The new spending frontier: how US data-centre construction overtook the country's airports, seaports and transit systems combined

The arithmetic is stark, and the people presenting it are not given to grandstanding. According to a 30 June 2026 data brief circulated by Unusual Whales and drawn from federal construction-put-in-place series, United States data-centre construction spending now exceeds what the country allocates, in a single year, to airports, marine terminals and mass-transit systems combined. The figure is not a forecast. It is a tally of cranes, concrete and switchgear already on order, already pouring, already drawing power from grids that, in several states, were not designed for the load.

For a half-century the canonical American infrastructure debate has run through highways, runways and rail. That triad is no longer the centre of gravity. The capital that once paid for a new terminal apron or a metro extension is now paying for server halls, cooling plants and the high-voltage transmission lines that connect them to substations originally built for aluminium smelters and paper mills. The shift is a measure of where returns are expected, where policy has chosen to lean, and where the political economy of the next decade is being written into concrete.

What the new line item actually buys

A modern hyperscale campus is, at its core, a power-and-cooling problem dressed up as a software problem. The buildings are large but unremarkable. The interesting engineering is upstream: long-lead switchgear, on-site substations, water-cooled heat exchangers, and increasingly, behind-the-meter generation. None of that comes cheap, and almost none of it appears on a balance sheet as a single tidy asset class. The Unusual Whales tally works because the federal construction series does treat the shell, the mechanical, the electrical and the site work as one capital flow.

The result is a category that has scaled past categories the political class still talks about as if they were ascendant. Airport terminal modernisation moves in five- to ten-year cycles, funded by a mix of passenger facility charges, municipal bonds and federal grants. Marine terminal work is lumpy, project-driven, and constrained by the rhythm of port authority capital plans. Mass-transit construction has not, in real per-capita terms, returned to the levels of the post-war urban renewal era. All three still matter. None of them is where the marginal construction dollar is going in 2026.

The scale is the point. When a single private-sector capex line crowds out, in the national accounts, the combined civil-works budgets of three legacy infrastructure modes, something has changed about how the country prices the future. The line item is not public spending, and the comparison is not strictly apples-to-apples. But the fact that private capital is willing to commit at this scale says something that a thousand op-eds about industrial policy cannot.

The political economy behind the build

The build did not arrive by accident. The policy scaffolding behind it includes the CHIPS and Science Act of 2022, the Inflation Reduction Act's manufacturing credits, and a patchwork of state-level incentive packages that have collectively underwritten fabs in Ohio, Arizona and Texas, and the data-centre campuses that orbit them. The logic of the bet is straightforward: if the country is going to host the next generation of compute-intensive industry, the physical plant has to be on-shore, and the grid has to be able to feed it.

The harder question, which the build itself does not answer, is whether the bet is being priced correctly. The hyperscale operators that are signing ten-year power purchase agreements are not charities. They are placing a specific wager: that the demand for AI inference, for cloud-hosted enterprise workloads, and for the training runs that precede both, will continue to compound at rates that justify multi-billion-dollar campus builds. The wager may pay off. It may not. What is unusual is the size of the cheque being written before the demand curve is fully visible.

The supply side is also concentrated in ways the headline figure does not show. A small number of contractors, a smaller number of switchgear suppliers, and a still-smaller number of utility-scale generation developers are doing the bulk of the work. The bottleneck is no longer capital. It is transformers, turbines, skilled electrical labour, and the multi-year queue at the interconnection authority. Several of the largest announced campuses are now sitting in a holding pattern, fully permitted and fully financed, waiting for the grid to be reinforced to the point where they can be energised.

What the old categories still buy

It would be a mistake to read the new arithmetic as a verdict that airports, seaports and mass transit no longer matter. They matter; they are simply no longer the leading edge of capital allocation. Airports are still the binding constraint on regional air-service growth in much of the country. Marine terminals are still the chokepoint for agricultural exports and for the containerised imports that keep retail prices down. Mass-transit systems in the largest metros are still running rolling stock that was supposed to be replaced a decade ago.

What has changed is the opportunity cost. The same engineering and construction firms that once competed for terminal modernisations and light-rail extensions are now allocating their senior project managers to data-centre work, where the schedule pressure is tighter and the margins are fatter. The labour that used to flow into public-works projects is being bid away by private campuses offering premium wages for the same skills. None of this is a conspiracy. It is the operation of a tight labour market in a sector that has found, for the first time in a generation, a buyer willing to pay a premium.

The structural question the reordering poses is not whether the country is over-investing in compute, but whether it is under-investing in everything else. The construction-put-in-place series does not, on its own, settle that question. It does, however, put a number on the trade-off that, until now, has been argued in adjectives rather than in dollar signs.

The grid as the binding constraint

For all the attention lavished on chip supply, on cooling technology, and on the geopolitics of advanced packaging, the binding constraint on the data-centre build in 2026 is the grid. The campuses that have broken ground are, in aggregate, asking for more new firm capacity than the country's regional transmission organisations have queued. The queue is measured in years, not months. The most ambitious projects have begun to bring their own generation: on-site gas turbines, fuel cells, in some cases small modular reactor procurements that will not be energised until the early 2030s.

This is where the legacy of the old infrastructure categories becomes most visible. The transmission lines that a hyperscale campus needs to draw from a distant wind farm, or to back up its on-site generation, are the same transmission lines that an airport expansion or a port electrification would have needed. They take a decade to permit and another decade to build. The data-centre build is consuming that lead time at a rate the system was not designed to absorb.

The honest summary is that the United States has, in effect, run a very large unfunded mandate on its transmission planners. The mandate was issued not by Congress but by the capex committees of a handful of cloud and model-lab operators. The planners are catching up. They are not there yet.

Stakes and what remains uncertain

The stakes of the reordering are not symmetric. The regions that host the new campuses gain construction jobs, tax-base expansion, and a seat at the table in the next round of AI-related industrial policy. The regions that lose out — typically those whose grids cannot accommodate the load, or whose workforces lack the specialised trades the build requires — are left with the question of how to compete for the next cycle. There is no guarantee there is a next cycle, or that it will look like this one.

The most uncertain variable is demand. The construction-put-in-place number is a real, sunk commitment. The revenue model that justifies it is a forecast. If the rate at which AI inference substitutes for other forms of software spend is lower than the operators' internal models assume, the campuses will be built but underutilised. The capital will have been spent. The grid will have been reinforced. The competitive advantage will have been conferred on whichever jurisdictions hosted the build. The expected revenue may not arrive.

That is the bet, written into concrete and switchgear across half a dozen American states, with a cheque that now exceeds the combined annual outlay on the country's airports, seaports and mass transit. It is the largest single reordering of the country's capital priorities in a generation, and it is being carried out, in the main, without a single line of contested appropriations legislation.

This publication framed the Unusual Whales data brief as a marker of a structural reordering in US capital allocation rather than as a one-off anecdote. The wire services have not yet published a comparable synthesis; the federal construction-put-in-place series is the underlying primary source.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://en.wikipedia.org/wiki/Construction_spending_in_the_United_States
• https://en.wikipedia.org/wiki/Data_center
• https://en.wikipedia.org/wiki/Electricity_sector_in_the_United_States
• https://en.wikipedia.org/wiki/CHIPS_and_Science_Act