The Take — The AI rules are being written in Ohio, not Washington
The most consequential AI regulation in America this year was passed by township trustees, county commissions and one governor with an executive order. I think that is the correct outcome, and the industry's plan for the morning after Tuesday's vote is an attempt to trade a working veto for a press release.
Start with what the ballot actually contains. Thirty-six governorships are up, and at least one candidate calling for a data center moratorium appears in 16 states, according to the MultiState tracker that both The Information and Wired are working from. Fewer than ten candidates nationwide pledge unconditional support. Wired counted at least 20 moratorium-friendly candidates; our own count of the moratorium races put the competitive set at five or six, with Texas the only large state where a governor's record is the live argument.
But the governor races are the small print. Ohio has at least 125 local moratoriums in effect and 14 local ballot measures on Tuesday alone, per MultiState. Twenty-eight localities across six states vote on data center measures this cycle, and Kansas towns are weighing ten-year bans. This is not a national debate about AI safety. It is a few hundred land-use decisions, and they bind in a way no federal framework currently does.
The only AI rule that has actually stopped anything
Here is the test I keep applying: name a rule that measurably changed what a lab or a hyperscaler built in the last twelve months. The federal answer is thin. A June executive order asks labs to submit frontier models for government testing up to 30 days before release — a window cut from 90 days after industry lobbying — with an August 1 deadline for the framework that governs the tests. The framework was finished in early August. It has not been published, and the administration told The Information that it need not release the draft, citing national security and commercial exemptions. No published criteria, no published participants, no published results. As a constraint on behaviour, that is a suggestion with a filing cabinet.
The local answer is not thin. Local opposition blocked $68 billion of US data centers in a single quarter this year, which we tracked project by project. Texas Governor Greg Abbott froze data center permits pending a grid review in August and told an audience the industry had "dug their own grave." New York has a statewide moratorium by executive order; a September Siena poll found 63 percent of New Yorkers would oppose a data center in their area even after it expires. Those are outcomes. They moved steel and capital.
That asymmetry is the argument. A rule a regulated party drafts, applies to itself and can revise is not a rule; it is a communication. A county board that can deny a conditional-use permit is a regulator whether or not anyone in Washington calls it one — and its accountability story is better, because the people who bear the substation and the water draw are the people who vote on the permit.
What is being offered in exchange
OpenAI's post-election campaign, which we covered when it leaked Sunday, has four planks: mandatory safeguards for the most capable systems under the name "speed limits," community compacts plus model state legislation for data centers, a worker-disruption agenda, and US-shaped global standards. Two of them move authority away from the level that has been saying no.
"Speed limits" has no number, no road and nobody holding a ticket book. OpenAI has not said what the threshold is, whether it binds training runs or releases, who sets it, or which agency enforces it. The phrase does political work, not regulatory work — Chris Lehane's framing, that people should "always remain in control," is a mood, not a mechanism.
The data center plank is the more revealing one. "Model state legislation" means a template, and templates written by the largest company in an industry have a way of becoming the default text in forty legislatures that lack the staff to draft their own. "AI compute campus" is a rename of the thing voters are objecting to. Lehane calls the backlash healthy — fair enough — but a community compact drafted by the applicant is not the community setting terms.
The steelman
The case against me is strong, and it is not stupid. A single national market for models is worth a lot: a lab cannot ship one system into fifty regulatory regimes, and the compliance cost falls hardest on smaller developers, not on OpenAI. Local vetoes are also regressive in a way their defenders rarely admit — wealthy counties refuse and poorer ones take the water bill and the tax base, which makes an anti-data-center vote in Marin partly a decision to put the load in Kern. County supervisors are not equipped to judge transformer loading or interconnection queue position, and a process that rewards the loudest public comment selects for the most organised neighbour, not the best site. And a federal floor can in principle be stricter than the modal state.
And there is a fair-reading problem with my framing: OpenAI is not asking Congress to preempt the states. Model legislation is the opposite of preemption. If the company wanted to kill local control it would say so, and there is precedent — West Virginia in 2025 and South Dakota this year passed laws stripping local authority over data centers, and Maine's governor vetoed a statewide moratorium on local-control grounds. Those are the real preemption moves, and OpenAI did not propose one.
Why the take still holds
Because a standard with no number, no enforcer and an unpublished test regime is not a floor. It is a ceiling on what anyone else can demand. Once "we support speed limits" is the industry position, the next state legislator who proposes an actual threshold is arguing against the labs and against the labs' stated position — the debate shifts from whether to constrain to how to word a press release. That is the practical function of a voluntary framework: it consumes the political demand for a mandatory one.
And the local layer is not a permanent patchwork. It is converging: New York wrote a statute, Virginia rewrote its approval process, Texas paused permits administratively, and the states that stripped local control did it with legislation. Counties test the politics; states codify the winners. Slower than a federal rule, and far more likely to survive a change of administration.
The tell is the vocabulary. Nobody renames a building because the physics changed. They rename it because the permit hearing changed. If the constraint were federal, the industry would be lobbying for a threshold; instead it is hiring people to write model bills and rebranding the product. Both are responses to local power, and both are cheaper than fixing a rate case.
What would change my mind
Three things, in order of weight. A federal statute with a published threshold, a named independent inspector and a public test report — not an executive-order framework behind a national-security exemption. A lab that lobbies for a stricter state law than the one it drafted, in a state where it has a project pending. And evidence that local denial only relocates capacity without delaying it: if the $68 billion that got blocked shows up as $68 billion built two counties over and eighteen months later, the veto is theatre and the industry's complaint about patchwork is the honest one.
Until then, the honest description of AI governance in the United States is this: the labs write the principles, the counties decide the projects, and the only rule anyone can point to that changed a decision came from a zoning hearing in Ohio.
If a county can stop a data center but nobody can publish a test result, which level of government is actually regulating AI? Write to us via our contact page.
