The AI Race Has Entered Its Super PAC Era

The Short Version

AI policy is becoming campaign infrastructure.

Not a white paper.

Not a panel.

Not a vague "Washington is watching" sentence in a model launch post.

Campaign infrastructure.

Business Insider reported on July 26 that groups tied to major AI and tech figures have already spent more than $65 million ahead of the 2026 U.S. midterm elections, with tens of millions more in reserve. The story names networks connected to OpenAI president Greg Brockman, Anthropic, and Meta, while noting that the full picture is hard to see because some affiliated groups do not disclose spending on the same schedule as super PACs.

That is the useful part.

Not "AI companies are lobbying."

Of course they are lobbying.

The useful shift is that AI regulation is becoming an electoral product surface.

Who gets elected now affects which state AI bills live, whether federal rules preempt state rules, how hard model reporting gets, whether advanced models face external testing, how data centers are permitted, how open-weight models are treated, and whether the next big safety incident becomes a law or a press cycle.

Very normal industry maturation. First you ship the chatbot, then you ship the super PAC.

Update: the money is not only campaign money now

OpenAI just added the quieter version of the same move.

On August 17, the company announced 14 grants for independent policy projects tied to its "Industrial Policy for the Intelligence Age" agenda. The total is small by frontier-lab standards: $1 million in funding and up to $1 million in API credits. That is not data-center money. It is not election money. It is not even one impressive GPU cluster.

But it matters because of what the grants are trying to turn into objects.

OpenAI is funding work on AI labor disruption scenarios, a "Right to AI," AI dividend ideas, data-center policy, person-based benefits, tax policy, clinical AI infrastructure, research-institution readiness, recursively self-improving AI incident taxonomies, AIxBio information sharing, national-security safety training, auditable government policy simulation, and AI-enabled democratic accountability.

That list is not a normal corporate social-responsibility pamphlet.

It is a map of the policy surfaces AI companies expect to fight over.

Jobs.

Energy.

Tax.

Hospitals.

Biosecurity.

Government simulation.

Public oversight.

Model self-improvement.

Very casual. The chatbot has started funding the committee agenda.

Policy prototypes are power too

This is the useful extension of the Super PAC story.

Campaign money shapes who gets to write the rules.

Policy grants shape what counts as a reasonable rule once those people arrive.

That distinction matters. A white paper can be ignored. A funded pilot with a dataset, schema, cost model, prototype, evaluation framework, or public report is harder to dismiss. It gives lawmakers something to point at. It gives agencies vocabulary. It gives journalists a frame. It gives industry allies a safer sentence than "please regulate us less."

OpenAI says these choices should not be made by technology companies alone. Good. It also says more than 400 people and organizations responded to the call, and that projects will run for six months with results reported in 2027.

Also good.

But independence is not magic.

The agenda, money, credits, convening power, and public framing still come from the company whose products sit at the center of the debate. That does not make every project captured. It does mean readers should track the dependency.

Who chose the questions?

Who gets API credits?

Which problems are made legible?

Which ones are left unfunded?

Which pilots become the default vocabulary when government finally moves?

This is not a conspiracy.

It is institution-building.

And institution-building is political.

The skeptical reading is necessary

There is a reason the source mix matters here.

OpenAI's own post frames the grants as a way to broaden the circle of people developing answers for the Intelligence Age. Its earlier Industrial Policy post said the ideas were exploratory and invited others to build on, refine, or challenge them. The OpenAI Foundation has separately said it expects to invest at least $1 billion across life sciences, jobs and economic impact, AI resilience, and community programs.

Those are real commitments.

They are also brand, strategy, and political positioning.

Tech Policy Press called the industrial-policy agenda a "policymercial," arguing that OpenAI was wrapping product expansion and public-interest language together. The Guardian reported that major AI companies were using policy papers, institutes, and lobbying to reshape the public narrative while trust in AI declined.

That critique should not make the grants worthless.

It should make the evaluation stricter.

Do the projects publish methods?

Do they challenge OpenAI's assumptions?

Do they produce evidence that a government, union, hospital, school, utility regulator, or civil-society group could use without becoming dependent on OpenAI?

Do they study AI as a product that can be regulated, not only as an inevitable force that society must adapt around?

If yes, useful.

If no, the policy ecosystem just got a nicer brochure.

Builders should watch the boring outputs

The practical takeaway is not "OpenAI gave some grants."

The practical takeaway is:

Watch which policy objects become reusable.

If AEI and Urban produce AI labor-disruption indicators, companies may eventually be asked whether their deployment matches those indicators. If the Abundance Institute builds a data-center atlas layer, local energy fights may get a more standardized vocabulary. If NTU produces privacy-preserving agent reports for government policy simulation, public agencies may start asking vendors for auditable model-backed forecasts. If IST creates an incident taxonomy for uncontrolled recursive self-improvement, future safety disclosures may inherit that language.

That is how soft power becomes hard paperwork.

Not immediately.

Not automatically.

But often enough to matter.

The AI policy fight is not just happening in Congress, state races, and agency offices. It is happening in grants, prototypes, research agendas, measurement frameworks, and the innocuous-looking API credits that let people build the first version of the system everyone later cites.

First the lab funds the idea.

Then the field tests the idea.

Then the government borrows the idea.

Then the builder discovers the idea is a requirement.

This is not one AI industry

The lazy version of this story says "Big AI wants less regulation."

That is too simple.

There are at least two visible political strategies now.

One is the pro-innovation, lighter-regulation strategy. Leading the Future, a pro-AI network backed by tech executives and investors, says on its own site that it wants to identify, maintain, and grow pro-AI candidates at the state and federal level. Axios reported that Leading the Future ended Q2 with $31 million after transferring $20 million to affiliated groups, including Think Big PAC and American Mission PAC.

