1GW data center could cost $50B

A 1-gigawatt AI data center is no longer a thought experiment. Google confirmed one near Detroit in March, and OpenAI and Oracle say Stargate is already past 5 gigawatts under development. (datacenterdynamics.com) (openai.com) A gigawatt is roughly the output of a nuclear reactor. The Financial Times reported in December that five U.S. facilities due from 2026 will each draw at least that much electricity. (ft.com) Google said its Michigan campus will sit in Van Buren Township on about 130 acres and be backed by 2.7 gigawatts of new resources with DTE Energy. Those resources include 1.6 gigawatts of solar, 400 megawatts of four-hour storage, 50 megawatts of long-duration storage, 300 megawatts of other clean resources, and 350 megawatts of demand response. (datacenterdynamics.com) (blog.google) OpenAI said in July 2025 that an expanded Oracle partnership added 4.5 gigawatts of Stargate capacity, taking the project to more than 5 gigawatts under development. In September, OpenAI, Oracle, and SoftBank said five additional U.S. sites pushed planned capacity to nearly 7 gigawatts and more than $400 billion in investment over three years. (openai.com 1) (openai.com 2) The basic problem is power density. AI clusters pack far more chips into each building than older cloud data centers, so the bottleneck shifts from land and concrete to substations, transmission lines, and long-term electricity contracts. (ft.com) (semianalysis.com) That is why the headline number on a gigawatt campus is not just construction cost. Developers now have to secure generation, storage, grid interconnection, and cooling systems at the same time they line up Nvidia chips and contractors. (blog.google) (ft.com) The grid math is tightening fast. The Financial Times, citing S&P Global Energy, reported that new U.S. data centers will need 44 gigawatts of additional capacity by 2028, while only about 25 gigawatts is likely to be available in time. (ft.com) That shortfall is already changing company behavior. SemiAnalysis reported on April 28, 2025 that Microsoft froze 1.5 gigawatts of near-term self-build projects that had been scheduled for 2025 and 2026. (semianalysis.com) Older crypto-mining sites are part of the search for faster capacity, but they are not plug-and-play AI campuses. Ramboll said mining facilities typically lack the reliability, backup power, and cooling systems required for AI-focused data centers, even when they already have access to cheap electricity. (ramboll.com) So the real story behind the “$50 billion gigawatt campus” debate is simpler than the meme. The AI race is now being priced in megawatts, transmission upgrades, and years of grid work as much as in chips. (ft.com) (datacenterdynamics.com)