The Day PJM Broke Its Demand Record

At 5:55 p.m. on July 2, the largest power market in the United States was using effectively everything it had. Air conditioners were pulling harder just as power plants were failing in the heat. PJM had recalled maintenance outages, called demand response, raised every available generator, and obtained federal permission for some plants to run beyond their environmental permit limits.

The meter reached 162,713 megawatts. But that was not the record.

The record was the electricity PJM estimates would have been consumed if it had not paid customers to stop consuming it. Add the curtailed demand back and the system wanted 168,158 MW between 5 and 6 p.m. - 2,595 MW above a mark that had stood since 2006. The estimate remains preliminary while PJM completes its demand-response accounting, but the shape of the event is already clear in the public grid data.

Ask the Grid's hourly series reaches 162,569 MW at 5 p.m., almost exactly matching PJM's 162,713 MW instantaneous reading at 5:55. The difference is measurement: our line is an hourly average assembled from PJM's load areas; PJM's figure is the single highest instant.

Neither number includes all the electricity consumers wanted. PJM estimates that demand response delivered 6,113 MW of reduction on July 2. It is not correct to simply add that number to the metered peak - the reductions have different baselines and are not all coincident - but demand-response performance is what allowed PJM to estimate an unrestricted peak of 168,158 MW.

That distinction matters. PJM did not serve 168 GW through wires at one moment. It kept the system balanced while serving 162.7 GW and making a material block of demand disappear on purpose.

The public operations log lets us rebuild the critical 37 minutes. At 5:18 p.m., PJM declared a Maximum Generation Emergency Action for the Mid-Atlantic and Southern regions. Every available online unit was to move toward emergency maximum output.

At 5:32 p.m., PJM issued an Emergency Use of Back-Up Generator Warning. Large loads in the two regions were being told to prepare for the possibility that they would have 15 minutes to disconnect from utility power and move onto their own generators.

At 5:36 p.m., it issued a Deploy All Resources Action specifically for BGE, PEPCO and Dominion. Generators were ordered to full output, offline generation was to start as quickly as possible, and dispatched load-management customers were to reduce consumption immediately. The instantaneous peak arrived 19 minutes later.

The word warning is doing crucial work.

Three days earlier, PJM had asked the Department of Energy for unusual authority: if the grid reached the last step before voltage reduction or forced outages, PJM wanted to direct large loads with backup generation to leave the grid. The application specifically discussed data centers and hyperscalers. DOE granted that authority in Order 202-26-33.

But authority is not activation. PJM's public emergency log contains the 5:32 warning and no Emergency Use of Back-Up Generator Action. PJM's own post-event report is explicit: the action was not needed and was never issued.

The public record therefore does not support the claim that PJM ordered data centers onto diesel generators on July 2. It supports a narrower and more revealing statement: the system got close enough that operators warned large loads to be ready.

This does not mean behind-the-meter generators stayed idle. PJM's separate, market-based demand-response programs were already running. Pre-emergency resources began reducing load at 1:30 p.m. Emergency products followed later, including resources PJM defines as behind-the-meter generators whose permits require an EEA2 condition before they can operate.

The difference is between a registered resource performing under an existing demand-response contract and the federal last-resort order aimed at large loads. The first happened. The second did not.

Public data gives the program-level result, but not the customer list. It does not say which participants were data centers, how many megawatts they curtailed, or which facilities started engines. A data center could have participated as registered demand response without being ordered to do so under the federal authority. PJM does not publish enough detail to separate those cases.

While demand moved down, the generation stack moved up. At the 5 p.m. hourly peak, natural gas supplied 70,435 MW, nuclear plants 31,704 MW, and coal 28,761 MW. Solar was still producing 11,477 MW. The unusual resource was oil.

At 6 a.m. Eastern, oil was supplying 639 MW. Eleven hours later it was producing 6,194 MW, and it reached 6,425 MW at 7 p.m. Oil is expensive and emissions-intensive, and much of this fleet exists for exactly these hours: when the system has exhausted cheaper options.

