The four hyperscalers run a racket on your power bill to fund their AI buildout.
To be fair — and the concession is worth making on the record, because the temptation to deny the technology and skip the political economy is real and unhelpful — the load these companies are pulling is real. The training runs that produced the current generation of large language models consumed electricity on a scale no previous commercial computing buildout had approached. The twenty-year power purchase agreement Microsoft signed for the restart of Three Mile Island Unit 1 (since renamed the Crane Clean Energy Center) for its roughly 835-megawatt output; the deal Meta wrote for Constellation’s entire 1.1-gigawatt Clinton Clean Energy Center in Illinois, announced in June 2025; the AWS co-location arrangement at Talen’s Susquehanna nuclear plant in Pennsylvania — these are not paper instruments. The reactors are restarting. The transmission lines are being built. The compute is being shipped. To refuse to acknowledge that the underlying technology works, in defined domains at substantial cost, is to forfeit the standing from which to critique the political economy around it. The technology works. The political economy is still the racket.
The trouble is that the people paying for it, in the granular and unbudgeted sense that matters at the kitchen table, are not the hyperscalers. They are ratepayers in Georgia and Ohio and Virginia and Texas and every other state where a data center has been sited within range of a substation that the local utility could be persuaded to upgrade. The $45 million of campaign advertising NPR documented this week — running since January across gubernatorial, House, and Senate races, with the two parties’ totals within a hair of each other (over $22 million Republican, more than $21 million Democratic, independents picking up the rest) — is not, despite the breathless framing in some of the coverage, an argument about technology. It is an argument about who pays. AdImpact’s count of mentions — a handful in January, 62 in June, more than 110 in July, about 164 in August — tracks the season. The grievance is older than the season.
Here it is worth being precise about what “the grid” actually is, because the public discourse has the unhelpful habit of treating “the data center pays for its own power” as if it settled the matter. A hyperscale data center drawing two hundred megawatts from a regional grid does not, in any meaningful sense, pay for the new transmission line that the regional grid operator had to build to bring those two hundred megawatts to the substation. It pays for the electrons — under a power purchase agreement, a long-term contract with a generator. It does not generally pay for the high-voltage wires, the substations, the right-of-way, the gas-turbine readiness charges the utility incurred to be prepared for hyperscale load that might or might not arrive, or the multi-year rate-case proceeding in which the utility asked the public utility commission for permission to recover those costs. Those costs go into the rate base. The rate base is approved, in most U.S. states, by a public utility commission whose members the sitting governor appointed. The sitting governor is now running — or being attacked in — a data-center ad. Same governor. Same rate base. Different quarter.
A twenty-year power purchase agreement at a fixed price is, in energy-economics terms, a hedge — and a hedge that socializes the construction risk onto the rate base while privatizing the upside is the kind of hedge the financial-innovation literature has a name for. It is called a subsidy, and it is the kind of subsidy that does not show up in the appropriations bill, which is exactly why it survives.
Cui bono? The four hyperscalers — Google, Microsoft, Meta, Amazon — have locked in twenty-year power purchase agreements on restarted nuclear units. Microsoft signed its PPA in 2024 for Three Mile Island Unit 1; the deal drew federal financing to accelerate the 2027 restart. Meta followed with the Clinton Clean Energy Center deal. Amazon’s AWS subsidiary arranged co-located compute capacity at Susquehanna, an arrangement first reported in 2024 and subsequently contested before FERC. The generators carry guaranteed offtake. The utility shareholders carry a rate base that grew. The transmission-line construction contractors carry the build. The household paying the meter at the end of the wire pays for everything that did not get pre-paid under the contract.
The ratepayer pays. The hyperscaler captures the upside. The utility gets a guaranteed off-take that justifies the capital expenditure it would otherwise have to justify on residential-demand growth rates alone. The local government gets a ribbon-cutting and a jobs claim that, if the historical pattern holds, will be quietly revised downward about a decade later when the actual employment numbers are audited. The data-center operator, which is to say the same four firms whose downstream products are the subject of every other AI-policy debate in Washington, captures the market capitalization. Kate Crawford’s Atlas of AI made the structural point years before the polling caught up to it: AI is “made from vast amounts of natural resources, fuel, and human labor.” The “natural resources” item in that sentence is now showing up in your electricity bill.
