Building the Shed

Everything else in this index has to be installed somewhere. The somewhere has a vacancy rate, and in the first half of 2026 that rate was 1.4% across North American primary markets — down from 1.6% a year earlier, despite total supply growing 33.7% to 10,903 MW, because absorption kept pace with everything built.

Under-construction capacity reached 7,481.1 MW, up a quarter in six months and past the previous peak of 6,350.1 MW set in late 2024. Of that pipeline, 80.4% was already committed before completion, against 74.3% a year earlier. Less than 1,500 MW remained available across all primary markets — roughly six months of demand at the prevailing absorption rate. Building at a record rate, vacancy falling anyway, four-fifths of the pipeline pre-sold. That is not a market clearing; it is a queue.

The market-level figure understates the scarcity where demand actually is. Northern Virginia held 4,496.5 MW of inventory — the largest concentration in the world — against 10.8 MW available, while absorbing 467.7 MW in the half, about a third of all net absorption across primary markets. Hillsboro, Oregon: 491.4 MW of inventory, 1.1 MW available. Phoenix: 1,069.2 against 12.1. Competition has moved upstream, to securing power and land rather than leasing buildings.

Buildings here are priced in dollars per watt rather than per square foot, and the unit is the point: the constraint is electrical, not spatial. A 2025 benchmark index spans roughly 60%, from $9.5 per watt in Charlotte, North Carolina through $9.8 in Phoenix and Columbus to $12.9 in New Jersey, $13.3 in Silicon Valley and $15.2 in Tokyo. The year's increase for a traditional air-cooled facility was put at 5.5%, itself down from 9.0% the year before.

Turner & Townsend's 2025 analysis reports average construction costs 7 to 10 per cent higher for liquid-cooled US facilities than equivalent air-cooled ones of similar IT capacity. It is a construction-cost comparison, not a breakdown of total AI infrastructure investment or operating costs; the quoted premium alone does not establish where those wider costs sit.

The constraint capital cannot relieve is people. A global construction study reported in July 2026 found data centres the most constrained sector globally for contractor capacity, with over 70% of 112 markets analysed tightening or overstretched — against 79% of traditional sectors balanced or with spare capacity. Some 71% of markets reported labour shortages outright. The same study found input costs had stabilised, which locates the pressure precisely: on people, not materials. And the work is being pulled out of housing and commercial building, because technology work pays better.

The last constraint is consent. The market research states that local opposition has become a serious obstacle, with community resistance and zoning delays stalling projects, and that securing local approval is now as critical to siting as power and fibre. That would be a real change in an industry that spent two decades choosing sites on latency, power price and tax treatment — but it is a brokerage's characterisation of a planning environment, not a count of refusals, and the county records that would settle it were not retrieved.

A word on whose numbers these are, because it matters more here than anywhere else in this index. There is no government statistical series for data-centre supply the way there is for minerals. The best available figures come from a commercial brokerage that earns fees on the transactions it measures, and the best cost benchmarks from a consultancy that sells cost management to the people building. Both are used because the alternative is asserting nothing. Neither is disinterested, and no competing brokerage's series was retrieved to check either.