A Fallen Power Line Exposes AI Data Centers’ Grid Instability

A minor power line failure outside Washington, D.C., this week escalated into a significant grid disturbance, exposing a critical vulnerability caused by the concentrated power demands of modern data centers, particularly those powering AI. The event, which caused lights to flicker across a region stretching to Chicago, was triggered when over 3 gigawatts of data center load vanished from the grid in seconds as facilities switched to backup power. While it did not cause a blackout, experts warn it is a harbinger of more frequent and severe instability as data center electricity consumption is projected to quadruple on some grids by 2040.

Key takeaways

• A downed power line caused data centers to disconnect nearly 3.5 gigawatts of load from the PJM grid, leading to a voltage spike and a recovery time of over 10 minutes.
• The mass disconnection, representing about 3% of PJM’s total demand at the time, was twice the size of a similar event in 2024, highlighting a rapidly growing problem.
• Data centers in Northern Virginia, the world’s densest concentration, made split-second decisions to switch to backup power, removing their load and destabilizing the grid’s delicate supply-demand balance.
• Startups like ON.Energy are developing campus-wide battery systems to “hide” data centers from the grid, allowing them to absorb fluctuations and ramp AI workloads without causing disruptions.
• Grid operators, including ERCOT, are moving to require large loads to “ride through” power disruptions, but the clock is ticking as data center load on PJM is forecast to reach 24% by 2040.

The cascading failure

When the power line failed, it created a voltage dip across the PJM Interconnection, the largest grid operator in the United States. Data centers, programmed to protect their sensitive computing hardware, sensed the fluctuation and began autonomously switching to their backup power systems. According to PJM data, about 3.1 gigawatts of load disappeared in roughly 30 seconds. This initial disconnection was followed by additional loads dropping off, peaking at 3.49 gigawatts of excess electricity suddenly on the grid. The electrical grid operates on a knife-edge of balance; this sudden, massive drop in demand caused a corresponding voltage spike, manifesting as flickering lights across the region. Experts note that such events are becoming more common as data center capacity expands.

Engineering a solution

The core issue is the simultaneous, uncoordinated reaction of thousands of data center systems. Ali Zain Banatwala, a senior market models specialist, emphasized the need for these large, co-located loads to disconnect or reconnect in a sequential, orderly fashion to give grid operators time to respond. A more robust solution, however, may be to engineer data centers to absorb grid disturbances rather than instantly abandoning it. ON.Energy is deploying such a system: a campus-scale uninterruptible power supply that uses banks of batteries and sophisticated power conversion equipment. This system presents a single, stable load to the grid and can use excess power to charge batteries or dispatch power during a dip, all within milliseconds. The company is currently installing 3 gigawatts worth of these systems at four data center campuses.

Why it matters

The incident is a stark warning for the AI boom’s infrastructure. Northern Virginia’s data center alley is a bellwether, and the problem is scaling rapidly. The 3.49-gigawatt disconnection this week was double the 1.5-gigawatt event from just two years prior. With data center load on the PJM grid projected to soar from about 6% in the recent past to 24% by 2040, the potential for catastrophic grid instability grows exponentially. If thousands of data centers continue to disconnect en masse during minor disturbances, they risk triggering widespread blackouts. The push for solutions—from grid operator mandates for “ride-through” capability to innovative battery buffering—is now a race against time to prevent the very engines of the digital economy from destabilizing the power networks they depend on.