing them to disconnect.

When data centers in Northern Virginia felt fluctuations caused by a failed power line, they switched to backup power, taking their load off the grid. As more data centers make the switch, they take large amounts of load off the grid. What started as a small drop in supply turned into a big drop in demand, sending inventory soaring and causing the lights to flicker.

Most data centers make decisions in a split second, and the ones that shut down this week appear to be no different. When the power dip hit them, they all decided to disconnect for a few seconds on their own, Ali Zain Banatwala, senior market model specialist at the Independent Electricity System Operator, told TechCrunch.

“We need to find a way for these loads that are placed next to each other to disconnect or reconnect,” he said. A more streamlined process would allow grid operators to develop robust processes in advance.

Alternatively, data centers can be built to absorb disruptions and not turn their backs on them. One startup, ON.Energy, has been working on a product to help data centers – and the grid – ride out events like the one that happened this week.

The company has established an uninterrupted power supply for the entire data center of the facility, which includes not only servers but also chillers and other equipment. The company actually hides the data center behind a bank of batteries connected to a complex power converter. All the grid “sees” is one constant, well-behaved load rather than peaks and valleys from each part of the data center. The ON.Energy system allows data centers to jump computing workloads up and down, including AI training, without disrupting the grid.

Perhaps more importantly, it also means data centers can absorb power fluctuations from the grid. Rather than disconnecting from the grid, the ON.Energy system can use any excess power to charge its batteries, and if the flow drops, the system can send power to the servers. Also, it can track the grid lead within milliseconds, preventing reverses or surges like the one that caused this week’s problem at PJM.

ON.Energy currently installs a total of 3 gigawatts of its systems on four data center campuses, de Azevedo said.

Grid managers also woke up to the problem.

ERCOT, for example, will need large loads such as data centers to “survive” the disruption, de Azevedo said.

However, the clock is ticking. The mass outage this week was twice as large as the same event in 2024, when 60 data centers went out simultaneously, pulling 1.5 gigawatts of load on the grid. At the time, data centers accounted for 6% of PJM’s workload, according to Synapse Energy Economics. By 2040, they are expected to make up 24%. If the problem is not solved quickly, things may get worse.