Trump and the hyperscalers capture the grid for AI and hand the bill to towns that didn’t ask.
It is worth being precise about what a data center is, because the public discourse has the misleading habit of treating the word as a single thing, when the engineering reality is at least three. A hyperscale training cluster — the kind built to train the next generation of large language models — draws on the order of 100 to 300 megawatts of continuous power, with cooling water measured in hundreds of millions of gallons per year, depending on climate and cooling technology. A persistent inference cluster, running a deployed model for end users, draws less per facility but runs twenty-four hours a day, which is what matters to the grid planner. A colocation facility, in which many smaller tenants rent space and power inside a larger building, has a load profile again. The press treats all three as “data centers.” The grid operator treats them very differently.
The four companies whose names come up most often in the buildout — Microsoft, Google, Meta, Amazon — have, between them, signed long-term power purchase agreements for nuclear baseload at a scale that would have seemed like science fiction five years ago. Microsoft took the whole output of Three Mile Island Unit 1 for twenty years, in a deal that included restarting a reactor that had been retired in 2019. Meta signed for the entire 1.1 gigawatts of Constellation’s Clinton Clean Energy Center. Amazon is the anchor tenant adjacent to Talen’s Susquehanna nuclear plant. These are real arrangements, and they are not the typical data center.
The typical data center is a load on a regional grid, not a generator. Of the major hyperscale facilities announced in 2024 and 2025, the overwhelming majority are net loads on their host utilities, not net contributors. And the data centers being commissioned this year will be powered by the same regional grids whose average residential electricity costs have risen more than 35% over the last five years, per the Bureau of Labor Statistics. The service territories where this matters most — Northern Virginia, the I-85 corridor in Georgia, central Ohio, the Phoenix exurbs — are the ones where the new load is large enough to be the marginal increment on the system, and the marginal increment is what sets the price. The industry’s argument that other large industrial facilities also consume resources is technically correct and substantively beside the point. A new aluminum smelter pays for its own power. A data center’s standard contract shifts the transmission and capacity upgrade costs to the rate base through utility cost-of-service ratemaking, which is the part the developers do not put in the press release. State utility commissions in some jurisdictions already have the tools to attribute a portion of system costs to a specific load class and route the remainder to that class’s customers. Done properly, that mechanism protects residential ratepayers from cross-subsidizing industrial load. Done as a voluntary pledge with no enforcement, it is something else.
The administration’s “pledge” deserves the closer read the press summaries have not given it. Voluntary. Non-binding. No private right of action for ratepayers. No clawback if a developer walks. No published methodology for what counts as an “AI-driven” increment on a residential bill. The text reportedly asks signatories to “explore” cost-allocation mechanisms — that is a different sentence from “implement” them. The structure resembles the standard posture of a regulator who has captured the regulated industry’s framing of the problem: the verb does the work that the noun is doing at the press conference.
The Vice President gave a sharper test on Monday. “If you build the data center you should be putting power back into the grid, not taking it out,” JD Vance told reporters. The statement is fine as far as it goes. The trouble is that the data centers being built today are doing the second thing, not the first. The hyperscaler-nuclear-PPA deals that Vance’s standard gestures at are real and consequential, but they are a small minority of the buildout — and most of them are running into permitting and licensing timelines measured in years, not quarters. The data center being commissioned in the next four quarters will, in the overwhelming majority of cases, be a net load on its host utility. Vance’s standard would catch this if it were the operating standard. The pledge is not the operating standard. The operating standard is the contract.
That is the engineering-substance read of the 70% figure the President dismissed as an “anti Data Center movement.” A recent Annenberg survey found roughly 61% opposed new data centers, and other recent polls put the share at more than 70%. These are not people who have been told what to think. They are people being told, by their own utility bills, what is happening. The bipartisan backlash reshaping the midterms is not, on closer inspection, a wave of local sentiment detached from the underlying engineering. It is the political system’s first attempt to make the bill visible.
Look at the three governors named in the past week as the most politically telling. Greg Abbott, the longest-serving Republican governor in the country, paused new data-center approvals in Texas. Josh Shapiro, who had been one of the most AI-friendly Democrats in the country before this month, paused approvals in Pennsylvania. Sen. Raphael Warnock of Georgia went to the site of a potential OpenAI facility, talked to locals, and joined the calls for a statewide moratorium. None of these are anti-infrastructure politicians. They are politicians who have noticed that the infrastructure being proposed does not, on the contract being offered, pay for itself. Cory Doctorow’s chokepoint-capitalism frame is useful here, though not in the way the buildout’s boosters would prefer: the buildout positions the hyperscalers as the buyer side of an effectively one-firm monopsony in advanced AI compute, with municipalities, utilities, and ratepayers as the locked-in side. A town that has already built a substation, run the transmission line, and committed the political capital to a project cannot easily walk that back. The hyperscaler knows this. The town’s lawyer knows this. The ratepayers paying for the upgraded substation over the next thirty years know this too.
Kate Crawford’s Atlas of AI names the buildout for what it is: a logistical-extractive system whose material costs are offloaded to places and people who are not in the room when the contracts are written. The towns reading their utility bills and signing moratoriums are not refusing the future. They are refusing to be the future’s silent partner. The buildout’s architects have not, on the contract being offered, given them a way to be a partner instead.
Trump is right that the towns refusing to host the buildout will, on current trajectory, be the places the AI economy bypasses. He is wrong about why. They will be bypassed because the architects have not figured out how to share the gains they intend to capture. The hyperscaler-nuclear deals that get the press will run; they are real and they are coming online. The standard deal — load on a regional grid, transmission upgrades paid by the rate base, water permits negotiated in private, host-community benefits written in language no host community’s lawyer drafted — is the one most of the buildout will actually be. Vance’s standard would catch this if it were the operating standard. It is not. The operating standard is the contract.
The AI century is going to be built. The question on the ballot in November is not whether. It is who is on the hook for the meter when the cooling towers go up.