A single power line failed outside Washington, DC this week, and the grid took more than 10 minutes to recover from something it normally shrugs off in seconds. The reason, as TechCrunch AI reports, was AI infrastructure: more than 3 gigawatts of data centers stopped drawing power at almost the same moment, spiking voltage across the largest grid in the country. No blackout followed. But lights flickered from Northern Virginia to Chicago, and that flicker is a warning.
This is significant because it’s the first time the data center load problem has shown up at this scale in a single, traceable event. Ricardo de Azevedo, CTO at ON.Energy, put it bluntly to TechCrunch: “It’s the canary in the coal mine.” These mass disconnections are “happening more and more,” he said.
What actually happened
When the line went down, data centers in Northern Virginia sensed the voltage dip and did what they’re built to do: switch to backup power to protect their servers. The problem is they all did it at once.
- About 3.1 gigawatts of load vanished from the PJM grid in roughly 30 seconds.
- At its peak, PJM had an extra 3.49 gigawatts of electricity sloshing around with nowhere to go.
- Stabilizing took another 11 minutes.
- The disconnected data centers were only about 3% of total demand at the time.
Three percent sounds small. It isn’t. The grid runs on near-perfect balance between supply and demand. When a big chunk of demand disappears in seconds, supply surges, voltage spikes, and failsafes start tripping. What began as a modest drop in supply snowballed into a much larger drop in demand.
Why it matters for the AI industry
Northern Virginia holds the highest concentration of data centers on the planet, and it sits inside PJM, which serves 67 million customers from New Jersey to Illinois. As AI training and inference workloads pile in, that concentration keeps growing.
The scale of the risk is climbing fast. Consider the trajectory, per data cited by TechCrunch AI:
- In 2024, a similar event knocked 1.5 gigawatts offline when 60 data centers disconnected together.
- This week’s event was twice as large.
- Data centers were about 6% of PJM’s load in 2024. By 2040 they’re projected to hit 24%.
The status quo was simple: data centers protect themselves first and let the grid sort out the rest. That worked when they were a rounding error. It doesn’t work when they’re a quarter of demand.
The fixes on the table
Experts point to two paths, and they aren’t mutually exclusive.
- Coordinated disconnection. Ali Zain Banatwala of the Independent Electricity System Operator told TechCrunch the goal is to get neighboring loads to “sequentially either disconnect or reconnect” instead of dropping off in unison. An orderly process lets grid operators plan ahead.
- Ride-through hardware. ON.Energy has built an uninterruptible power supply for an entire data center campus, covering servers, chillers, and everything else. It hides the facility behind a bank of batteries and power conversion gear, so the grid sees one steady, well-behaved load instead of jagged peaks and valleys. When the grid surges, the system charges its batteries. When it dips, it dispatches power to servers. It can follow the grid’s lead within milliseconds.
ON.Energy is now installing 3 gigawatts of these systems across four data center campuses, de Azevedo said. Regulators are moving too. ERCOT, the Texas grid operator, plans to require large loads like data centers to “ride through” disruptions rather than bailing out at the first sign of trouble.
What to expect next
Watch for “ride-through” requirements to spread from ERCOT to other grid operators, PJM included. If you’re building or siting AI compute, grid behavior is becoming a design constraint, not an afterthought. The engineering answer already exists. The open question is whether deployment keeps pace with demand.
The clock is the real story here. This week’s event was double the size of the one two years ago, and the load curve only bends upward from here. Get the fixes in early and these become footnotes. Wait, and the flickering lights turn into something a lot harder to reset. Full details are available at the original TechCrunch AI report.