Power, not chips, is the AI bottleneck
Researchers and industry analysts say the biggest constraint on scaling AI is now electricity, siting and physical infrastructure — not just processors or models. That shift raises new capital and operational questions: utility exposure, grid capacity, and where to site datacenters as energy becomes the binding constraint on AI growth. (brookings.edu) (apnnews.com)
A year ago, the artificial intelligence race looked like a hunt for more chips. In 2026, the scarcer item is often the electricity to turn those chips on. (csis.org) A data center is just a warehouse full of computers, and every computer turns electricity into calculations and heat. When companies add more artificial intelligence servers, they do not just buy processors; they need substations, transmission lines, transformers, backup systems, and enough power plant output to feed the whole site every hour. (brookings.edu) The International Energy Agency said on April 10, 2025 that global data center electricity use could rise to 945 terawatt-hours by 2030 from 415 terawatt-hours in 2024. That is roughly the electricity use of Japan today, coming from a sector that was once treated as a rounding error in grid planning. (iea.org) (spglobal.com) The United States is where this gets tight fastest. The Electric Power Research Institute projects U.S. data centers could use 9% to 17% of national electricity by 2030, up from 4% in 2024, depending on how aggressively artificial intelligence keeps expanding. (powering-intelligence.epri.com) That changes what “capacity” means in artificial intelligence. A company can have money, land, and a purchase order for the best graphics processing units, but if the local grid cannot deliver another 300 megawatts, the machines sit in boxes. (csis.org) Researchers and grid planners now use a blunt phrase for the new race: speed-to-power. It means the winning site is not the place with the cheapest dirt or the best tax break, but the place that can get real electricity first. (csis.org) That is why artificial intelligence companies are suddenly acting like utilities. Google signed a deal with Kairos Power in October 2024 to support up to 500 megawatts of advanced nuclear capacity by 2035 for its data centers, and Microsoft signed a 20-year agreement in September 2024 tied to restarting Three Mile Island Unit 1 with about 835 megawatts. (blog.google) (kairospower.com) (constellationenergy.com) Amazon went even closer to the wire. Talen Energy said its March 2024 deal with Amazon Web Services covered a data center campus next to the Susquehanna nuclear plant, and the relationship was expanded in June 2025 to supply additional electricity through the grid for artificial intelligence and cloud operations in Pennsylvania. (talenenergy.com) The bottleneck is not only generation. Lawrence Berkeley National Laboratory found the median time from a grid interconnection request to commercial operation had stretched to 5 years for projects built in 2023, up from less than 2 years for projects built in 2000 through 2007. (lbl.gov) Geography is starting to shift with the power map. The Electric Reliability Council of Texas said in an April 2025 presentation that its 2025 long-term load forecast jumped sharply, with most of the increase tied to future data center demand, while Northern Virginia utilities and grid planners are filing special forecasts because data center load is overwhelming old assumptions. (ercot.com) (pjm.com) So the artificial intelligence business is starting to look less like pure software and more like railroads or steel. The hard question is no longer only who has the smartest model, but who can secure a gigawatt, wait out a five-year grid queue, and keep the lights on at industrial scale. (brookings.edu) (csis.org)