The stress was not evenly distributed. PJM's deploy-all-resources action named BGE, PEPCO and Dominion, the same Mid-Atlantic corridor where heat, outages and transmission constraints converged.

Day-ahead prices had already anticipated trouble. In BGE, the July 2 hourly price reached $1,598.77/MWh. PEPCO reached $1,533.62/MWh. The PJM-wide reference reached $1,219.91/MWh an hour later. Most days in the preceding week averaged tens of dollars per megawatt-hour at PJM-RTO.

This is evidence of wholesale system stress, not a direct measure of anyone's electric bill. Retail customers buy through utilities, suppliers, hedges and regulated procurement plans. Emergency costs may also be allocated later through uplift. The event created expensive wholesale hours; the public data cannot turn those hours into a universal household bill increase.

Demand alone did not create the emergency. Supply was eroding underneath it. PJM reported 18,363 MW of generation outages at the July 2 peak, compared with roughly 12,800 MW on the ten highest summer peak days of the previous three years. Across July 2-4, forced outages ranged from 18.1 to 19.4 GW. Plant equipment was the largest reported cause, followed by ambient-air limitations and other operating problems.

DOE's other order, 202-26-32, addressed the supply side. It covered 594 units at 51 plants, about 20.4 GW of installed capacity, whose environmental permits or operating restrictions could limit them during the emergency.

PJM now says 94 units at 17 plants, representing approximately 3.25 GW, actually ran beyond permit limits under that authority. Together they accumulated 1,107 unit-hours above those limits.

That is an unusually concrete environmental finding. It is also the boundary of the public record. PJM says unit owners reported waiver-dependent operations daily, but the unit-level notices are not posted alongside DOE's order. DOE publishes changing lists of eligible units; those lists are not proof that every listed plant ran or exceeded a permit.

AirNow monitors recorded unhealthy-for-sensitive-groups ozone in parts of Maryland on July 2, including AQI 136 at Edgewood. Those readings cannot attribute a monitor result to the federal waiver. Heat and sunlight drive ozone formation, and power plants are only one source among several. The supportable conclusion is that plants operated beyond permit limits. The unsupported one is that the order caused a particular community's bad-air day.

Data centers are not incidental to PJM's problem. They are one of its largest new facts. PJM's May forecast-accuracy report puts the 2026 data-center July peak at 11,479 MW, up from 8,167 MW observed in July 2025 - an increase of about 41 percent in one year. The forecast is equivalent to 6.8 percent of a peak the size of 168,158 MW.

The load is geographically concentrated. Dominion's 2026 data-center July peak forecast is 7,066 MW. AEP accounts for 2,699 MW. The rest of PJM combined accounts for roughly 1,714 MW. Northern Virginia is therefore central to the region's new load shape and to the transmission constraints that matter on its hottest afternoons.

But a structural driver is not the same as an event trigger. PJM does not publish actual data-center demand for 5:55 p.m. on July 2. It does not publish a site-by-site curtailment list. It does not identify operators or owners among the 6,113 MW of demand response. The forecast can tell us how large the sector has become, not which server hall created the final megawatt of the emergency.

PJM got through the hottest regional day it has measured without voltage reduction, manual load shed, or the special backup-generator action. It did so by calling every available generator, paying roughly 6.1 GW of demand to move out of the way, and allowing 94 units to operate beyond permit limits. Emergency actions concentrated in the corridor where data-center growth is largest.

That is more consequential than saying data centers caused an emergency. It shows exactly how close the system came, exactly which levers operators pulled, and exactly where public transparency stopped.

We know the grid warned large loads. We know registered behind-the-meter resources were eligible and deployed within demand-response programs. We know 17 plants exceeded permit limits. We know PJM's forecast data-center peak is almost 7 percent of a system peak this size.

We still do not know which data centers reduced load, which ones started generators, or what they emitted. For a resource PJM may increasingly depend on during emergencies, that is not a small missing field. It is the next story.

The charts above are frozen from the PJM data behind Ask the Grid at publication. Open PJM on the main map and replay July 2, 2026.