The data center consuming a nuclear reactor’s worth of electricity in eastern Ohio is not a campaign-trail political liability. It is an engineering problem with a documented cost distribution: the hyperscaler captures the compute and the revenue, the ratepayers absorb the grid upgrade, and the public service commission holds the bag when the thermal load exceeds what the local distribution transformer was built to carry. Ohio Sen. Jon Husted’s failed amendment Thursday — an attempt to shield consumers from additional energy costs tied to data center operations — was not a political miscalculation. It was the rare instance in which an engineer-turned-politician tried to address the actual extraction architecture, and his own party’s Senate leadership killed it before it reached the floor.
This is the distinction that a growing bloc of vulnerable Republicans keeps failing to make. They are breaking with President Trump on tariffs, on the Iran war, on the AI-driven data center buildout — but they are breaking on politics, not on substance. The breaks are visible and repeated: Sen. Susan Collins of Maine distancing herself from tariffs she says are crushing small businesses; Sen. Roger Marshall of Kansas separating himself from duties that are strangling aircraft production in Wichita; Sen. Pete Ricketts of Nebraska questioning whether consumers will see lower prices from allowing more foreign beef into the country. In the House, seven Republicans — including two Iowa members who had previously sided with the president on war powers — voted this week to remove U.S. forces from hostilities with Iran unless Congress authorizes the conflict. Rep. Maria Elvira Salazar of Florida released a television ad calling Trump’s immigration enforcement a betrayal of the Hispanic voters who delivered him the White House.
The pattern is consistent: Republicans in competitive races are picking fights with the administration from states with concentrated exposure to specific industries. They are not rejecting tariffs as a principle. They are not rejecting data centers as a technology. They are rejecting the political cost of staying in line — and calling that a policy position.
The engineering questions remain untouched.
On data centers: Trump’s recent dismissal of communities opposed to the AI buildout — “want to end up being backwards and poor” — is not merely offensive rhetoric. It is a factual claim about economic development that does not survive contact with the documentary record. Data center operators negotiate power purchase agreements for baseload electricity at rates that must eventually be socialized across the rate base if they exceed what distributed generation can supply. Microsoft’s twenty-year PPA with Constellation for the Crane Clean Energy Center at Three Mile Island — 835 megawatts, accelerated to 2027 with a billion-dollar federal loan guarantee — is a lock-in event, not a market signal. Meta’s 1.1-gigawatt commitment to Constellation’s Clinton Clean Energy Center in Illinois follows the same architecture. The hyperscalers are buying dedicated nuclear capacity to feed GPU clusters that serve inference workloads for a generation of large language models whose revenue projections remain unverified in SEC filings. The public subsidizes the grid upgrade; the hyperscaler captures the inference rent. That is the engineering reality the Husted amendment was trying — and failing — to address. Husted’s own party’s Senate campaign arm now calls the issue “the anchor hanging around Husted’s neck.” It is an anchor, but the weight is not political. The weight is the kilowatt-hours per token and who pays for them.
On AI regulation: Trump’s posture — dismissing calls for tighter federal regulation of artificial intelligence — is not an argument about innovation policy. It is an argument about enforcement architecture. The current generation of large language models are statistical pattern-matching systems trained on datasets of contested provenance and deployed at scales that introduce failure modes the systems’ own developers have not fully characterized. The alignment problem is not a hypothetical about future superintelligence. It is a present-tense engineering defect: a system trained to produce plausible continuations of text will, in deployment, produce confident-sounding outputs that are wrong in ways that are not visible to the operator without independent verification. Any engineer who has formally verified a cryptographic protocol — who has proven with mathematics that a system does what it says it does, and not what an adversary wishes it did — will tell you that the current generation of LLMs cannot meet that standard, because their specification cannot be written down. You cannot verify a system whose behavior you cannot enumerate. The AGI claim is not a technical specification. It is a fundraising document. Trump’s refusal to regulate is not light-touch governance. It is the absence of a verification requirement for systems that are already being deployed in hiring, lending, criminal-legal, and medical-diagnostic contexts with no mandatory audit, no mandatory disclosure of training data provenance, and no mandatory right of appeal for the people the systems affect.
