PJM eyes 15 GW for AI data centres

The grid operator for 13 states and Washington, District of Columbia is preparing for a jump in electricity demand so large that it is talking about lining up 15 gigawatts of new power, mainly because artificial intelligence data centers are arriving faster than the system was built for. Bloomberg reported the proposal on April 10, 2026, and PJM Interconnection has already been warning that large new loads could threaten reliability if supply does not show up quickly. (bloomberg.com) (pjm.com) Fifteen gigawatts is not a rounding error. It is roughly the output of about 15 large nuclear reactors, or enough electricity for millions of homes, which is why this is being treated as an emergency planning problem instead of a normal forecast update. (bloomberg.com) PJM Interconnection is the air traffic controller for the power system across all or parts of Delaware, Illinois, Indiana, Kentucky, Maryland, Michigan, New Jersey, North Carolina, Ohio, Pennsylvania, Tennessee, Virginia, West Virginia, and Washington, District of Columbia. It says it serves about 67 million people, and its job is to make sure enough electricity is available every hour, not just on average over a year. (pjm.com 1) (pjm.com 2) The pressure point is not ordinary internet traffic. It is clusters of specialized chips for artificial intelligence training and inference, packed into warehouse-size buildings that can demand hundreds of megawatts each, which makes one campus look more like a steel mill than an office park. (ferc.gov) (whitecase.com) PJM’s own forecasts show why the operator is moving now. In its January 14, 2026 load report, PJM said summer peak demand is now expected to grow 3.6% a year to about 222 gigawatts by 2036, which is about 65.7 gigawatts of added demand over 10 years. (pjm.com) (utilitydive.com) That forecast actually got lower in the next few years and higher later on. PJM cut some near-term numbers after applying stricter screens to proposed data centers, but it still expects the long-term surge to be bigger than it thought a year earlier. (utilitydive.com) (pjm.com) The stricter screens matter because a lot of announced projects never become real power demand. In a February 25, 2026 presentation, PJM said it filtered projects based on the strength of their commitments, assumed at least 36 months for ramp-up, and used a 70% utilization rate unless a developer could support a higher number. (pjm.com) Even after that trimming, PJM’s board said on January 16, 2026 that it needed a backstop generation procurement process for short-term reliability needs. The board also said new large loads may need to bring their own generation or accept curtailment, which means a data center could be told to cut demand when the grid is tight. (pjm.com) Federal regulators are already involved because the old rulebook was not written for this kind of load. On December 18, 2025, the Federal Energy Regulatory Commission ordered PJM to create clearer rules for artificial-intelligence-driven data centers and other large loads that want to colocate with power plants, saying the existing tariff was too unclear to protect consumers and reliability. (ferc.gov) The fight underneath this proposal is about who pays and who waits. If a 500-megawatt or 1-gigawatt data center connects before enough new generation and transmission are ready, the system can end up leaning on older plants, tighter reserve margins, and higher capacity prices that spread across the region. (ferc.gov) (utilitydive.com) So the 15-gigawatt figure is really a signal that artificial intelligence is no longer a software story inside PJM territory. It is now a steel-and-concrete story about gas turbines, transmission queues, capacity auctions, and whether the power system can add the equivalent of several big cities’ worth of demand without pushing everyone else’s bills higher. (bloomberg.com) (pjm.com)