AI data‑centre bottleneck
Nearly half of planned U.S. data‑centre projects for 2026 face delays or cancellations, creating a potential supply bottleneck for AI capacity and the risk that commercial AI subscriptions could become much more expensive. That infrastructure squeeze flips the usual debate: demand for AI is real, but physical capacity — power, real estate, and build schedules — may constrain pricing and delivery. For businesses budgeting AI spend, the result could be higher unit costs or longer lead times independent of software performance. (x.com) (x.com)
The surprise in artificial intelligence right now is not a lack of customers. It is a lack of buildings with enough electricity, permits, and electrical gear to run the chips those customers want to use. (cbre.com) A data center is a warehouse full of computers, and modern artificial intelligence needs far more of them than a normal website or payroll system. In the biggest United States markets, vacancy fell to 1.4% at the end of 2025, which means almost every suitable room with power is already spoken for. (cbre.com) That squeeze showed up in construction numbers. Capacity under construction in primary North American markets fell to 5,994.4 megawatts at the end of 2025 from 6,350.1 megawatts a year earlier, the first decline since 2020. (cbre.com) The reason was not weaker demand. CBRE said planned projects kept running into permitting, zoning, and power-procurement hurdles, which is like having buyers lined up outside a store while the shelves are still stuck at the factory. (cbre.com) Power is the hardest part because a large data center can draw electricity on the scale of a small city. The United States Energy Information Administration said national electricity demand grew about 1.7% a year from 2020 through 2025 after growing just 0.1% a year from 2005 through 2019, with data centers driving that acceleration. (eia.gov) Even when a developer has land and money, the grid connection can take years. JLL said in its 2026 outlook that power, not location or cost, is now the main site-selection criterion because grid-connection wait times are measured in multiple years. (jll.com) Prices are already moving with that scarcity. CBRE said the average asking rate for a 250-to-500-kilowatt requirement in primary markets rose 6.5% year over year to $195.94 per kilowatt per month, marking a fourth straight annual increase. (cbre.com) That feeds straight into artificial intelligence bills because the software sits on top of the physical machines. If landlords charge more for powered space and developers wait longer for new capacity, model providers either absorb the hit or pass it on through higher subscription prices, usage caps, or slower rollouts. (cbre.com) (jll.com) The bottleneck is not just wires in the ground. Bloomberg reported on April 1 that transformers, switchgear, and batteries are also in short supply, which delays projects even after financing is secured. (bloomberg.com) Developers are still trying to build at enormous scale. Wood Mackenzie said the United States had data centers in development requiring 241 gigawatts of electricity at the end of 2025, but it also said utilities and grid operators are making it harder to connect new sites because there is only so much power to go around. (bloomberg.com) So the next fight in artificial intelligence is less about whether people want the tools and more about who gets access to the limited physical capacity behind them. In 2026, the scarce resource is not code; it is megawatts. (ft.com)
