US warns of AI power strain

The U.S. warning on AI is no longer just about chips. It is now about whether the grid can serve warehouse-size data centers fast enough. (politico.com) (belfercenter.org) A data center is a building full of servers, plus the cooling gear and backup equipment needed to keep those computers running around the clock. Training and running artificial intelligence models happens inside those facilities, so more AI means more electricity, more cooling and more grid connections. (iea.org) (energy.gov) The Department of Energy said U.S. data centers consumed about 176 terawatt-hours in 2023, up from 58 terawatt-hours in 2014. Its estimate for 2028 is 325 to 580 terawatt-hours, equal to roughly 6.7% to 12% of total U.S. electricity use. (energy.gov) The International Energy Agency said on April 16 that electricity demand from data centers rose 17% in 2025, with AI-focused sites climbing even faster. It also said supply chains for transformers, gas turbines, advanced chips and other equipment have tightened over the past year. (iea.org) The U.S. pinch points are physical. A new Harvard Belfer Center brief said AI-driven load is already outpacing available capacity in some regions, pushing companies to delay projects, buy power directly and install on-site gas generators. (belfercenter.org) That brief pointed to a July 2024 incident in northern Virginia, where a voltage fluctuation disconnected 60 data centers at once. The sudden drop created a 1,500-megawatt power surplus and forced emergency action to avoid wider outages. (belfercenter.org) Federal regulators are now trying to change how these giant loads get onto the system. Politico reported on April 20 that the Federal Energy Regulatory Commission set itself a June deadline for a proposal tied to Energy Secretary Chris Wright’s plan to speed AI infrastructure and decide who pays for grid upgrades. (politico.com) FERC has already ordered PJM, the grid operator serving 67 million people across 13 states and Washington, D.C., to write rules for AI-driven data centers co-located with power plants. The commission said the aim is to protect reliability and consumers while making service rules more transparent. (ferc.gov) The Energy Department is also trying to shorten the build timeline. It said 16 federal sites are being evaluated for data center and energy infrastructure partnerships, with a target for selected projects to begin operating by the end of 2027. (energy.gov) The supply response is spreading beyond wind and solar. The International Energy Agency said the tech sector signed about 40% of all corporate renewable power purchase agreements in 2025, while conditional offtake deals tied to small modular reactor nuclear projects grew from 25 gigawatts at the end of 2024 to 45 gigawatts this month. (iea.org) The policy argument has widened with the infrastructure strain. Commerce Department officials have previously tied export controls to access to advanced computing power for military and strategic uses, and the current grid fight shows that “compute” now also depends on power, land and cooling that governments can no longer treat as abundant. (commerce.gov) (iea.org) The U.S. warning, then, is not that AI might need more power someday. It is that regulators, utilities and federal agencies are already rewriting grid policy because the load is here now. (politico.com) (energy.gov)