PJM Proposes Curtailing Power to New AI Data Centers During Grid Emergencies

The artificial intelligence boom has a physical footprint, and it is colliding with the hard limits of the American electrical grid. PJM Interconnection, the operator of the largest wholesale electricity market in the United States, has directed its staff to file a landmark proposal with federal regulators to manage the strain.[1]

PJM manages the flow of power for 67 million people across 13 states and the District of Columbia, a territory that includes Northern Virginia—the world's densest concentration of hyperscale data centers. As these facilities expand to train and run massive AI models, their power requirements are threatening to outpace the grid's ability to supply them.[2][3]

To protect residential and commercial customers from rolling blackouts, PJM's board of managers has approved a framework called the Interim Resource Adequacy Service. If approved by the Federal Energy Regulatory Commission, the policy will fundamentally change how the grid treats major computing loads starting in June 2027.[1][4]

The proposal establishes a strict "bring your own power" mandate for new, massive energy consumers. Any new facility drawing 50 megawatts or more—a threshold that captures virtually every modern AI data center campus—must prove it has secured sufficient independent or contracted power supplies.[2][6]

If a 50-megawatt facility fails to secure its own capacity, its electricity access becomes conditional. During a grid emergency, PJM will order these unhedged data centers to drop offline before it asks residential customers to conserve power or initiates rolling blackouts.[1][4]

This represents a two-tier reliability structure. Historically, grid operators have used Pre-Emergency Load Management—paying various industrial customers to voluntarily reduce consumption—as their first line of defense. Under the new rules, unhedged data centers will bear the first-mover risk, facing mandatory curtailment before those broader emergency measures are deployed.[4]

The mechanism of curtailment does not mean grid operators will casually switch off servers without warning. Affected data centers would be compensated for reducing demand, similar to existing demand-response initiatives. Facilities would typically receive an emergency signal and have roughly 15 minutes to transfer their operations to on-site backup power.[5][7]

For most data centers, that backup power comes from industrial diesel generators. While diesel is easy to source and store locally, it introduces significant environmental friction. Federal rules generally limit these generators to 50 hours of operation per year for demand-response programs, and local communities have increasingly pushed back against the localized air pollution they create.[5][7]

PJM's aggressive regulatory posture is driven by stark grid mathematics. The operator forecasts that new large loads will add approximately 70 gigawatts of demand to its system by 2038. Simultaneously, the grid is losing its older, baseload power sources; roughly 15 gigawatts of generation, primarily coal and older natural gas plants, has retired across the PJM footprint since 2022.[2][6]

This divergence between surging demand and retiring supply culminated in PJM's most recent capacity auctions, which failed to secure the system's required 20 percent reserve margin. To fix the immediate shortfall, PJM's board approved a second measure: a Reliability Backstop Procurement.[1][5]

The backstop mechanism will launch a specialized central capacity auction in September and October 2026, targeting 6,831 megawatts of new supply. To incentivize rapid construction, PJM is offering a price ceiling of $555 per megawatt-day and allowing selected power plants to lock in commitments for up to 15 years, provided they come online by 2032.[1][4]

The threat of data center curtailment is not purely theoretical. During severe heat waves in May and July of 2026, the U.S. Department of Energy granted PJM emergency authority to force data centers onto backup generators if necessary to prevent wider grid failures.[7][8]

On July 2, 2026, as peak demand approached an all-time record of 166,147 megawatts, PJM issued an emergency warning to utilities to prepare for the possible curtailment of data centers in Maryland and Virginia. While the grid stabilized and the final step was not needed, the event proved that data centers are now firmly inside the grid's emergency playbook.[7][8]

For the technology sector, the implications are profound. Cloud computing contracts are built around guarantees of continuous uptime and availability zones. PJM's policy signals that computing capacity can now only grow as fast as the physical power system can support it, unless tech companies are willing to accept interruptions.[8]

The June 2027 effective date gives data center operators a narrow window to adapt. Facilities must now decide which workloads can be paused or shifted to other regions during a power crunch, and developers are increasingly looking to fund their own dedicated power sources, such as advanced geothermal or small modular nuclear reactors, to guarantee their AI ambitions are not unplugged.[4][8]