Grid planning for AI: PJM’s 15 GW idea
PJM proposed adding 15 gigawatts of new power to meet AI data‑center demand and avoid shortages, showing grid operators are being asked to plan capacity specifically for compute loads. That kind of proposal signals long lead times and coordination needs between utilities, developers and regulators. (x.com)
PJM runs the high-voltage grid and wholesale power market across all or parts of 13 states and the District of Columbia, and on April 10 it said it wants 15 gigawatts of new power supplies to keep up with artificial-intelligence data centers. The plan would pair proposed data centers with new power plants in a special process running from September 2026 through March 2027. (pjm.com) (bloomberg.com) Fifteen gigawatts is utility-scale power, not a rounding error. PJM’s own board said in August 2025 that peak demand in its region could grow by 32 gigawatts from 2024 to 2030, with almost all of that growth coming from data centers, while the system’s recent peak was about 153 gigawatts. (utilitydive.com) A data center is just a warehouse full of computers, but an artificial-intelligence data center behaves more like a new factory town than a normal office park. Thousands of graphics-processing chips run day and night, so the electricity draw is large, steady, and hard to shed at 5 p.m. when the grid is already stressed. (bloomberg.com) (insidelines.pjm.com) That is why PJM is moving away from the old assumption that demand just shows up and the market sorts it out later. In January, PJM’s board said new large loads may need to bring their own new generation or accept earlier curtailment, which means the computers could be told to power down first when the system is tight. (insidelines.pjm.com) (pjm.com) PJM is also talking about a backstop generation procurement process, which is grid-operator language for “buy extra supply before the shortage shows up.” Its January plan bundled that idea with faster forecasting, more state involvement, and a quicker interconnection track for some projects. (insidelines.pjm.com) (pjm.com) The politics around this have gotten unusually direct. On January 16, the Trump administration and a bipartisan group of governors asked PJM to hold a one-time emergency auction so data-center owners could sign 15-year power purchase agreements for new plants, with the data centers paying whether they used the power or not. (utilitydive.com) That matters because PJM normally buys capacity through auctions held years ahead for the whole system, not by matching one class of customer to one new block of generation. The new idea treats artificial-intelligence campuses less like ordinary demand and more like a giant industrial project that needs its own supply plan before it plugs in. (ferc.gov) (utilitydive.com) (bloomberg.com) The pressure is landing on a grid that already serves about 67 million people and was already warning about reliability and rising wholesale costs before this latest artificial-intelligence wave. PJM’s January filing said the supply-demand imbalance was already pushing up costs that can hit consumer bills. (pjm.com) (insidelines.pjm.com) So the real story is not just “more power for more servers.” It is that grid operators, utilities, regulators, governors, and data-center developers are now trying to schedule power plants, curtailment rules, and market contracts around compute demand on a 2026-to-2027 clock, because waiting for the normal grid process now looks too slow. (bloomberg.com) (insidelines.pjm.com)
