Trump makes ratepayers finance the AI buildout.
It is true, in the narrow sense in which talking points are usually true, that a gigawatt of new load buys something. It buys hyperscaler capacity to run inference at scale — the kind of compute that, on the documented evidence of any benchmark you can read, does a few useful things and a great many things it does not. The trouble is that the bill for that gigawatt — the transformers, the transmission upgrades, the generation, the water and sewer capacity, the twenty-year power contract that locks a baseload unit into serving one buyer — does not land at the hyperscaler’s accounting department. It lands at the regulator, in the form of a rate case, and from there, where new transmission and generation costs are socialized into general rates, at the ratepayer, in the form of a monthly bill.
The Trump administration talks about the buildout as if it were free. The arithmetic says otherwise, and the arithmetic is not a matter of opinion.
A note on what the people pushing this are actually pushing, because the public discourse has the unhelpful habit of treating “an AI data centre” as a single object rather than as what it actually is: a campus drawing hundreds of megawatts, with its own substation, its own water rights, its own security perimeter, and a long-term contract with a utility that will need to build new capacity to deliver the power. The building is only the visible part. Behind it sit transformers, cooling systems, transmission lines, nuclear plants, roads, municipal services and a rate base large enough to absorb whatever the contract leaves behind.
Microsoft signed a twenty-year power-purchase agreement with Constellation in 2024 to underwrite the restart of Three Mile Island Unit 1 — now called the Crane Clean Energy Center — at roughly 835 megawatts. That is the sort of number that used to mean a city’s worth of industry. It is now committed to one corporate buyer for two decades. Meta followed with a long-term agreement covering roughly 1.1 gigawatts from Constellation’s Clinton Clean Energy Center in Illinois. AWS placed a data-centre campus beside Talen’s Susquehanna nuclear plant in Pennsylvania and contracted for power from the same site. Google has signed comparable arrangements.
These are real contracts, signed by named companies, backed by real financing. The electricity has to come from somewhere on the grid the rest of us also use.
That somewhere is the question the rate cases are now answering, and the answer depends on where you live. When a utility signs a hyperscaler PPA — a long-term contract to sell a fixed block of power at a fixed price — the utility’s costs of upgrading transmission and generation can enter the rate base, the regulated asset value against which rates are set to recover them. Where state regulators allow those costs into general rates, the hyperscaler pays its contracted price and the rest of the ratepayer base subsidizes the buildout the hyperscaler requires. Where regulators force the hyperscaler to pay for its own interconnection and grid upgrades, the arithmetic lands differently.
The mechanism is not subtle.
The hyperscaler’s PPA pays for the megawatt-hours the data centre consumes. It does not necessarily pay for the grid upgrades its existence forces, the marginal cost of reliability work, the expanded substations, the water infrastructure, the roads, or the increased exposure carried by the host municipality. Those costs do not evaporate. They move — onto households and small businesses that cannot relocate their electrical load to a state with a better tax structure.
This is where the political economy becomes more useful than the press release. Harold Innis wrote about staples economies — cod, fur, wheat, timber — in which a metropolitan buyer captures the value of an extracted resource while the producing region manages the social costs. The geography is different. The math is identical. An outside firm arrives, pays the marginal cost of extraction, books the revenue elsewhere and leaves the fixed cost to the public.
The AI version is more sophisticated than the nineteenth-century cod-fishery version because the staple is compute rather than fish. The corporate campus sells a service rather than taking a material away. But the structure survives the change in nouns. Revenue is booked to the hyperscaler’s corporate parent. The cost of the megawatt-hours, the grid, the water and the municipal services is booked where those costs fall: in the host county, across the utility’s rate base, on the monthly bills of people who did not sign the PPA.
That is the deal.
Cory Doctorow has a useful four-stage account of platform decay: good to users, then good to business customers, then value clawed back from both for shareholders, then collapse. He calls the pattern enshittification. The data-centre version is happening in real time to the ratepayer, who is not the hyperscaler’s user but is nevertheless being placed inside its cost structure.
To be fair — and I mean this in the Canadian sense, as a real concession rather than rhetorical throat-clearing — the AI buildout is real. The grid does need to expand. Some of the transmission upgrades the hyperscalers are forcing will be needed anyway. Hyperscaler power contracts have kept nuclear units open that might otherwise have closed, and the restart of Three Mile Island Unit 1 is the most visible example. New load can justify new generation that, in time, lowers wholesale prices. The PPAs are not on their face an unreasonable way to finance new nuclear.
