Put the numbers on one denominator

A transmission line went down outside Washington. According to PJM data, data centers switched to backup power and about 3.1 gigawatts of load left the grid within thirty seconds; at its peak the system carried 3.49 gigawatts of surplus electricity, and stabilizing took eleven minutes rather than a few seconds. Data from Ting Labs, which measures through sensors plugged into household outlets, show voltage rising from Northern Virginia to Chicago. Citing Reuters, the report puts the disconnected load at about 3% of PJM demand at that moment; PJM serves 67 million customers.[1]

A 3% share sounds small and misleads if read on its own. What decides the outcome here is how quickly the load departed, more than how large it was. Ali Zain Banatwala, a senior market models specialist at the Independent Electricity System Operator, says the data centers disconnected within a few seconds of one another and that a method is needed for neighboring loads to disconnect and reconnect in sequence. The same 3.1 gigawatts withdrawn over minutes would have been a balancing problem; withdrawn in thirty seconds it produced a supply surplus, and lights flickered across the region.[1]

Who buys the buffer

The proposed fix is physical too. ON.Energy has built a campus-scale uninterruptible power supply covering not only servers but chillers and other equipment; the data center sits behind a bank of batteries and power conversion gear, and the grid sees a single well-behaved load. The company's chief technology officer, Ricardo de Azevedo, calls the event the canary in the coal mine and says they are installing 3 gigawatts of these systems across four data center campuses. According to de Azevedo, ERCOT will require large loads such as data centers to ride through disruptions.[1]

This is a constraint moving house: a problem that would be solved on the grid side with transmission and balancing investment turns into batteries and converters on the site side. Whoever pays for it as data center capital expenditure also gets the ability to ramp training workloads up and down without disturbing the grid. Another explanation is available: the same batteries can be justified purely by workload swings and continuity, in which case riding through a disturbance is a by-product of the purchase rather than its purpose.[1]

Where the scale is heading

We have a comparison point: in a similar event in 2024, 60 data centers disconnected at once and pulled 1.5 gigawatts; this week's departure was twice that size. According to Synapse Energy Economics, data centers were about 6% of PJM load at that time and are expected to reach 24% by 2040. If that share triples while the disconnection behavior stays the same, the next event will not be absorbed by balancing. The concrete threshold to watch: whether PJM publishes a binding ride-through requirement for large loads, as ERCOT is set to do. If no such requirement appears by the end of the year, the buffer stays a voluntary purchase and procurement decides where it gets built.[1]