AI’s growing physical bill
Global AI compute and power use surged in 2025, with a Stanford-cited estimate putting AI energy demand at about 29.6 gigawatts and global compute capacity near 17.1 million GPUs by early 2026. The same reporting links that growth to rising needs for generation, cooling and grid upgrades—basically the electricity and infrastructure that run large-scale models. (businesstoday.in)
Artificial intelligence’s boom now comes with a utility-scale power bill: Stanford’s 2026 AI Index says AI data center power capacity reached 29.6 gigawatts in 2025. (hai.stanford.edu) That same report says global AI compute capacity has grown about 3.3 times a year since 2022, reaching 17.1 million Nvidia H100-equivalent chips by early 2026. Nvidia supplied more than 60% of that compute, with Google and Amazon providing much of the rest. (hai.stanford.edu) A data center is a warehouse of servers, storage gear and networking equipment. The International Energy Agency says servers now account for about 60% of electricity use in modern data centers, while cooling ranges from about 7% in efficient hyperscale sites to more than 30% in less-efficient enterprise facilities. (iea.org) The International Energy Agency estimated data centers used about 415 terawatt-hours of electricity in 2024, or about 1.5% of global electricity consumption. Artificial intelligence model training and deployment happen mainly inside those facilities. (iea.org) The spending wave is already visible. Alphabet told investors on February 4, 2025 that it expected to spend about $75 billion in capital expenditures in 2025, with most of that tied to technical infrastructure such as servers and data centers. (abc.xyz) Power supply is moving with it. Constellation said on September 20, 2024 that a 20-year agreement with Microsoft would support restarting Three Mile Island Unit 1, renamed the Crane Clean Energy Center, adding about 835 megawatts of carbon-free power to the grid. (constellationenergy.com) OpenAI, SoftBank, Oracle and MGX announced the Stargate Project on January 21, 2025, saying it intended to invest $500 billion over four years in United States artificial intelligence infrastructure and begin deploying $100 billion immediately. (openai.com) The bottleneck is no longer just chips. Stanford’s report says the buildout now depends on electricity generation, transmission lines, cooling systems and the physical limits of the grid that connects new facilities. (hai.stanford.edu) MIT Technology Review reported in May 2025 that major tech companies were pursuing new nuclear, gas and grid deals as artificial intelligence moved from occasional chatbot use into search, software and everyday consumer apps. (technologyreview.com) The result is that each new model now needs more than code and chips. It also needs substations, transmission approvals, water for cooling and enough power to keep millions of accelerators running at once. (hai.stanford.edu)
