Artificial intelligence is a physical industry. It runs on land, water, electricity and silicon, and building the places it runs in has become one of the largest capital programmes in industrial history — the fourteen biggest operators alone are on course to spend roughly $750 billion in 2026. The buildout began in a shortage: through 2023 the accelerators to fill the buildings could not be had at any price, and the race to secure them pulled campuses out of farmland in Louisiana, Wisconsin, Texas and rural Georgia at a pace no other kind of construction matches. What is being built where, what it consumes, what it pays for and who pays for the rest are separate questions, and the discourse around them swings between alarm and dismissal — usually untethered from the permits, utility filings and market data that actually settle them.

Those documents keep adjudicating in both directions. On one side, the scale is real and it is arriving: concrete energised at Abilene and Mount Pleasant, a restarted reactor at Three Mile Island bought outright for one company's load, campuses that doubled in cost between announcement and groundbreaking, and two decades of flat American electricity demand ending. On the other, a great deal of what is announced never gets built — a gigawatt campus turns out to hold a fraction of that in signed leases, an expansion is scrapped, half of a year's planned capacity slips — and the benefits promised to host communities shrink under measurement. Causal study and company disclosure agree that the permanent workforce is a small fraction of the construction one, in ratios that keep landing near twelve or fifteen to one. The environmental record is equally two-sided: the most-quoted figures for the sector's power and water use often turn out not to be in the reports they are credited to, while the number the coverage omits is that water consumed at the power plants supplying data centres runs about twelve times their direct draw.

What changed is that the buildout stopped being something the system absorbed and became something the system regulates. Communities moved first — a statewide permit pause in New York, a snap 120-day stop in Oregon's densest data-centre city, a discharge permit withdrawn in Ohio after thousands of objections, and pre-emption laws in the other direction where states wanted the investment. Then the grid itself moved. A single line fault in Northern Virginia dropped more than three gigawatts of data-centre load in seconds, exposing that no reliability standard governed how loads that large should behave; federal regulators responded by ordering mandatory standards and a registry for them. And the largest American grid operator, after three capacity auctions that cleared at their price ceiling and still came up short, proposed to register every large site, buy the missing supply itself, and cut new data-centre load before the emergency measures that reach everyone else. Its own market monitor puts billions of the resulting charges directly on data-centre demand. And the federal government now sits on both sides of its own policy: an executive order directs agencies to open public land to data centres and accelerate their permits, while the Interior department's own appeals board has stopped one of the first for an environmental review that never examined the project it approved.

Three things remain unsettled, and they are the ones worth watching. The first is who pays: nothing in the federal architecture can assign retail costs to a particular customer, so whether the buildout's grid costs land on its operators or on ordinary bills is being decided state by state, with the default in at least one operator's own filings falling on everyone. The second is whether the measurement improves. The first requirement of its kind is now on the statute book: New Jersey obliges every data centre to file its energy and water consumption with a regulator twice a year — the kind of per-site figure no American operator has published and no regulator has held. But the reports are self-declared, no audit is written into the law, and access to them is partly carved out of the state's public-records act, so how much the new regime actually learns depends on rules still being drafted, in Trenton and in the federal registries alike. The third is politics, and it is hardening fastest. Opposition to a data centre in one’s own area is now the majority American position, having risen twelve points in four months while views of artificial intelligence itself did not move at all — the objection is to the building rather than the technology, it crosses both parties, it is strongest among the young, and it does not soften with familiarity. It is being written into local ordinance faster than any legislature is writing it out, which makes siting an electoral question before it is a technical one.