PJM seeks ~15 GW to meet AI demand

PJM runs the power grid for 67 million people across all or parts of 13 states and the District of Columbia, and it now says it needs about 15 gigawatts of new electricity supplies to avoid shortages as artificial intelligence data centers pile onto the system. That is roughly the output of 15 large nuclear reactors or about 15 million homes’ worth of average electric demand. (pjm.com) (bloomberg.com) This is not a normal “more servers, more power bills” story. PJM said on April 10 that it wants to match proposed data centers with new power plants in a special process running from September 2026 through March 2027 because the usual queue is too slow for the load arriving now. (bloomberg.com) The pressure point is summer 2027. Bloomberg reported PJM is already bracing for a multi-gigawatt supply shortage by then, which means the grid operator is planning around a problem that is less than 18 months away, not a distant 2035 forecast. (bloomberg.com) The strange part is that PJM actually cut some of its near-term demand forecast in January 2026 after it started vetting data-center projects more strictly. Even after that cleanup, Utility Dive said PJM still expects summer peak load to grow 3.6% a year to about 222 gigawatts by 2036, adding about 65.7 gigawatts over 10 years. (utilitydive.com) Most of that jump is coming from data centers. Data Center Dynamics reported that PJM had earlier forecast peak load growth of 32 gigawatts from 2024 to 2030, with 30 gigawatts of that tied to data centers, which tells you how heavily one industry is now steering grid planning. (datacenterdynamics.com) PJM’s board spent January building a new rulebook for this exact problem. It proposed a path for large new customers to bring their own generation, a “connect and manage” option that lets them hook up earlier but face curtailment in tight conditions, and a backstop procurement process for short-term reliability needs. (pjm.com) That “bring your own generation” idea is the key shift. Instead of a data-center developer waiting years for the grid to find spare capacity, the developer can show up with a dedicated gas plant, battery system, or other supply and make the case that the new load will not crowd everyone else off the system. (pjm.com) (lw.com) The likely winners are projects that can be built fast. Data Center Dynamics said PJM’s earlier fast-track review selected 51 projects totaling 11.8 gigawatts, with gas-fired generation at 69%, battery storage at 19%, and nuclear at 12%, which is a much more practical mix for a three-year deadline than waiting for massive new transmission lines and slower generation builds. (datacenterdynamics.com) Consumers are already seeing the cost of this scramble. Data Center Dynamics reported PJM’s latest capacity auction pushed wholesale capacity prices up 22% from the prior year, with residential bills in some places expected to rise more than 5% over the next year and Columbus, Ohio, facing an average increase of $27 a month. (datacenterdynamics.com) This is why artificial intelligence infrastructure is starting to look less like cloud software and more like steel mills or aluminum smelters. OpenAI paused part of its Stargate project in the United Kingdom this week, and Network World reported that electricity costs were one of the main reasons, which is the same constraint PJM is now trying to solve in the United States. (networkworld.com) The practical result is that the next big artificial intelligence campus will not go wherever land is cheap or tax breaks are generous. It will go where a grid operator can prove there is power, where a generator can be financed for 15 years, and where the local system can survive a heat wave without telling the new customer to power down. (bloomberg.com) (pjm.com)