Jeffries expects North American hyperscalers to deliver just 4.46 gigawatts of new datacenter capacity in 2030 against stated plans of 13.82 GW — roughly a 68% shortfall in the base case of a forecast issued October 9.

The analysis, built with datacenter market researcher Aterio and led by Jeffries managing director Jonathan Petersen, models installed IT gigawatts in the United States and Canada only. It does not price buildings, IT gear, or operating power.

"The AI industry has a datacenter problem," Petersen wrote. "Demand for inference and frontier model training is growing faster than developers can bring capacity online."

How much datacenter is actually in the pipeline?

Jeffries' dataset counts roughly 65 IT gigawatts of datacenters currently under construction in North America — about 140% of installed capacity — with another 361 GW of announced future projects, equal to roughly 760% of what runs today. Pipeline is fat. Delivery isn't.

Base-case buildout in the Jeffries model:

• 2026: 12.9 GW of new IT capacity
• 2027: 18 GW
• Beyond 2027: decelerating, year by year, as constraints tighten

"Any of those discussions in isolation misses the bigger picture," Petersen said. He argued it takes grid interconnects, behind-the-meter power, chip supply, labor, electrical and cooling manufacturing, and a "cooperative political and regulatory environment" to move a datacenter from pipeline to online capacity.

Where does the hyperscaler plan meet reality?

The gap between stated aspirations and Jeffries' expected build widens sharply after 2027:

• 2027: Hyperscalers plan 8.37 GW; Jeffries expects 7.53 GW
• 2028: Plans 10.13 GW; Jeffries expects 6.54 GW
• 2029: Plans ~18 GW; Jeffries expects 5.86 GW
• 2030: Plans 13.82 GW; Jeffries expects 4.46 GW

A 2026 convergence is mostly cosmetic. Hyperscaler planned capacity lands at 6.21 GW for the year, and with December approaching, that figure is effectively baked in.

What changes for operators and investors?

The bottleneck rotates by year, per Jeffries:

• 2026: Labor is the biggest damper on datacenter spending
• 2027–2029: Cooling equipment becomes the primary constraint
• 2030: Available power dominates as the gating factor

All three constraints run in parallel. The model discards operators' own announcements and applies completion probabilities based on observed supply, capital, and labor inputs. That's why even the bull case — which assumes some easing of construction supply chains — still produces heavy delays and cancellations. Neither variant treats hyperscaler capex or chip-forecast projections as a sufficient predictor on their own. Labor shortages, political pressure for datacenters to pay their share of generation costs, and rising community skepticism drive the bear case.

Why is this story surfacing now?

US Census Bureau data, surfaced by Harvard economist Jason Furman on X, shows datacenter and power-utility construction growing fast enough through August 2026 to offset declining commercial and residential construction in dollar terms. Inflation-adjusted (to 2021 dollars) datacenter construction spending began its steep climb in November 2022, coinciding with the GenAI build-out.

Census data covers labor, materials, and on-site power distribution and backup equipment — not compute, storage, or networking inside the hall. Federal, state, and local government datacenter spending is excluded, though Jeffries treats it as immaterial to the trend.

What does this change for buyers planning AI deployments? Announced capacity is running at multiples of installed base, but installed growth is bound by the slowest rung on the ladder — and that rung shifts year to year.