AI data centers vs. the grid

Texas is discovering that the hardest part of an artificial intelligence boom is not chips or code. It is finding hundreds of megawatts of electricity, plus the land, wires, permits, and local consent to deliver it. (news.bgov.com) In Austin this week, Texas lawmakers and industry officials said artificial intelligence growth could strain infrastructure, raise costs, and force new tradeoffs only five years after the state’s 2021 grid failure. Texas Senator Tan Parker said the state is now prioritizing “dispatchable” power, especially natural gas, for round-the-clock data center demand. (news.bgov.com) The scale is the shock. University of Southern California professor Shon Hiatt said data centers already under construction across the United States could require more than 75 gigawatts of power within about two and a half years, which he compared to adding another Texas-sized electricity load. (news.bgov.com) A data center is basically a warehouse full of computers, and the newest artificial intelligence sites are much larger than the old internet-server kind. The Electric Power Research Institute says new centers now commonly range from 100 megawatts to 1,000 megawatts, or about 80,000 to 800,000 homes’ worth of demand. (restservice.epri.com) Artificial intelligence changes the math inside those buildings. The Electric Power Research Institute estimates a ChatGPT request uses about 2.9 watt-hours of electricity, versus about 0.3 watt-hours for a traditional Google search, which is roughly a tenfold jump. (restservice.epri.com) That does not mean artificial intelligence alone “broke” the grid. The United States Department of Energy said in its 2023 transmission study that the grid was already dealing with aging infrastructure and insufficient transmission capacity before the latest data center rush. (energy.gov) Texas is where those two stories now collide. The Electric Reliability Council of Texas, the operator for most of the Texas grid, told state senators on April 1 that it is tracking about 410 gigawatts of large-load interconnection requests, and about 87% of that queue is data centers. (ercot.com) Even that queue does not mean 410 gigawatts will all get built. It means developers are racing to reserve a place on the grid, and Texas is moving toward a batch-study process because the old one-by-one review kept getting blown up by repeated restudies when new projects landed in the same transmission zone. (ercot.com) The bottleneck is not only generation. Bloomberg Government reported that Parker warned the United States still depends heavily on China for high-voltage transformers, while the Department of Energy has separately said the country faces an ongoing shortage of transformers and other grid components. (news.bgov.com, energy.gov) The price effects are starting to show up in research. A March 2026 Federal Reserve Bank of Dallas paper found existing data centers have already raised wholesale electricity prices by about 2% to 6% on average nationwide, with bigger effects in regions that host major data center corridors. (dallasfed.org) That is why this is turning into a land-use fight as much as an energy one. From Wisconsin to Massachusetts to Illinois, local residents are objecting to projects over power bills, water use, diesel backup generators, and the fact that a huge facility can arrive with fewer permanent jobs than a factory of similar size. (politico.com, wbur.org, nprillinois.org) So the new strategic question for artificial intelligence companies is no longer just who has the best model. It is who can secure power, survive interconnection delays, line up transformers, and get through county hearings before the neighbors decide the project is too big for the grid next door. (news.bgov.com, ercot.com)