TL;DR · 30-second read
The Short Version
When one power line failed near Washington, DC, a wave of data centers (the warehouse-sized buildings full of computers that run the internet and AI) unplugged from the power grid within seconds and switched to their own backup power.
Together they had been using about 3% of the electricity on a grid that serves 67 million people. Losing that much demand at once threw the system off balance. Lights flickered from Virginia to Chicago, and it took more than 10 minutes to settle.
It was twice as big as a similar scare in 2024, and data centers keep multiplying. The proposed fix is to make them stay connected through brief hiccups, often by using giant batteries.
A single power line failure outside Washington, DC, this past week caused a wave of Northern Virginia data centers to switch to backup power at nearly the same moment, TechCrunch reported. According to PJM data, about 3.1 gigawatts of demand disappeared from the PJM Interconnection grid in roughly 30 seconds. After a partial recovery, more facilities dropped off, and the grid’s surplus peaked at 3.49 gigawatts. Stabilizing took another 11 minutes. A line fault like this one normally takes only a few seconds to recover from.
According to Reuters, the disconnected load was about 3% of PJM’s demand at the time. Sensor data from the startup Ting Labs showed voltage spiking across the grid from Northern Virginia to Chicago. There was no blackout, but lights flickered across the region. The event was roughly twice the size of a 2024 PJM incident in which 60 data centers disconnected at once and removed 1.5 gigawatts.
Executive Summary
Most discussion of data centers and the power grid asks whether there is enough electricity to feed them. This event shows the opposite failure: a cluster of very large customers that all leave the grid at once when they sense trouble. Each facility’s protective logic behaved as designed and switched to on-site backup power when voltage dipped. Together, those individually sensible decisions turned a routine line fault into a regional disturbance that lasted more than ten minutes.
Two remedies are being proposed. The first is coordination. Ali Zain Banatwala, a senior market models specialist at the Independent Electricity System Operator, argues that neighboring loads need a way to disconnect and reconnect one after another rather than all together. The second is ride-through, which means keeping facilities connected through brief disturbances. Ricardo de Azevedo, CTO of ON.Energy, whose company sells campus-scale battery systems built for that purpose, said ERCOT, the Texas grid operator, will require large loads to ride through disruptions.
The stakes rise with the load. Synapse Energy Economics put data centers at about 6% of PJM’s load in 2024 and projects 24% by 2040. The potential size of a synchronized dropout grows with that share unless the facilities are built to behave differently.
AI’s Other Grid Problem Is Load That Leaves
An electric grid has to keep supply and demand in near-perfect balance from moment to moment. When a line goes down, the grid normally absorbs the loss and rebalances within seconds. This time, the voltage dip from the fault reached many data centers sitting close together. Their protection systems are designed to make split-second decisions to safeguard the computers inside, and they moved the facilities onto backup power within a few seconds of each other. The grid lost about 3.1 gigawatts of demand in roughly 30 seconds. Generators were suddenly producing more electricity than customers were using, and voltage jumped.
The second wave is the detail grid planners should focus on. PJM’s grid appeared to recover partway before more loads dropped off, pushing the surplus to a peak of 3.49 gigawatts. In that pattern, the disturbance triggers disconnections, and the disconnections deepen the disturbance. That feedback is what stretched a seconds-long event into an 11-minute one. A gigawatt is roughly the output of a large power plant. In effect, the grid had to absorb the sudden departure of several power plants’ worth of customers.
That is why the grid risk from the data center buildout is not only a question of supply. Operators are used to planning around the sudden loss of a large generator. Clustered data centers create a mirror-image contingency: a large block of customers lost all at once because of a single triggering event. Everyone on the grid is exposed to that. Households saw flickering lights, generators must ramp down fast, and the operator must hold capability in reserve for a contingency that grows as data center load grows.
Concentration Turns a Local Fault Into a Regional Event
Northern Virginia has the highest concentration of data centers in the world. That density is what made the disconnection synchronized, because one voltage dip reached many facilities almost at once. The effect traveled well beyond Virginia. Ting Labs runs a sensor network through household electrical outlets, and it recorded the voltage spike as far away as Chicago. PJM covers the territory from New Jersey to Illinois, serves 67 million customers, and is the largest grid operator in the United States.
