AI data centers hit power limits

The constraint on new artificial intelligence data centers is increasingly electricity: getting enough power, at the right site, on the right timeline. (epri.com) The Electric Power Research Institute said on February 26, 2026 that United States data centers could consume 9% to 17% of national electricity by 2030, up from about 4% to 5% today. Its report said artificial intelligence workloads already account for an estimated 15% to 25% of data-center electricity use. (epri.com) The U.S. Department of Energy’s 2024 recommendations said hyperscale facilities are now seeking 300- to 1,000-megawatt connections, with lead times of one to three years straining local grids. The department’s advisory group said generation, storage, grid technologies and closer coordination with utilities are now central to meeting that demand. (energy.gov) A data center is a warehouse of servers, and the newest artificial intelligence servers draw far more electricity than older cloud-computing racks. The International Energy Agency said on April 10, 2025 that there is “no AI without energy,” specifically electricity for data centres. (iea.org) That power has to reach a specific parcel of land through substations, transmission lines and utility approvals, not just exist somewhere on the grid. The Department of Energy said those connection requests are stretching local systems, and EPRI said permitting, siting and transmission delays could constrain both generation and grid buildout. (energy.gov) (epri.com) The bottleneck is showing up first in regional queues. In Texas, the Electric Reliability Council of Texas said planners were assessing more than 233 gigawatts of large-load interconnection requests near the end of 2025, with more than 70% coming from data centers. (utilitydive.com) Kristi Hobbs, the grid operator’s vice president of system planning and weatherization, told the Electric Reliability Council of Texas board on December 9, 2025 that the large-load queue had grown almost 300% from a year earlier. She said the system had “outgrown the process” used to review those requests one by one. (utilitydive.com) Virginia shows how concentrated the problem can get. EPRI said Virginia is already the only state where data centers use more than 20% of electricity, and that share could rise to 39% to 57% by 2030. (epri.com) The growth forecasts remain steep even with those physical limits. The International Energy Agency projects global electricity demand from data centers will more than double to about 945 terawatt-hours by 2030, with artificial intelligence as the biggest driver of that increase. (iea.org) The practical result is that the race to build more computing power now runs through transformers, substations, batteries, gas plants, wind and solar projects, and new transmission lines. The companies that can secure power first are the ones most likely to expand first. (epri.com) (energy.gov)