AI data centres strain grids

AI data-centre builders are no longer waiting for the grid; they are bringing power to the site or buying land that can get it faster. (datacenterknowledge.com) At Data Center World 2026 in Washington, D.C., Prologis Mobility vice president JT Steenkamp said U.S. operators are pairing on-site generation with large real-estate portfolios to close the gap between AI demand and available electricity. He said the U.S. could face about 200 gigawatts of new AI load by 2030 while retiring about 104 gigawatts of existing power capacity. (datacenterknowledge.com) Chris Crosby, chief executive of Compass Datacenters, wrote in POWER on April 22 that utilities and data-centre companies now need to “build together,” “invest together” and “innovate together” as AI campuses seek power at a scale beyond earlier cloud builds. He pointed to utility partnerships, grid upgrades and on-site solutions as the operating model taking shape. (powermag.com) A data centre is a warehouse full of computers, and AI packs more chips into each room, which pushes electricity demand up faster than in older cloud facilities. JLL said on January 5 that power, not location or land cost, is now the main site-selection filter because grid connections in some markets take multiple years. (jll.com) That scramble follows a broader jump in electricity use. The International Energy Agency said global power consumption rose 3% in 2025, with electric vehicles and data centres among the drivers, while the U.S. Department of Energy said data centres used about 4.4% of U.S. electricity in 2023 and could reach roughly 6.7% to 12% by 2028. (msn.com) (energy.gov) The bottleneck is not only generation; it is also the queue to connect new power plants and grid equipment. Lawrence Berkeley National Laboratory said in its 2024 “Queued Up” report that the typical project built in 2023 took nearly five years from interconnection request to commercial operation, up from three years in 2015 and less than two years in 2008. (emp.lbl.gov) Utilities are seeing the same pressure from the customer side. An Electric Power Research Institute survey published in September 2024 said data-centre growth is straining utility planning because both total load and the size of individual AI-related requests are rising, while utilities are still refining how they study and process those interconnection requests. (epri.com) That is why the shopping list has shifted from just land and fiber to substations, voltage controls, backup systems and batteries. JLL said data-centre operators are engaging utilities and behind-the-meter generation partners in parallel with real-estate decisions, and some markets are pushing “bring your own power” strategies. (jll.com) The industry argues this is not the first time efficiency has softened a power crunch. Crosby cited the cloud era, when power-usage effectiveness improved and Google’s DeepMind cut cooling energy in one data-centre application by about 40%, but he said AI’s current buildout still requires earlier utility coordination and more physical infrastructure. (powermag.com) (deepmind.google) For data-centre developers in 2026, the fastest route to new AI capacity is often not finding more servers. It is finding a site where electricity can arrive on time. (datacenterknowledge.com)