On Flock Safety: the endorsement of automated license plate reader networks and facial-recognition-adjacent surveillance platforms by the president of the United States is not a law-and-order position. It is a surveillance architecture decision with specific engineering characteristics. Flock Safety’s system captures vehicle make, model, color, and partial plate data at scale, stores it in a cloud database, and makes it queryable by any law enforcement agency with access — no warrant required for bulk queries in most jurisdictions. The system is a textbook instance of the shitty technology adoption curve that Doctorow describes: surveillance technology deployed first in contexts where subjects lack political power — traffic enforcement, parking lots, suburban homeowners’ associations — normalizing the apparatus before it reaches contexts where the subjects include journalists, protesters, and political opponents. A senator or representative who “takes a position” on Flock Safety without addressing the engineering architecture — the storage model, the query permissions, the absence of a warrant requirement, the lack of a data-retention limit — has not engaged the issue. They have performed engagement while avoiding it.
The political structure of these breaks is transparent. Sens. Collins and Marshall are protecting specific industries in their states from tariff damage. Rep. Hinson is protecting Iowa’s agricultural economy. The seven House Republicans who voted on the Iran war powers resolution represent districts where the energy price shock from the conflict has made the war’s cost visible at the gas pump. Sen. Husted tried to address the data center energy-cost question and was defeated by his own party’s leadership. Salazar is protecting her Hispanic-voter coalition in south Florida. Each break is a rational calculation that distance from Trump on one issue is the price of general-election survival in a competitive district. That is politics. It is not policy.
The difference matters because the engineering questions are the ones with structural consequences. Tariffs on specific industries are bad trade policy, but they are adjustable by executive action and do not create path-dependent lock-in. A war in Iran can be ended by a cease-fire, and its energy-price effects will dissipate with the conflict. But the data center buildout — the hyperscaler PPAs for dedicated nuclear capacity, the grid-upgrade costs socialized across ratepayers, the surveillance infrastructure deployed without warrant requirements, the AI systems deployed without mandatory verification — these are architecture decisions that, once made, are expensive and politically difficult to reverse. They are the stack: cloud, CDN, DNS, submarine cable, chip, mobile duopoly, ad-tech, surveillance layer. Each layer consolidates further with each quarterly earnings cycle. A politician who breaks with Trump on tariffs but cannot explain the kilowatt-hours-per-token cost of the GPU cluster replacing the shuttered paper mill is not addressing the extraction. They are managing the politics of it.
Sen. Cornyn called it a “mistake” for Trump to put himself at the center of the midterms. “I don’t see any of that appealing to independents or any disaffected Democrats,” he said. That is a political observation, and it is probably correct. But the deeper mistake is the one neither party is willing to name: that the midterms are being fought over the political surface of issues whose actual substance is engineering, and that the engineering substance — who pays for the grid, who owns the data, who verifies the algorithm, who warrants the camera — is the thing that will determine whether these extraction architectures become permanent.
Collins, the lone Senate Republican seeking re-election in a state Harris won in 2024, estimated Trump’s proposed $5,000 dividend checks would cost $1.2 trillion. “I don’t know where that money comes from,” she said. A fair question. But the money is also coming from the ratepayers who will pay for the grid upgrades that serve the data centers Trump wants to build, from the users whose data trains the models Trump refuses to regulate, and from the communities whose streets are patrolled by the cameras Trump endorses without a warrant requirement. Collins can see the fiscal arithmetic. She is not looking at the engineering one.
There is a public consultation open at the FCC on data center interconnection and grid reliability. The deadline matters, because deadlines are the only part of regulatory processes that the regulated actually respect.