The headline number from this midterm cycle is small enough to be overlooked. Political spending tied to artificial intelligence safety reached roughly $110 million, according to a Bloomberg tally published this week, and more than half of it came from a single network.
Public First, the group backed by Anthropic chief executive Dario Amodei and others, put more than $63 million behind candidates who favour stricter rules for AI developers. Most of the districts where those candidates ran returned results favourable to them. On the same ballot, 40 of 41 local referendums on data centers passed measures restricting construction. Two results, one lesson. The money moved a handful of races. The zoning changed the map.
Start with what $110 million buys. Total spending on the midterms runs into the billions, so the AI safety network was a minor participant by the standards of a party committee or a pharmaceutical lobby. What a focused donor can do with a sum that size is pick primaries, where turnout is low and a few million dollars of advertising is decisive, and support candidates in districts that are already competitive. That is a real achievement and a narrow one. The correlation between spending and outcomes also proves less than it appears. Candidates who favour AI regulation tend to run in urban and suburban districts where voters already hold those views, and a group that chooses its races well will look effective whether or not its advertising changed any minds. The honest reading is that Public First bought access and credibility with a set of likely winners, which is what a new political operation needs in its first cycle.
The referendums are the more interesting result, because 40 out of 41 is not a partisan number. It is a land-use number. Voters who were asked whether a data center should be built near them said no, and the reasons they gave were water, electricity, noise and property values. Those objections do not divide along the lines that AI policy divides along. A county supervisor who has to answer for a substation and a drying aquifer is not making a statement about model evaluation. That distinction explains why the two results can coexist without contradiction. A voter can want the government to slow down artificial intelligence and also want the data centers that make it work somewhere else. Both positions are about who gets to decide, and both are locally popular.
For the industry, the referendums are the expensive half. A federal disclosure requirement is a compliance cost that a large laboratory can absorb and pass on. A county moratorium is not appealable to Congress, cannot be litigated on First Amendment grounds, and arrives with a planning calendar rather than a statute. Forty restrictions at the local level will shape where compute gets built more than any bill the next Congress passes. That is the asymmetry the AI safety money did not address. Public First spent its resources on federal races, where the levers are visible: hearings, disclosure rules, appointments to the agencies that review models. The decisions that will determine whether a training cluster can be powered in Ohio or Virginia are made by boards of supervisors who have never been lobbied by a laboratory and do not expect to be.
There is a case that the spending worked anyway. A network of candidates who owe a new political group a meeting is an asset that compounds across cycles, and $63 million is a cheap entry price for a sector that expects to be regulated. Anthropic’s own posture, publicly in favour of testing regimes and transparency requirements, becomes easier to defend when legislators who share that view hold seats. The counterargument is that the industry has now learned which message travels. Candidates who talked about safety in the abstract won where the districts were friendly. Ballot measures that talked about water and electricity won almost everywhere.
The second message is more durable because it is about a specific building, and no amount of Washington spending changes a local planning decision. A laboratory can hire a lobbying firm and a compliance department. It cannot easily hire the county board that has to approve a substation, and it cannot argue that a moratorium is preempted by a federal law that does not exist. The campaigns that beat the data centers did not need to know what a training run is.
What happens next is likely to look less dramatic than either side expects. A Congress with more members sympathetic to disclosure will hold hearings, introduce bills and probably pass something narrow, most likely around reporting requirements for large training runs. Preemption of state and local rules, the outcome the industry most wants, will be harder to sell to members who just watched 40 of 41 referendums go the other way.
The $110 million bought a seat at the table. It did not buy the table. That distinction will matter more in 2028, when the same donors will have to decide whether to keep funding federal races or start showing up at county planning meetings.