The lobbying record helps explain why the bipartisan panic is so structurally telling. The same four companies that locked in the PPA terms spent the last decade ensuring that state legislatures passed sales-tax exemptions on data-center equipment, that county commissions agreed to property-tax abatements worth the substantial majority of assessed value for a decade or more, that federal tax provisions and Department of Energy programs subsidizing data-center buildout remained on the books, and that the trade-association roundtables in D.C. and the state capitals kept their own talking points coordinated. They did so through the Data Center Coalition and through direct filings under the Lobbying Disclosure Act. The legal record is the lobbying record. The lobbying record was largely bipartisan. Which is why the present $45-million panic is so structurally telling: the political class that built the architecture, in roughly equal shares between red states and blue states, is now confronting seventy-seven percent of Americans telling it so.
The political convergence is the tell, and it is worth pausing on. Michael Franz of the Wesleyan Media Project called data centers “a unique issue where MAGA Republicans and liberal Democrats land in the same place.” He is correct, and the reason is not complicated: both parties are responding to the same polling. A June Reuters/Ipsos survey found 77 percent of Americans concerned about data centers raising electricity costs, with two-thirds of Democrats and half of Republicans saying they would oppose a facility in their own community. This is not a coalition assembled by messaging professionals — though the messaging professionals, on both sides, are now very busy indeed. It is a coalition assembled by the light bill. The visible convergence of Bannon, Sanders, and Ocasio-Cortez on the same side of the question, and the National Republican Senatorial Committee’s warning to the tech industry that a Husted loss would make politicians nationwide wary of them, are downstream effects of that underlying arithmetic. So is Byron Donalds in Florida pledging to “fight for a ratepayer protection plan” against data centers that “jack up utility rates,” which is roughly the same sentence a Democrat in Wisconsin or Iowa is running in the same week, with the words rearranged. Wisconsin Democrats put $634,795 behind August advertising promising that David Crowley would “stop AI data centers from making your energy bills skyrocket.” Florida Republicans, through the Friends of Byron Donalds PAC, spent more than $2.5 million on Donalds pledging a “ratepayer protection plan.” The script reads like one document. The bipartisan character of the architecture’s opponents is no longer deniable.
Inside the Republican Party, the response has been to read the polls and adjust. The NRSC’s internal memorandum — obtained by NPR, first reported by Axios — called data centers “a sleeper issue for the entire election cycle.” That is unusually direct language from a party committee to the tech industry: we will not be able to defend you in November if the ratepayer revolt continues. A Republican Party operative who spoke to NPR on condition of anonymity — the party is not commenting publicly because views within the party vary — was blunter still: “We could potentially have data centers as long as they pay for their own energy.” That conditional acceptance is the most honest sentence in the entire ad campaign on either side, and the only one that names the actual point of leverage. As Darrell M. West of the Brookings Institution told NPR, Republican incumbents and nominees “want to slow down the movement towards data centers because they can see so many of their own voters are opposed.” President Trump has been the conspicuous outlier, urging states to keep building while admitting the buildout could “use a little public relations help.” West put the analytical point cleanly: “Typically, when the public shifts, politicians shift with them. Trump is not doing that on data centers.” That sentence is doing more analytical work than its length suggests, and the work it is doing is naming the structural fact that the AI buildout is now the most legible single indicator that the infrastructure of the next decade is being assembled on terms negotiated by a very small number of firms, with the political system being asked, at town-hall volume, whether the terms were fair.
