AI growth runs into power limits
Firms building industrial‑scale AI are bumping into a simple problem: most data‑centre space isn’t designed for power‑hungry AI workloads — JLL estimates under 10% of U.S. capacity is AI‑ready. Big tech is already directing capital toward reliable, high‑output generation — including next‑generation nuclear projects — and policy analysts warn electricity demand is now a core planning issue for AI rollout. (datacenterknowledge.com) (reuters.com) (brookings.edu)
The bottleneck for artificial intelligence is starting to look less like chips and more like electricity. On April 10, Jones Lang LaSalle said less than 10% of United States data center capacity can handle “true artificial intelligence-dense” loads today. (datacenterknowledge.com) A data center is just a warehouse full of computers, and older warehouses were built for lighter jobs. Jones Lang LaSalle says many enterprise sites still use traditional air cooling, while artificial intelligence systems are being packed into far denser rooms built around liquid cooling. (datacenterknowledge.com) (jll.com) That density jump is huge. Deloitte says a five-acre site that adds graphics processing units to central processing units can see power use rise from 5 megawatts to 50 megawatts, which is the difference between a small industrial load and a miniature power plant. (deloitte.com) The biggest projects are already moving past that scale. Deloitte says the largest United States data centers now being built or planned are expected to need up to 2 gigawatts of power, and some early-stage campuses could reach 5 gigawatts, which it compares to the electricity used by 5 million homes. (deloitte.com) That is why site selection has changed. Jones Lang LaSalle says power, not land price or location, is becoming the main filter for new projects because grid connections can take years and global data center occupancy is already around 97%. (jll.com) The industry’s own growth forecasts show how fast this problem is arriving. Jones Lang LaSalle projects nearly 100 gigawatts of new data center capacity between 2026 and 2030, says the sector could grow at a 14% compound annual rate through 2030, and estimates up to $3 trillion of investment will be needed. (jll.com) The power squeeze is now pushing technology companies upstream into generation. Reuters reported on April 10 that Meta agreed in January to help fund two TerraPower units with up to 690 megawatts of capacity and also signed a deal with Oklo for a 1.2 gigawatt nuclear technology campus in Ohio. (usnews.com) Amazon is making the same move at larger scale. Reuters reported that Amazon is working with X-energy to bring more than 5 gigawatts of small modular reactor capacity online in the United States by 2039, while Google signed an agreement with Kairos Power targeting its first small modular reactor by 2030. (usnews.com) Small modular reactors are nuclear plants built in smaller pieces, more like factory-made modules than one giant custom build. Reuters says none are producing commercial electricity yet, but tech companies are giving developers something banks want: long-term customers with strong balance sheets. (usnews.com) Policy researchers are now treating electricity as part of artificial intelligence governance, not a side issue. Brookings wrote on April 10 that advanced artificial intelligence models have become a fast-growing source of global energy use and that future power availability and grid capacity could limit how much computing can actually be deployed. (brookings.edu) So the new race in artificial intelligence has two finish lines. One is getting more chips into data centers, and the other is getting enough electrons to the building to turn them on. (datacenterknowledge.com) (brookings.edu)
