AI's power problem

A chatbot answer feels weightless, but the machine behind it is not. A single new data center can need as much electricity as a small city, and utilities are now planning around server farms the way they used to plan around factories. (epri.com) That shift is showing up in national forecasts. The United States Energy Information Administration said on April 8 that its 2026 Annual Energy Outlook now expects stronger long-term electricity demand partly because of data centers, while its short-term outlook projects U.S. power use rising from 4,195 billion kilowatt-hours in 2025 to 4,244 billion in 2026 and 4,381 billion in 2027. (eia.gov) (energynow.com) The old assumption was that computing would keep getting more efficient and electricity use would stay manageable. The new problem is that artificial intelligence training and artificial intelligence inference work like running thousands of race cars at once instead of a few fuel-efficient sedans. (epri.com) (goldmansachs.com) The numbers are getting big fast. The Electric Power Research Institute said on February 26 that data centers now use 4% to 5% of U.S. electricity, and could reach 9% to 17% by 2030 under its latest scenarios. (epri.com) This is not spread evenly across the map. In Virginia, where the country’s biggest cluster already sits, the Electric Power Research Institute said data centers use about 25% of the state’s electricity today and could reach 39% to 57% by 2030, with Arizona, Indiana, Iowa, Nebraska, Nevada, Oregon, and Wyoming also at risk of crossing 20%. (epri.com) Electricity is only half the fight. Cooling those buildings can require millions of gallons of water a day, and a 2024 Lawrence Berkeley National Laboratory estimate put direct U.S. data-center cooling water use at 17 billion gallons in 2023, with the possibility of doubling or quadrupling by 2028. (eesi.org) (theconversation.com) That is why local politics has turned hostile. Reuters reported on April 6 that Amazon, Microsoft, and Alphabet’s Google had each recently walked away from multibillion-dollar U.S. data-center projects after community opposition over water, land use, and power strain. (usnews.com) Some states are now trying to slow the buildout or at least force a clearer accounting. Minnesota lawmakers are pushing separate water permits for hyperscale data centers, and Virginia’s 2026 legislative session was dominated by bills tied to electricity demand, water use, emissions, and land use. (mprnews.org) (virginiamercury.com) While U.S. towns are arguing over substations and aquifers, China is building. CNBC reported on April 8 that Alibaba and China Telecom opened a data center in Shaoguan, Guangdong, built around 10,000 of Alibaba’s own Zhenwu artificial intelligence chips, with plans to expand to 100,000 chips. (cnbc.com) That detail matters because the facility is not just a warehouse full of servers. It is a stack: Alibaba designed the chips through its T-Head unit, runs a major cloud business, builds the data centers, and sells the artificial intelligence services that sit on top. (cnbc.com) The race in artificial intelligence now looks less like a software contest and more like a contest over who can secure transformers, turbines, transmission lines, water rights, and chip supply at the same time. The companies that solve that physical bottleneck get to train bigger systems and serve more users, even if their models are not the only good ones. (epri.com) (cnbc.com)