A single downed power line outside Washington, DC this week caused more than 3 gigawatts of data center load to disappear from the PJM grid in roughly 30 seconds, sending voltage spiking from Northern Virginia to Chicago and taking the grid more than 10 minutes to stabilize. PJM Interconnection, the largest grid operator in the United States, serves 67 million customers across a footprint stretching from New Jersey to Illinois. The disconnected data centers accounted for about 3% of total PJM demand at the moment of the fault.
The mechanics matter. When the line failed, data centers in Northern Virginia — home to the highest concentration of data centers in the world — sensed the voltage dip and switched to backup power almost in unison. About 3.1 gigawatts of load vanished in 30 seconds. The grid briefly appeared to recover, then more facilities dropped off. At peak, PJM had an extra 3.49 gigawatts of electricity on a system that had suddenly lost its buyers, and it took another 11 minutes before supply and demand rebalanced.
Ricardo de Azevedo, CTO at ON.Energy, called the incident "the canary in the coal mine" and said events involving large loads like data centers are "happening more and more." The scale confirms it: this week's mass disconnection was twice as large as a similar 2024 event, when 60 data centers simultaneously disconnected and pulled 1.5 gigawatts off PJM.
“It's the canary in the coal mine”— Ricardo de Azevedo, CTO at ON.Energy
Key facts
- 01A single downed power line outside Washington, DC triggered 3.1 gigawatts of data center load to vanish from the PJM grid in about 30 seconds.
- 02Grid recovery took more than 11 minutes, with 3.49 gigawatts of excess electricity on the system at peak.
- 03The event was twice as large as a 2024 incident, when 60 data centers dropped 1.5 gigawatts simultaneously.
- 04Data centers made up 6% of PJM's load in 2024 and are projected to reach 24% by 2040.
- 05ON.Energy is installing 3 gigawatts of grid-buffering battery systems across four data center campuses.
The trajectory is steeper than the current numbers suggest. Data centers made up roughly 6% of PJM's load in 2024, according to Synapse Energy Economics. By 2040, they are expected to account for 24% — a fourfold share increase on a grid that is already the largest in the country. A 3% load event today becomes a 12% load event on the same relative footprint at that point, with far less margin for the grid to absorb it.
The core engineering problem is that data centers make protective decisions in milliseconds and make them independently. When a voltage dip propagates through Northern Virginia, hundreds of facilities each decide, in isolation, to switch to backup power within the same few seconds. The grid experiences that as a coordinated demand collapse even though no one coordinated it. What starts as a small supply shortfall inverts into a much larger demand shortfall, and voltages surge in the other direction.
Ali Zain Banatwala, senior market models specialist at the Independent Electricity System Operator, framed the fix bluntly. Grid operators need a protocol so co-located facilities disconnect and reconnect in sequence rather than as a swarm, giving operators predictable behavior to plan around.
“We need to figure a way for these loads that are located next to each other to sequentially either disconnect or reconnect”— Ali Zain Banatwala, Senior market models specialist at the Independent Electricity System Operator
The alternative approach is to make data centers stop reacting to grid disturbances at all. ON.Energy is building uninterruptible power systems that cover an entire data center campus — servers, chillers, and supporting equipment — behind a bank of batteries and power conversion hardware. The grid sees one steady, well-behaved load instead of a jagged composite. When the grid dips, the batteries dispatch power inward; when it surges, they absorb it. The system can track grid conditions within milliseconds, so a fault that would have triggered a mass disconnect instead gets swallowed by the buffer.
The same architecture also decouples AI training workloads from the grid. A data center can ramp compute up or down without pushing that variability onto the utility. ON.Energy is currently installing 3 gigawatts of these systems across four data center campuses, de Azevedo said — a footprint roughly equal to the load that dropped off PJM this week.
Regulators are starting to respond. ERCOT, the Texas grid operator, is moving to require large loads including data centers to "ride through" disruptions rather than reflexively disconnecting, de Azevedo said. PJM has not yet announced comparable rules, and the incident this week will sharpen the pressure to do so. The 2024 event drew attention but limited action; a repeat at twice the scale, 18 months later, is harder to file away.
The lights did not go out this week — bulbs flickered, and the grid held. But the pattern is the concerning part. Data center load is the fastest-growing category on PJM's system, driven almost entirely by AI training and inference buildouts, and the facilities are being sited in dense clusters that maximize the coordinated-disconnect risk. Each new campus that comes online without ride-through capability adds to the size of the next event.
The infrastructure story that matters for the AI market is not whether there's enough power — it's whether the power that exists behaves predictably when a hyperscale AI campus is sitting on top of it. The companies building AI data centers have optimized for latency, cooling, and cost per megawatt. Grid civility has been someone else's problem. That's about to change: either through mandates like the one ERCOT is drafting, or through liability exposure the first time a cascade like this week's actually causes a blackout. Buffering hardware from vendors like ON.Energy is one path; sequencing protocols coordinated with grid operators is another. Both cost money the industry has so far preferred to spend on GPUs. The bill is now arriving.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