The Ohio Senate race shows the ratchet turning fastest. Sherrod Brown’s supporters have spent millions describing his opponent, Republican Senator Jon Husted, as “the face of data centers,” accusing him of giving corporations more than $2.5 billion in tax breaks and pushing electric bills up by thousands of dollars. Husted, who championed data centers as Ohio’s lieutenant governor, has been introducing congressional legislation to protect ratepayers from the higher electric rates data centers can cause. As Husted told Spectrum News: “It’s not really a shift. I was always trying to hold these companies accountable by making them pay local property taxes.” The Republicans are aware of the exposure: the NRSC memo singled out the Ohio race and warned the tech industry that a Husted loss would make politicians nationwide wary of them. As West told NPR, the federal government does not really play a role in permitting new data centers; the decisions sit with governors and the public utility commissions the governors appoint. Which is to say: the same statehouses that wrote the property-tax abatements and approved the rate-base expansions are the only statehouses that can undo them. The presidential preference, in this picture, matters at the margin. It does not change the underlying fact that the statehouses are where the votes are.
The political-economy shape here is older than the technology. Harold Innis, writing about the staples economies of the Canadian and American frontiers in the middle of the last century, identified the recurring pattern in which a single extractive commodity — cod, fur, wheat, timber, oil — pulls the rest of the economy into its service, the rents captured at the extraction point flowing outward to the metropole while the costs — environmental, social, infrastructural — accumulate at the site. The hyperscaler AI buildout is not, in this sense, a novel political economy. It is the staples thesis updated for compute. The grid is the new railway. The ratepayer is the new homesteader being charged for the right-of-way.
Cory Doctorow has called the pattern “chokepoint capitalism” when it shows up between creative workers and the audiences they want to reach: a small number of large companies positioned between two groups that must transact through the chokepoint, with the chokepoint holder capturing the surplus at every transit. The grid is now a chokepoint. The hyperscaler sits on one side. The residential ratepayer sits on the other. The transmission wires the ratepayer paid for move electrons the ratepayer did not order to a building the ratepayer did not invite. Tim Wu, writing about a different kind of monopoly infrastructure in The Master Switch, called this the curse of bigness; the pattern rhymes. Both parties, in roughly equal shares, helped build the chokepoint. Both parties are now spending record sums to disclaim it.
The “China” frame the hyperscalers prefer — recasting an ordinary cost-allocation dispute as a civilizational contest to foreclose the boring legislative debate that would otherwise settle it — does not survive contact with a ratepayer who has opened the bill.
The remedy frame, when it lands, will be a stack. State-level cost-causer-pays rules for new transmission tied to hyperscale load, written into the rate-design tariff by the same public utility commissions that wrote the present tariff. A federal override available where the state commissions decline to act, with Congress holding commerce-clause authority over wholesale electricity markets and able to condition any future federal support for the buildout on cost-causer-pays at the substation level. Public-utility-commission proceedings that require utilities to attribute the marginal cost of new data-center-driven generation to the data center rather than the residential customer. Federal tax-policy changes that close the depreciation and property-tax loopholes through which the bulk of the $2.5 billion in Ohio and similar sums elsewhere have flowed. And interoperability and competition requirements that prevent the same four firms from owning every layer of the AI stack from chip foundry to inference endpoint — the lever Tim Wu has been arguing for across two decades of antitrust writing, and the piece of the remedy stack most likely to be structurally decisive if it lands. The labour side of the bargain — the electricians, the line workers, the building trades whose threatened withdrawal of political support in closely contested races is one of the few credible sources of countervailing pressure in this picture — is treated for the first time as a partner in the buildout rather than a cost to be minimized. None of this is more radical than the electricity regulation the United States has deployed in every decade since 1935. None of it requires waiting on the present panic to subside.
The present panic will not subside on its own. Forty-five million dollars is the symptom of a political class newly awake to what it built in the quiet of committee rooms and rate-case dockets over the last decade, now legible to the voters whose bills arrive in their kitchens. The deadline that matters is not the one on the campaign-finance report. It is the one on the rate-case filing in your state’s public utilities commission, where the next data-center load is being added to your bill without your name on the docket. Deadlines for comment in those dockets matter because deadlines are the only part of regulatory processes that the regulated actually respect. The work is to be done.