The trouble is that the mechanism by which the buildout’s costs are distributed is also the mechanism by which its gains are captured.
The hyperscaler gets reliable baseload on terms the open market might not provide. The plant gets a creditworthy customer. The utility gets a large new load. The state gets a story about reindustrialization. The ratepayer gets the rate case.
The political moment arrived when that story reached the electric bill. The Associated Press described artificial-intelligence data centres as bipartisan shorthand for tech-sector excess. The shorthand did not form because of a particularly bad press cycle. It formed because the bills started arriving.
In August, a Senate Republican campaign arm circulated polling warning that voter anger at the buildout was no longer a coastal-elite story. Donald Trump continued promoting the data centres as the future of American industry. The contradiction is not a messaging problem. It is a math problem now visible in monthly bills, which is the chain connecting a power contract in a corporate boardroom to a ballot in a congressional district.
Republican voters in Trump country were sold the buildout as patriotic reindustrialization. They are discovering that the reindustrialization is also being built on their dime, for the explicit purpose of training and running models whose capabilities are described in the vaguest possible terms in the same press releases announcing the next eighty billion dollars of capital spending.
The data centre has stopped being an abstraction. It has become the line item.
This is also why the usual language about innovation is inadequate. An AI data centre is not a cloud in the figurative sense. It is a physical industrial installation with a negotiated claim on land, energy, water, roads and public credit. Treating it as a weightless digital service is not merely imprecise. It is how the cost disappears from the sentence.
The four forces that, as Doctorow argues, once constrained concentrated industrial power were competition, regulation, self-help through interoperability, and labour. Each has been weakened over the same decades in which platform and energy decisions have been consolidated. Competition does not constrain a firm that can secure the scarce power, land and financing before anyone else can enter. Regulation does not constrain much when the regulated party arrives with a legal department larger than the public-interest staff reviewing its application. Interoperability is not a serious remedy for a substation. Labour cannot bargain over a rate base it does not control.
The first two still matter. They are simply being applied to the wrong side of the meter.
The relevant question is not whether artificial intelligence is useful. Some of it is. The question is who pays when a private company makes a public infrastructure claim, who owns the resulting assets, who receives the subsidy, and who can leave when the promised revenue does not arrive. A claim that the buildout will eventually produce public value is not a rate-allocation formula. “Eventually” is not a transmission tariff.
Speaker Mike Johnson’s remark last week that “some of these things are out of our control” landed in the middle of a list that, according to the AP’s account, included war with Iran, the demolition of the East Wing of the White House, the threatened demolition of the Kennedy Center, the appearance of democratic socialists in mainstream debate and the question of Lake Ontario’s name. He did not name data centres.
The omission is diagnostic.
The data-centre buildout is not a force of nature. The federal government writes loan guarantees. The Department of Energy’s loan office writes conditions. The Federal Energy Regulatory Commission oversees interstate transmission. State public-utility commissions decide how costs enter the rate base. Utilities propose the allocation. Hyperscalers negotiate the contracts. Each decision has an author, a beneficiary and a docket.
The counterparty, it turns out, was the ratepayer.
That is the part of the arrangement the slogan leaves out. “American leadership” does not specify who pays for the transformer. “Reindustrialization” does not say who owns the transmission line. “Energy security” does not explain why a household with no data centre on its street should finance one.
The public lever is still in the docket. State commissions can require hyperscalers to pay the transmission upgrades their load requires. Interconnection rules can price the load honestly. Procurement standards can account for water withdrawals, land and reliability costs. Federal loan guarantees can require rate-base protections for host states rather than treating the ratepayer as an invisible source of collateral.
Most utility commissions run public comment periods on rate cases touching hyperscaler load. The costs that need challenging are precisely the costs a hyperscaler has an incentive to socialize: transmission upgrades, generation siting, water infrastructure and reliability work. The public record is where those costs can be named before they become a monthly bill.
Deadlines are the only part of regulatory processes the regulated actually respect, and submissions are the only part of the record subsequent administrations have to read. The submission portal, for those who know which docket to file in, works.
The work is to be done.