AI data‑centre power crunch
Deploying and operating advanced models is being gated by power and physical capacity as much as by model design, with less than 10% of U.S. data‑centre capacity deemed AI-ready. (datacenterknowledge.com) Utilities and regulators are already feeling the strain—some states are pausing transmission bids and debating policy changes to manage demand. (wisconsinwatch.org) When compute and power are scarce, teams are more likely to spend on data and evaluation that improve model reliability without more retraining runs. (carboncredits.com)
A data center used to be a warehouse full of servers. An artificial intelligence data center is closer to an aluminum smelter with computers inside, because the limiting factor is often megawatts of electricity, not floor space. (datacenterknowledge.com) That is why a striking number landed this week: Sean Farney of JLL said less than 10% of United States data-center inventory can handle “true AI-dense” loads today. Most older sites were built for air cooling and lower power draw, not racks packed with graphics processors running around the clock. (datacenterknowledge.com) A rack is the metal cabinet that holds the computers. JLL says new artificial intelligence racks are approaching 100 kilowatts each, which is enough power for dozens of homes concentrated into one cabinet, and that pushes builders toward liquid cooling and heavier electrical gear. (jll.com) This is not a niche side problem. JLL says nearly 100 gigawatts of new data-center capacity could be added worldwide between 2026 and 2030, doubling global capacity, and it expects up to $3 trillion of investment to chase that buildout. (jll.com) The power appetite is already visible at the grid level. The International Energy Agency estimates data centers used about 415 terawatt-hours of electricity in 2024, and its new outlook says demand could reach about 945 terawatt-hours by 2030, roughly around Japan’s current electricity use. (iea.org) In the United States, the Department of Energy says the Electric Power Research Institute projects data centers could rise to as much as 9% of annual electricity generation by 2030, up from about 4% of total load in 2023. That kind of jump turns server demand into a utility-planning problem. (energy.gov) Wisconsin shows what that looks like on the ground. Wisconsin utilities asked federal regulators to pause parts of a transmission bidding process, arguing that delays could slow the power infrastructure needed for large new data centers, while consumer advocates warned the changes could leave households paying more. (wisconsinwatch.org) Another Wisconsin fight is about who pays for the wires. Wisconsin Watch reported in March that American Transmission Company was folding substations into a broader $1.3 billion buildout tied to the Port Washington campus, turning one data-center project into a regional cost debate. (wisconsinwatch.org) The bottleneck is not only the grid connection. JLL says power has become the primary site-selection criterion ahead of location or rent, because companies are facing multiyear waits for enough electricity to come online. (jll.com) That changes how artificial intelligence teams spend money. If every extra training run competes for scarce graphics processors, scarce cooling, and scarce power, then buying cleaner data, stronger testing, and better evaluation becomes a cheaper way to improve reliability than brute-forcing another giant retraining cycle. (carboncredits.com) The next phase of the artificial intelligence race is starting to look less like a software contest and more like a race to secure transformers, turbines, transmission lines, and water permits. The companies that get power first will be able to train and serve models first. (jll.com)
