A Data Center Digest Original StoryThis article was researched and written entirely by AI, without human review or editing, as part of the Data Center Digest's ongoing experiment in AI-powered journalism.

At about 7:56 a.m. on July 22, 2026, a mechanical failure took a 230-kilovolt line out of service in Northern Virginia. The grid handled the fault correctly. The data centers did not wait to find out. About 3,800 megawatts of their load disconnected within moments and moved to on-site backup generation, the largest such transfer in PJM's history, according to PJM. Over the same year, the federal government went the other way and asked data centers to run those same generators so the grid could stay up. In 2026 the data center was, in the space of a few months, both a source of grid trouble and a reserve for it. This piece sets out what happened, what the rules say about running the generators, and the alternative of asking data centers to use less.

When data centers leave on their own

The July event was the third of its kind in PJM's Dominion zone, the territory that holds Northern Virginia's data center cluster. PJM says similar events of about 1,500 megawatts each occurred on July 10, 2024, and February 17, 2025. The 2024 event began when a lightning arrestor failed on a 230-kilovolt line during a thunderstorm. According to GridStatus, six rapid voltage drops followed, and NERC found that data centers shifted from grid power to uninterruptible power supply systems. The facilities stayed off the grid for hours, and reconnecting them took manual intervention. Voltage and frequency moved outside normal ranges, and utilities heard complaints about flickering lights and noisy appliances, the Union of Concerned Scientists noted. NERC's president called it a “five-alarm fire when it comes to reliability.”

The 2026 event was larger. PJM reports the first wave moved 2,970 megawatts to backup generation as system voltage rose, and a second wave moved another 1,099 megawatts, causing further voltage disturbances. PJM's total load fell from 99,984 megawatts to 96,205 megawatts. The grid operator reported no significant reliability impacts and restored its area control error limit within nine minutes, against a NERC standard of 30. Operating Committee Chair Emanuel Bernabeu said the fault was cleared normally and the data centers “are too sensitive” to voltage changes, disconnecting “too early.”

Regulators are now writing rules. On July 16, FERC ordered NERC to submit the first phase of enforceable standards for computational loads by December 31, 2026, according to PJM, and minimum voltage and frequency ride-through requirements, which would keep a facility connected through a brief disturbance, are expected in 2027. PJM is also evaluating interconnection requirements for large loads, including ride-through standards, Utility Dive reported. Until those exist, an emergency that starts with one transmission fault can end with several thousand megawatts of demand disappearing at once.

When the grid asks them to run the generators

The second thread runs in the opposite direction. In December 2025, Energy Secretary Chris Wright argued that backup generators at data centers were sitting idle while utilities struggled to meet demand. “We have 35 gigawatts of backup generators that are just sitting there today,” he said, according to E&E News. In late January 2026, as Winter Storm Fern brought severe cold, DOE sent emergency orders to PJM and Duke Energy that could require data centers to run backup diesel generators, covering about a dozen states. PJM did not need to use the measure during that storm, Data Center Knowledge reported.

DOE issued a second order on May 18, 2026, running through May 20, after PJM requested it the day before. The agency cited a heat wave and transmission and generation outages tied to seasonal maintenance, and repeated its estimate of more than 35 gigawatts of unused backup generation nationwide. Another order covered the early-July heat wave. PJM issued backup-generator warnings on July 2 and July 3, the first of their kind, and no generators were dispatched, Data Center Knowledge reported. PJM called the warnings precautionary. Served peak load reached 162,713 megawatts, and unrestricted demand, which counts emergency actions and customer load reductions, was estimated at 168,158 megawatts, an all-time unrestricted peak. Independent Market Monitor Joseph Bowring said PJM “was getting very close to a shortage of primary reserves.” The coverage reviewed for this piece did not report that any data center generators ran in the January or May events either, so the 35 gigawatts remains an estimate of what is available, not a record of what has been used.

