AI Labs Bought Both Sides of the AI Regulation Fight

AI-focused super PACs have raised more than $200 million for the 2026 US cycle and spent over $55 million in federal races alone. The most expensive fight so far was not the industry against regulators. It was one frontier lab’s political network against another’s, over a single New York state assemblyman, and the side arguing for regulation outspent the side arguing against it and still lost.

That result breaks two comfortable stories at once. It breaks the story that AI money simply buys outcomes, because the bigger spender lost. And it breaks the story that there is a public-interest side in this fight with independent funding, because there is not. Both sets of millions trace back to people who build frontier models.

The scoreboard

Two networks dominate. Leading the Future wants a single federal AI framework that preempts state law. Public First Action opposes federal efforts to freeze state progress without adequate federal safeguards. Both are funded from inside the industry they are arguing about.

Leading the Future has backed 28 candidates and 25 of them won their primaries. It has endorsed 18 Republican state legislative candidates across Arizona, Florida, Georgia, Michigan, Mississippi, Pennsylvania, Texas and Utah. Public First Action backed 11 candidates and 10 won. Read on win rates alone, the industry money looks unstoppable. The most expensive individual race says otherwise.

The race that defines the fight

Alex Bores sat in the New York State Assembly and co-authored the RAISE Act, the state frontier-model safety law. He ran for the Democratic nomination in New York’s 12th congressional district. Leading the Future spent more than $7.6 million against him by June, in what has been described as an $8 million campaign against one state legislator seeking a House seat.

The other side spent more. Public First Action and its Jobs and Democracy vehicle put roughly $10 million behind him, and total independent expenditure on the race reached about $12 million. Bores finished second and lost the primary to Micah Lasher. The New Yorker described the contest as a proxy battle between OpenAI and Anthropic, which is the most accurate three-word summary available and also the most uncomfortable one.

Why “proxy war” is the wrong thing to feel reassured by

The usual reading of a proxy war is that it is a balance. One lab funds restraint, another funds speed, the two cancel out, and the public gets a genuine contest of ideas. That reading requires ignoring what both sides have in common.

Both networks are financed by organisations whose revenue depends on frontier models being built and sold. They disagree about the rules. They do not disagree about who should be at the table when the rules are written, and the answer both are funding is: the labs. A disagreement between two well-capitalised factions of the same industry is not a check on that industry. It is a narrowing of the debate to the questions that industry finds worth $200 million.

The absent party is anyone whose interest in AI policy is not commercial. Teacher and professor unions funded the Guardrails Alliance, which put about $285,000 behind Bores. Set that next to $7.6 million on one side and $10 million on the other and the shape of the room is clear.

The law they fought over had already been weakened

The RAISE Act passed both chambers of the New York legislature in June 2025. Governor Kathy Hochul significantly weakened it before signing it in December 2025. What survived requires large frontier developers to publish annual risk assessment and mitigation frameworks, to disclose serious safety incidents within 72 hours, and it gives the state attorney general enforcement up to $1 million for a first violation and $3 million for subsequent ones.

Those are disclosure duties with modest penalties attached. For a company raising at tens of billions in valuation, a $3 million repeat penalty is a rounding error, and the $8 million spent attacking the law’s author exceeded the maximum fine the law can impose by a factor of nearly three. That comparison is the clearest statement of priorities in the entire episode.

Hochul also gutted a separate AI protection bill hours before a federal executive order aimed at stamping out state AI laws. State-level AI regulation in the United States is being negotiated down before it reaches a governor’s desk, and then contested at the ballot box anyway.

What preemption actually decides

Preemption sounds procedural. It is not. A single federal framework replacing 50 state regimes is a genuine efficiency argument, and it is the argument Leading the Future makes in public: maintaining compliance state by state in piecemeal fashion is difficult. That is true.

The part that is not said out loud is that preemption also collapses 50 chances to pass a binding rule into one chance, in the venue where the industry’s lobbying advantage is largest and where nothing has passed yet. Removing the states removes the only level of American government that has actually legislated on frontier models. The efficiency argument and the deregulation outcome are the same policy, and only one of them appears in the advertising.

This is the mirror image of what the European Union chose, and the contrast is useful for anyone building on either side of the Atlantic. The EU wrote one binding framework with real obligations rather than arguing about which level of government gets to write it, which we covered in our EU AI Act guide for builders. The American question is not what the rules should be. It is whether there will be enforceable ones at all.

What this changes if you build with AI

Political spending stories usually stay in the politics section. This one has direct consequences for architecture and procurement decisions being made now.

• US regulatory certainty is further away, not closer. A contested preemption fight running through a midterm and into a general election means the American rule set stays unsettled into 2027. Plan for the rules to arrive late and then arrive quickly.
• The EU framework is now the stable one. If you need a compliance target that will still exist in eighteen months, the European obligations are the only ones with a fixed shape. Building to them is the lower-variance choice even for a US-first product.
• Vendor policy positions are now a procurement input. Your model provider has a declared position on whether the jurisdiction you operate in should be allowed to regulate it, and has spent real money on that position. That is legitimate information when choosing a long-term dependency.
• Disclosure duties are the surviving control. What made it through in New York was reporting: annual risk frameworks and 72-hour incident disclosure. Whatever passes elsewhere is likely to have the same shape, so the capability worth building now is the ability to produce that evidence on demand.

Why it matters

The AI industry spent 2025 arguing that it wanted to be regulated. In 2026 it discovered it could afford to specify how, and it turns out the industry does not agree with itself. That disagreement is real and the money behind both halves of it is real.

What the New York result shows is that the outcome is not simply purchasable, which is the one genuinely encouraging fact in the whole picture. What it also shows is that a state legislator who writes a frontier-model safety law can expect an eight-figure campaign against him, funded by the companies the law would cover. Every other state legislator now knows the price of writing one.

The question worth asking of any AI policy position in the next two years is not whether it sounds responsible. It is who paid to put it in front of you, and whether anyone without a commercial stake in the answer had the budget to be heard at all.

Related reading

• The government stake in OpenAI on the other direction of the same relationship.
• The US switched off frontier AI and China open-sourced it on what happens when policy and capability diverge.
• The secret frontier AI review and the open-weight exemption on rules written without a public record.
• The open-weight window is closing on how regulation reshapes what you are allowed to self-host.
• EU data sovereignty for AI in 2026 on building for the jurisdiction with settled rules.
• AI agents with money and no spending limits on the controls nobody shipped while this argument ran.