A single downed power line exposed a deepening vulnerability in AI data center grid stability on 22 July, when more than 3.1 gigawatts of data centre load vanished from the PJM Interconnection in roughly 30 seconds, sending voltage surging from Northern Virginia to Chicago and causing lights to flicker across several states.
The line had tripped in the Ashburn area, the heart of the world’s densest concentration of data centres. The grid would normally recover in seconds. This time, according to ElectronEconomics, it took more than 10 minutes to stabilise. The cause was not a utility switching off the facilities: each data centre’s own internal controls detected the voltage disturbance and automatically transferred load to backup power. When enough of them did so simultaneously, the cure became a second problem.
At its peak, the PJM grid carried a surplus of 3.49 gigawatts. It took another 11 minutes before conditions normalised. The disconnected facilities represented around 3% of total demand on the system at the time, according to Reuters. Three percent sounds modest until you understand that the grid must maintain near-perfect balance between supply and demand at all times: even small mismatches trigger protective failsafes, and failsafes can cascade.
A Repeat Offence, Now Twice as Large
The July event was the second of its kind on PJM’s network. In July 2024, around 60 data centres disconnected simultaneously, pulling approximately 1,500 MW from the grid. The load loss this time was more than double that. Data centres accounted for roughly 6% of PJM’s load when the 2024 event occurred, according to Synapse Energy Economics. By 2040, that share is projected to reach 24% of total load.
The scale of the underlying growth is difficult to overstate. Synapse modelling projects PJM data centre consumption rising from 50 TWh in 2023 to 350 TWh by 2040, with data centres accounting for 49 GW of peak load, a 20% increase above the base-case peak. Meanwhile, PJM’s own projections have summer peak usage climbing by around 70 GW over the next 15 years, to 220 GW. The record summer peak for the entire PJM footprint was 165 GW in 2006.
The fiscal and environmental consequences compound the reliability risk. The Synapse factsheet puts the increase in net-present-value system costs attributable to data centre loads at $160 billion from 2025 to 2040, with near-term residential electricity bill increases of around 10%. Data centres are estimated to pay around 18% of PJM-wide energy costs by 2040, equivalent to approximately $23 billion annually.
AI Data Center Grid Stability and the Coordination Problem
The core engineering difficulty is that data centres make autonomous decisions at machine speed. Ali Zain Banatwala, senior market models specialist at the Independent Electricity System Operator, told TechCrunch: ‘We need to figure a way for these loads that are located next to each other to sequentially either disconnect or reconnect.’ A staggered response would give grid operators time to deploy countermeasures; a synchronised one does the opposite.
Ricardo de Azevedo, chief technology officer at ON.Energy, described the July event as ‘the canary in the coal mine,’ and told TechCrunch that events involving large loads are ‘happening more and more.’ His company has developed an uninterruptible power supply designed to cover an entire data centre campus, including servers, chillers and ancillary equipment. The system places a bank of batteries and sophisticated power-conversion equipment between the grid and the facility, so that what the grid sees is a single, stable load rather than the aggregate of thousands of independent switching decisions. ON.Energy is currently installing a total of 3 GW of its systems across four data centre campuses, de Azevedo said.
Regulators are moving, if not yet at the pace the grid’s trajectory demands. PJM’s Market Monitor has recommended that PJM establish a load queue for new large data centre connections, and stated plainly that PJM should not continue interconnecting large data centre load if that load cannot be served reliably. In February 2026, PJM filed a compliance submission with the Federal Energy Regulatory Commission in response to a December 2025 order, proposing three new transmission services for co-located loads willing to limit withdrawals under certain conditions, as outlined by White & Case.
Texas is further along. ERCOT was tracking approximately 226 GW of large load seeking interconnection as of November 2025, of which around 73% are data centres. Under rules effective from June 2025, ERCOT’s updated interconnection standards cap the load that can be lost from any single transmission line or generator trip at 1,000 MW per connection point, a direct attempt to prevent the kind of synchronised mass disconnection that struck PJM this month.
If the 2024 event doubled by 2026, the arithmetic for 2030 or 2035 is the one that should be keeping grid operators up at night. The next test of whether the fixes are working will almost certainly come before anyone is ready for it.
