A single downed power line near Washington, DC, turned into an object lesson in AI data centre grid stability, and why it is becoming one of the most pressing reliability problems facing electricity networks in the United States.
When the line failed, more than 3.1 gigawatts of data centre load vanished from the PJM Interconnection grid in roughly 30 seconds, according to PJM data. The grid operator, which serves 67 million customers from New Jersey to Illinois, is the largest in the country. It took more than 11 minutes to fully stabilise, and at its worst the network was carrying an excess of 3.49 GW it could not shed quickly enough.
Residents across Northern Virginia reported flickering lights and strange noises from air conditioners and refrigerators. One source told Reuters: ‘The power quality took a pretty big hit.’ PJM itself said it registered the sudden demand drop and a change in grid frequency but maintained there were no impacts to its overall reliability. Dominion Energy, whose network covers the affected area, told Data Center Dynamics: ‘Within minutes of the incident occurring, our system operations team was able to stabilize the situation and return to normal operating conditions.’
The data centres that disconnected represented around 3% of total PJM demand at the time, according to Reuters. That sounds modest until you consider that the grid must maintain near-perfect balance between supply and demand. A sudden loss of that much load is, electrically speaking, identical in its effect to losing a large power station, except it happened faster than almost any grid protection system is designed to anticipate.
Why AI Data Centre Grid Stability Is Getting Harder to Manage
What makes this incident particularly awkward is that it has happened before. A similar event in 2024 saw 60 data centres disconnect simultaneously, pulling 1.5 GW of load off the PJM grid. During that earlier episode, Grid Status analysis found PJM-wide frequency climbed as high as 60.047 Hz, well above the plus-or-minus 0.036 Hz band that NERC targets for stable operation. The current event was roughly twice the scale.
‘It’s the canary in the coal mine,’ Ricardo de Azevedo, CTO at ON.Energy, told TechCrunch. These kinds of events are ‘happening more and more,’ he added.
The trajectory makes that uncomfortable. Data centres accounted for around 6% of PJM load in 2024. A Synapse Energy Economics report projects that share will reach 24% by 2040, with data centre load growing from 50 TWh in 2023 to 350 TWh. By that point, data centres are projected to drive PJM peak demand up by 49 GW (a 20% increase) and annual energy consumption up by 313 TWh. The same report estimates data centres will be paying roughly 18% of PJM-wide system energy costs, or around $23 billion, by 2040.
Ali Zain Banatwala, senior market models specialist at the Independent Electricity System Operator, put the core problem plainly to TechCrunch: ‘We need to figure a way for these loads that are located next to each other to sequentially either disconnect or reconnect.’ Right now, most data centres make that call in a split second, and when a cluster of them all reach the same decision within seconds of each other, the cascade is essentially impossible for grid operators to pre-empt.
Ride-Through Rules and the Battery Buffer Approach
Texas has moved faster than most jurisdictions to address the problem. The Texas Public Utility Commission approved ERCOT’s ride-through rules for data centres after ERCOT staff warned that even a small fraction of the 438 GW of large load projects currently queued for study could trigger cascading outages if those facilities fail to ride through ordinary voltage or frequency disturbances. Under the approved ERCOT NOGRR 282 framework, any facility of 75 MW or more where at least half its demand is computational must meet specific thresholds: riding through two seconds at 80% voltage, half a second at 50%, and recovering to 90% of pre-disturbance load within one second. Utility Dive reported the rules are designed to reduce the reliability risk from large computational loads unexpectedly tripping during disturbances.
One company is betting on a hardware answer rather than waiting for regulation to catch up everywhere. ON.Energy has developed a whole-campus uninterruptible power supply that wraps batteries and power conversion equipment around an entire data centre site, so the grid sees a single, stable load rather than the volatile peaks and troughs generated by racks of servers, chillers, and cooling systems cycling independently. The system can absorb excess grid power into its batteries or dispatch stored energy to servers within milliseconds, smoothing out exactly the kind of voltage sag that triggered this week’s disconnections. De Azevedo said the company is currently installing 3 GW worth of its systems across four data centre campuses.
The incident is also drawing fresh attention from federal and state regulators already examining how to manage sudden demand swings from data centres and crypto mining operations, according to Data Center Dynamics. Whether PJM follows ERCOT’s lead with its own ride-through mandate will be the question to watch. The next event may not stop at flickering lights.
