One fallen power line exposed a growing AI data center problem. Here’s how to fix it.
TechCrunch Tim De Chant
A downed power line near DC caused 3+ gigawatts of data centers to bail off the grid at once, flickering lights from Virginia to Chicago. It's a preview of bigger grid trouble as AI data centers keep multiplying.
Based on reporting by TechCrunch, Tim De Chant — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
A single fallen power line outside Washington, DC this week should have been a non-event, the kind of blip the grid shrugs off in seconds. Instead it took more than 10 minutes to stabilize, because over 3 gigawatts of data centers in Northern Virginia sensed the disturbance and, almost in unison, switched to backup power and dropped off the grid. Voltage swung wildly from Virginia to Chicago. Lights flickered. Nobody lost power outright, but the episode exposed a fragility that grid experts have been quietly worried about for a while.
The mechanics are almost absurdly simple. PJM Interconnection, which runs the largest grid in the US and serves 67 million people, needs supply and demand to stay in near-perfect balance. When that line failed, roughly 3.1 gigawatts of data center load vanished within 30 seconds as facilities' failsafes kicked in. The grid tried to recover, then more load dropped, pushing the surplus to 3.49 gigawatts before things settled 11 minutes later. That's about 3% of PJM's total demand disappearing almost instantly — not because of a blackout, but because the data centers protected themselves the moment they smelled trouble.
Ricardo de Azevedo, CTO of ON.Energy, calls it the canary in the coal mine. He's not exaggerating for effect. This is the second time in two years a PJM disturbance has triggered a mass data center disconnection, and the first one, in 2024, only pulled 1.5 gigawatts from 60 facilities. This week's event was double that size. Ali Zain Banatwala of Ontario's grid operator points to the real design flaw: data centers clustered together all make the same split-second decision within seconds of each other, with zero coordination. Nobody engineered a sequence for disconnecting or reconnecting gracefully. They just all bail at once.
ON.Energy thinks it has a fix, and it's not about asking facilities to behave better during a crisis — it's about hiding them from the grid entirely. The company builds a battery-backed power buffer for an entire data center campus, chillers and all, so the grid only ever sees one smooth, predictable load. Instead of disconnecting during a dip, the system can draw down batteries to keep servers running, or soak up excess supply by charging them, responding within milliseconds. De Azevedo says ON.Energy is currently rolling out 3 gigawatts of this gear across four campuses. ERCOT in Texas is also moving to force large loads to
My take — AI-written commentary, not fact-checked reporting
The grid math here is brutal and not going away: data centers were 6% of PJM's load last year and are projected to hit 24% by 2040, and that growth curve is entirely AI-driven, so treating this as a one-off glitch would be foolish. What bugs me is that the fix — buffering campuses so they ride through disruptions instead of flinching — is a solved engineering problem sitting mostly unused while everyone races to build more GPUs first and ask permission later. Regulators mandating ride-through behavior, like ERCOT is doing, should have happened years ago, not after the second near-miss.
Read more about this at: TechCrunch