Each AI chatbot runs on servers in a warehouse on a power grid, and the industry's own 2025 tally shows roughly half of the largest planned AI datacentres will not be built, a bottleneck Google now admits by calling its cloud business
Two hundred and fifty new datacentre projects were announced worldwide between 2021 and 2024, each demanding more than 100 megawatts of electricity, the output of a small power plant and enough to serve roughly 300,000 homes. Roughly half will either miss their build window or never be built at all, according to the Uptime Institute's 2025 global survey, the body that inspects and rates the facilities.
Google has publicly described its cloud business, which rents out AI compute to other companies, as 'compute-constrained', unable to deliver all the AI capacity customers want because the rooms that house the chips cannot be built or powered fast enough.
The clearest recent example sits at the edge of a Virginia Civil War battlefield. The Prince William Digital Gateway, about 2,000 acres set aside for as many as 10 large datacentres next to the Manassas battlefield, was approved by county supervisors in late 2023 and briefly looked like a model for how fast datacenter land-use fights could move. Opponents took it to court. A March 2026 ruling halted the project; Prince William County dropped its appeal in April; and by July 2026 the project was officially declared dead.
The opposing filings, including a January 2026 amicus curiae brief from a regional conservation coalition, made the dispute concrete: "the solemn nature of this historic site would become marred by sitting in the shadow of the monstrous datacentres, along with their associated electrical infrastructure." Uptime's research director, Jay Dietrich, lists the recurring causes across the wider pipeline: developers without datacentre experience pitching projects on spec, with no committed tenants; proposals so large that the local grid, water utility, or road network cannot absorb them; concentration in a handful of "datacentre corridors" where permitting commissions have hit their processing limit; and financing that has tightened as the cost of capital has risen. None of these are technical AI problems. They are civil-engineering and politics problems arriving on the same timeline as the next model release.
The chokepoint does not have to be a cancellation. In Santa Clara, California, the city that gave Nvidia its name, the LA Times reported in November 2025 that newly built datacentres are standing empty for years because the local utility cannot deliver them enough power. In Amsterdam, an Australian developer has sued the Dutch grid operator after being refused a new connection, the datacentre equivalent of being denied a fire hydrant.
Hyperscalers have historically broken through such constraints with capital, financing the substations and power contracts their projects needed. That pressure relief still exists, but the friction piling up in 2026 is different in kind. The projects dying are overwhelmingly speculative, pitched by developers without anchor tenants, and the choke has moved from finance alone into grids, zoning, water rights, and heritage cases. Money cannot vote on a rezoning appeal, and the projects dying in 2026 are dying in places they have not died before.
The projects still moving are the ones with anchor tenants and signed power contracts. The ones dying are the speculative projects, pitched on behalf of an AI demand curve that did not arrive in time. The Uptime Institute counts roughly 125 such projects among the 250 it has been tracking since 2021.