What the air rules allow

Running an emergency generator outside a true emergency runs into air rules. EPA's current framework, summarized by Sidley, sets no hourly cap on emergency operation and limits nonemergency operation to 100 hours a year. Demand response under a financial agreement with a third party counts as nonemergency, with up to 50 of those hours allowed if the local balancing authority or transmission operator dispatches the engine and the operator keeps records, among other conditions. In a May 2025 fact sheet, EPA said grid operators can require customers to start generators during demand response events, and that an Energy Emergency Alert Level 1 can justify emergency demand response when escalation to Level 2 is imminent. The guidance is nonbinding, and EPA said it is evaluating more substantive changes to the underlying regulations.

States are moving the other way on permits. Virginia's environmental agency said the “limited operating hours” basis for exempting emergency generators from stricter controls “is no longer appropriate,” and expects most future data center backup fleets to be permitted as non-emergency units. For permit applications filed on or after July 1, 2026, it set a presumptive limit of 0.60 grams per horsepower-hour, which effectively requires selective catalytic reduction. A separate law, HB 507, requires emissions limits at least as strict as the federal Tier 4 standard. The numbers behind the dispute differ by who is counting. A state legislative review found data center generators account for less than 4% of regional nitrogen oxide emissions, and the Data Center Coalition cited a finding that they often run 10 to 30 minutes a month. The Piedmont Environmental Council counts thousands of diesel generators representing more than 11 gigawatts of backup capacity in Virginia. The Southern Environmental Law Center said tighter controls are an improvement but may let data centers run generators far more often. Compass Data Centers said it keeps using older Tier 2 engines because the catalyst systems on Tier 4 units can slow or shut down an engine if the controls fail during an outage.

The alternative: use less, not generate more

A different approach asks data centers to cut demand briefly. A Duke University study from 2025 estimated that the 22 largest U.S. balancing authorities could add 76 to 126 gigawatts of new flexible load if it could be curtailed for 0.25% to 1% of the time, or roughly 85 to 366 hours a year, Utility Dive reported. PJM, at 18 gigawatts, had the most headroom at a 0.5% curtailment rate, ahead of MISO at 15 and ERCOT at 10. The study did not account for transmission constraints, which could shrink the figure.

Some operators are trying it. Google announced demand response agreements with Indiana Michigan Power and the Tennessee Valley Authority in August 2025, the first to cover machine learning workloads, which can be shifted away from peak periods. Google's Michael Terrell wrote that “there are limits to how flexible a given data center can be,” because services such as Search and Maps need reliability. Texas made curtailment a condition of connection. Senate Bill 6, signed in June 2025, requires large loads of 75 megawatts or more that connect to ERCOT to install shutoff equipment, so utilities can disconnect them during firm load-shed events, and creates a separate voluntary program that pays loads to reduce demand with at least 24 hours' notice.

The trade-offs

The two approaches pull against each other. Asking a data center to run generators adds local emissions and relies on equipment built for rare outages. Asking it to curtail leaves the facility's own reliability promises in doubt. Staying connected through a disturbance, the goal of ride-through standards, makes a facility less likely to add to the problem but also keeps a large load on a stressed grid. Virginia has added a wrinkle: according to Data Center Knowledge, a separate state law bars operators from meeting voluntary demand flexibility standards by running carbon-emitting backup generators, so in that state the generator route and the flexibility route cannot count as the same thing.

For a community near a campus, the practical questions are narrow. Is the site covered by a shutoff or curtailment obligation? Are its generators permitted as emergency or non-emergency units? Will it ride through the next voltage dip, or trip off? The answers now differ by state and by grid operator.

What to watch

NERC's first-phase standards for computational loads are due by December 31, 2026, with ride-through requirements expected in 2027. EPA could change the underlying generator rules. Texas is still building the programs SB 6 created. And next summer will test whether the warnings PJM issued in July 2026 stay precautionary.

States in this story