The trend line gives the event its weight. In 2024, 60 data centers disconnected together and removed 1.5 gigawatts. This time the figure was about 3.1 gigawatts, roughly double. Data centers were about 6% of PJM’s load then, and Synapse Energy Economics projects 24% by 2040. Two events are not a statistical series, and nothing here shows the next one will double again. The arithmetic still matters: if protective behavior stays the same while the share of load it governs roughly quadruples, the ceiling on a synchronized dropout rises with it.
De Azevedo called the event “the canary in the coal mine” and said incidents involving large loads are “happening more and more.” That claim comes from a vendor of the remedy. The two PJM incidents, with their measured sizes, are the documented reference points.
Two Fixes: Choreograph the Exit, or Don’t Leave
The first approach accepts that data centers will disconnect but asks them to do it in order. “We need to figure a way for these loads that are located next to each other to sequentially either disconnect or reconnect,” Banatwala said. Staggered behavior would let grid operators write procedures in advance instead of reacting to a cliff. The hard part is coordination. The facilities belong to different owners, each has settings tuned to protect its own uptime, and they would need shared thresholds or a shared signal.
The second approach is ride-through, meaning the facility stays connected through brief disturbances. ON.Energy’s version is an uninterruptible power supply, the battery backup that keeps servers running through outages, sized for an entire campus, including chillers and other equipment as well as servers. Batteries and power-conversion equipment sit between the grid and the facility, so the grid sees one steady, predictable load. The company says the system:
- charges its batteries when the grid has surplus power;
- discharges to the servers when supply dips;
- follows the grid within milliseconds;
- smooths the swings of AI training workloads that ramp power up and down.
It says it is installing 3 gigawatts of these systems across four campuses.
Those claims are specific enough to test. The test is whether a campus behind such a system stays connected and grid-friendly during a real fault like this one. The economics matter too. Sizing batteries and power electronics for a whole campus adds capital cost that the developer bears, while most of the benefit goes to the grid. Mandates may be what shifts that calculation, such as ERCOT’s ride-through requirement, which de Azevedo described.
What Developers and Buyers Should Expect
From a tenant’s point of view, the July event worked: the data centers switched to backup, and the computers presumably kept running. The cost of that decision fell on the grid and everyone else connected to it. This asymmetry is what makes the event a policy question rather than just an engineering one. Individual facilities have little reason to change behavior that protects their own uptime unless interconnection rules require it.
For developers, the likely direction is toward grid connection terms that specify how a facility behaves during a disturbance, not just how much power it draws. ERCOT is moving that way, according to de Azevedo. PJM, where both incidents occurred, is where the question is now most visible. For suppliers of storage, power conversion and switchgear, the grid-facing layer of a campus becomes part of the design requirement. For colocation buyers, it is reasonable to ask providers how their facilities respond to voltage dips and whether ride-through requirements could change facility design or pricing.
Background
PJM Interconnection is a regional transmission organization, the independent body that runs the high-voltage grid and wholesale electricity market for an area stretching from New Jersey to Illinois. It serves 67 million customers and is the largest grid operator in the United States. Its territory includes Northern Virginia, which has the highest concentration of data centers in the world.
Data centers are built never to lose power. When grid voltage wavers, they switch to on-site backup, such as uninterruptible power supplies and generators, in a split second. That design protects the computers inside. When many facilities share the same stretch of grid, however, it also means they can react to the same disturbance at the same moment. PJM saw this in 2024, when 60 data centers disconnected simultaneously, and saw it again at twice the scale this July. ERCOT, the operator of most of Texas’s grid, is among the grid managers moving to require large loads to ride through disturbances. Source: One fallen power line exposed a growing AI data center problem — here’s how to fix it (TechCrunch): a report on the roughly 3-gigawatt data center disconnection on the PJM grid and proposed ride-through fixes.Sources