AINEWS
Half of US datacenters slated for 2026 are not under construction yet
A Jefferies research note finds 80 percent of 2027 28 planned capacity has not broken ground, and the gap is structural, not cyclical.
A Jefferies research note finds 80 percent of 2027 28 planned capacity has not broken ground, and the gap is structural, not cyclical.
Only 12 of the 24 gigawatts of US datacenter capacity officially slated to come online in 2026 is currently under construction. The other half exists on press releases, utility filings, and investor decks, not on job sites, according to a Jefferies research note shared with The Register (Only half of US datacenter capacity planned for 2026 is actually under construction).
That gap matters because the 24-gigawatt figure is not just an analyst projection. It is the number utilities, regional grid operators, and the AI labs themselves have used to plan power purchases, land acquisitions, and product roadmaps. When half the planned capacity slips, the consequences flow in two directions: electricity demand forecasts need to be redrawn, and the AI services pitched to investors as "coming next quarter" quietly stretch into the following year.
The picture gets worse further out. Jefferies estimates that as much as 80 percent of capacity planned for 2027 and 2028 has not started substantial construction, and aggregate announced US datacenter capacity is now approaching 160 gigawatts by 2032. The analyst's realistic build rate is 15 to 20 gigawatts a year, well short of the 40-plus gigawatts some forecasts assume for the back half of the decade.
The reason is not a single bottleneck. It is a stack of them, and most are old-economy infrastructure problems that predate the AI buildout.
The largest is interconnection queues. The process of connecting a new large load to the transmission grid has stretched to seven years in some US regions, with hyperscalers and colocation operators waiting behind other generators and industrial customers. The Department of Energy has directed FERC to speed up interconnection rules for datacenter-class customers (The Register — DOE powers up plan to get AI datacenters on the grid quicker), but rule changes take time and many of the 2026 projects were queued years ago under the old process.
On top of that sit transformers and skilled trades. Large power transformers have lead times measured in years, and the welders, electricians, and high-voltage technicians needed to build a hyperscale campus are drawn from a finite labor pool. Order a transformer today and you may not see it installed until the project you needed it for has already missed its window.
Then there is the offtake problem. A datacenter only pencils out if a buyer has signed for the compute or colocation space, and many announced projects are sitting on land and power reservations without a committed anchor tenant. Until an offtake is signed, the project is a real-estate play, not a datacenter.
The most damning issue is duplicative counting. Hyperscalers routinely submit multiple interconnection requests to different utilities for the same logical project, a hedging behavior that lets them keep options open but inflates the announced-capacity number when the requests are summed across the country. Jefferies argues that the headline 160-gigawatt 2032 figure overstates real build intent for exactly this reason.
Operators are not standing still. Behind-the-meter generation, where a datacenter builds its own gas turbines or fuel cells on site, and hybrid configurations that pair grid power with on-site batteries are increasingly the path forward for projects that cannot wait for the queue. The Register's reporting on the AI buildout's collision with the US grid describes some of these workarounds (The Register — AI datacenter boom collides with US grid reality). Texas announced 14 gigawatts of new datacenter projects in the second quarter of 2026 alone, a one-quarter tally that reflects how aggressively the pipeline is shifting toward states and utilities with faster permitting and shorter queues.
None of this means the buildout is vaporware. The constraint stack is real, but so is the demand. The 15-to-20-gigawatt realistic build rate is still enormous, larger than the entire US datacenter fleet a decade ago. What it means is that the announced-capacity number, the figure cited in earnings calls, regional economic-development press releases, and AI roadmap slides, is not a planning signal. It is a maximum possible, with most of the risk concentrated in the projects that have not yet ordered a transformer, signed a queue position, or committed an anchor tenant.
Watch those three things on the next datacenter announcement you read: the project's interconnection queue status, whether transformers have been ordered or delivered, and whether an offtake contract has been signed. If none of those is public, treat the gigawatt number as a marketing figure, not a forecast.