U.S. completes just 5 GW of 12 GW planned AI datacenter capacity for 2026
- U.S. developers entered 2026 with a much larger artificial-intelligence data-center pipeline than what had actually broken ground, as utility and equipment delays held projects back. - CBRE says operators now prioritize securing 300-megawatt-plus power deliveries within 36 months, while interconnection work can stretch to 24, 36 or 48 months. - Transformer shortages and grid queues are colliding with AI demand growth. (woodmac.com)
Artificial-intelligence data centers are bigger than older server farms, and the limiting factor in 2026 is often electricity, not land. (cbre.com) (woodmac.com) CBRE says U.S. operators now rank power cost and delivery speed ahead of fiber links when they choose sites, with 300-megawatt-plus service in under 36 months the new target. (cbre.com) That target is colliding with utility reality. CBRE says any need for new high-voltage transmission or extra generation can push interconnection timelines to 24, 36 or even more than 48 months. (cbre.com) Wood Mackenzie said in March that planned U.S. data-center capacity additions were cut in half from the third quarter to the fourth quarter of 2025 as load-queue challenges persisted. (woodmac.com) The firm said Texas, Virginia and Georgia still led in capacity under development, but a large share of the biggest projects remained speculative, especially where developers were counting on future behind-the-meter gas generation. (woodmac.com) (cbre.com) Equipment is another choke point. Wood Mackenzie said U.S. transformer demand has surged since 2020, imports now supply about 80% of power transformers, and domestic factories have not kept pace. (woodmac.com) Those transformers are the voltage converters that let a data center take power from the grid and use it safely inside the building. Without them, a finished shell is still not a working campus. (woodmac.com) Inside the building, AI hardware creates its own bottlenecks. Micron said high-bandwidth memory has become a critical performance component for graphics processing units, and newer Nvidia systems are being designed around larger memory footprints. (micron.com) SEMI said in March that U.S. data-center electricity use could rise from 4.4% of the grid to as much as 12% by 2028, while cooling and thermal-density problems are forcing more liquid-cooled designs. (semi.org) JLL said North America already had more than 35 gigawatts of data-center capacity under construction at year-end 2025, with 92% of the development pipeline precommitted and vacancy near zero. (jll.com) The result is a market where graphics chips still matter, but substations, transformers, switchgear, cooling systems and utility approvals increasingly decide which AI campuses actually open on time. (woodmac.com) (cbre.com)
