Roughly 65 gigawatts of IT capacity are under construction across North American datacenters — about 140 percent of installed capacity — with another 361 GW of announced future projects, or roughly 760 percent of what exists today. Yet Jefferies expects only 12.9 GW of new capacity to come online in 2026 and 18 GW in 2027.

That is the sobering core of an October 9 report by Jonathan Petersen, managing director at Jefferies, built on historical data from Aterio, a market researcher specializing in datacenter forecasts. The base forecast comes with bull and bear variants covering 2026 through 2030.

"The AI industry has a datacenter problem," Petersen writes. "Demand for inference and frontier model training is growing faster than developers can bring capacity online." Many announced projects, he notes, will be delayed or canceled.

Why does more capex not mean more datacenters?

The model does not take operator forecasts at face value. Instead, it weighs the probability that each project will actually be completed. The result is a stubborn gap between intention and concrete.

Petersen is blunt about the usual shortcuts in this debate. Grid interconnects alone, or hyperscaler capex, or chip sales projections — each in isolation — miss the point. "It takes all of those inputs, plus available labor, manufacturing of electrical and cooling equipment, and a cooperative political and regulatory environment to move a data center project from the pipeline to online capacity," he says.

The constraints shift over time. In 2026, labor is the biggest damper on datacenter spending. Cooling equipment takes over as the main bottleneck from 2027 through 2029. By 2030, available power becomes the binding constraint. All of these limits, naturally, are present throughout.

How large is the gap between plans and reality?

The hyperscaler numbers tell the story. For 2026, planned and estimated North American hyperscaler capacity converge at 6.21 GW, which is unsurprising this late in the year. After that, the divergence widens sharply:

• 2027: 8.37 GW planned; 7.53 GW expected to be built
• 2028: 10.13 GW planned; 6.54 GW expected
• 2029: just under 18 GW planned; 5.86 GW expected
• 2030: 13.82 GW planned; 4.46 GW expected

These are huge gaps. Even the bull case assumes delays and cancellations, because supply chains ease only somewhat. The bear case adds labor shortages, political pressure for datacenters to pay their share of electricity generation costs, and rising skepticism in host communities about the financial benefits.

Even hyperscalers, with strong balance sheets and ample cash, cannot build at the rate they want. Money, it turns out, does not pour concrete.

What does the construction data show?

US Census Bureau data, compiled by Oxford University's Our World In Data project, shows inflation-adjusted datacenter construction spending taking off from November 2022 — the start of the GenAI boom — and accelerating since. A chart posted on X by Jason Furman, economics professor at Harvard, tracks construction growth through August 2026.

The Census figures cover construction materials and labor, including power distribution and energy backup equipment inside facilities. They exclude compute, storage and networking gear. They also exclude government datacenter spending, though that would be a tiny part of the signal.

Datacenter and utility construction is now counterbalancing declines in US commercial and residential construction. Conveniently for builders, high mortgage rates are steering construction workers away from housing and toward datacenter projects.

What is the wider bottleneck?

No political, social or economic pressure can accelerate the installation of power generation equipment at datacenters or utilities. Nobody expected the AI boom to move this fast, and supply chains — from chipmaking equipment to chipmakers to the makers of grid and in-facility power distribution gear — were not ready to ramp at the pace the industry now demands.

The verdict here is measured but clear. The buildout continues, up and to the right, but at a pace set by electricians, cooling equipment factories and grid interconnects rather than by investor enthusiasm. Expect the gap between announced gigawatts and delivered ones to stay wide through 2030.