That side wants AI development to remain fast, national, competitive, and less fragmented by state-by-state rules.

The other visible strategy is the safeguards strategy. Anthropic said on July 21 that it is donating another $20 million to Public First Action, bringing its total support to $40 million. Anthropic says the money is for public education and policy work, not candidate election spending. Public First Action is tied to PACs that support candidates who favor AI safeguards, and Axios reported that the group said it had raised more than $80 million by the end of June.

That side wants transparency, independent evaluation, stronger oversight, and policy mechanisms that could slow or block especially risky model deployments.

Both sides speak the language of American AI leadership.

Both sides speak the language of public interest.

Both sides are trying to build political machinery before the law hardens.

That is the story.

AI regulation is not just a debate between government and industry.

It is becoming a fight inside the industry over which version of government should show up.

Anthropic is not just buying safety halo

Anthropic's move is easy to caricature.

You can say:

"The safety lab is funding the safety PAC because regulation helps incumbents."

There is some truth in that suspicion.

Regulation can become a moat. Reporting requirements, external testing, security programs, deployment thresholds, and government verification are easier for rich frontier labs than for tiny model companies or open-source projects. A safety framework can be genuine and still make life harder for competitors.

Annoying, but true.

But it would also be lazy to treat Anthropic's position as pure theater.

Its July 21 post is unusually explicit. Anthropic argues that frontier AI companies should be transparent about model capabilities and risks, that governments should be able to verify safety claims, that civil penalties should enforce safe practices, and that governments may ultimately need a way to slow or block deployments that pose serious catastrophic risk.

That is not just "please regulate our rivals."

That is a worldview.

You can agree with it or not. You can worry about capture. You can ask who designs the tests, who audits the auditors, what happens to open models, and whether "catastrophic risk" becomes a flexible phrase that only the biggest labs can navigate.

Good.

Ask all of that.

But the important point is that safety politics is no longer a blog-post genre. It has budgets, organizations, donors, field strategy, state targets, and election calendars.

The safety debate got a ground game.

Meta is fighting the state layer

Meta's role is different.

Meta has its own model ambitions, but its political exposure is also much broader. It has social apps, ads, youth-safety fights, content rules, privacy obligations, commerce surfaces, and now consumer AI agents that want to operate across ordinary life.

That makes state law terrifying.

Axios reported last year that Meta launched the American Technology Excellence Project, a nonfederal super PAC focused on state candidates, with tens of millions of dollars behind it. Meta framed the effort as a response to a patchwork of state-level tech and AI proposals that it argues could hurt U.S. competitiveness.

This is a useful reminder:

AI policy is not only federal frontier-model policy.

It is also California.

It is state privacy law.

It is child-safety law.

It is data-center permitting.

It is procurement.

It is employment law.

It is consumer protection.

It is whether a state legislature can create obligations faster than Congress can preempt them.

If you are Meta, that state layer is not some constitutional abstraction. It is product risk, ad-stack risk, recommender-system risk, training-data risk, youth-safety risk, and agent-permission risk.

The AI race is becoming local in very expensive ways.

OpenAI has a distance problem

OpenAI's position is more delicate.

Business Insider has previously reported that Greg Brockman and his wife gave $25 million to Leading the Future. OpenAI later said the company does not direct the group's activities, has not donated to super PACs or political campaigns, and does not have an employee-funded PAC.

That distinction matters.

Do not casually turn one executive's political giving into "OpenAI donated."

But the reputational problem remains.

When a senior leader at a frontier lab funds a major pro-AI political network, the company cannot expect the public to experience that as a purely private hobby. The model lab, the executive, the policy agenda, the donations, and the industry fight all blur together.

That is especially true because OpenAI is already politically central. It is working with government on model access, long-horizon safety, cyber capability, enterprise deployment, infrastructure, and the broader argument that the U.S. should keep frontier AI leadership.

So the distance line is real.

It is also thin.

This is one of the weird new responsibilities of frontier AI leadership:

The personal political behavior of senior executives can become part of the company's trust surface.

Very elegant. Governance now includes your cofounder's checkbook.

The policy fight is a product fight

Builders should care about this even if they never watch campaign ads.

Policy will change product surfaces.

If federal law preempts stricter state rules, AI companies get one kind of deployment environment.

If states keep moving, companies get another.

If advanced models require external testing before release, release cadence changes.

If open-weight models face restrictions, model choice changes.

If data centers hit moratoriums or local power fights, compute plans change.

If AI hiring, lending, insurance, education, health, or child-safety rules tighten, app design changes.

If export controls widen, access changes.

If the government creates reporting duties for frontier incidents, observability changes.

This is not background politics.

It is roadmap risk.

The companies funding these political networks know that. They are not spending because they enjoy FEC forms. They are spending because legal affordances become product affordances.

What your model is allowed to do is now partly an election outcome.

Read the money, not just the slogans

There is a trap here.

Each side wants the high ground.

The pro-innovation side says excessive regulation will slow American AI, help China, fragment markets, and deny people the benefits of new technology.

Sometimes that is true.

The safeguards side says uncontrolled deployment will put workers, children, critical infrastructure, and democratic institutions at risk.

Sometimes that is true too.

The money does not automatically make either argument false.

But the money changes how we should read the argument.

When a company says "innovation," ask which liability it wants to avoid.

When a company says "safety," ask which competitors it may accidentally or deliberately burden.

When a PAC says "parents," ask which product risk is being reframed.

When a group says "American leadership," ask whether it means open competition, closed national champions, data-center buildout, export controls, or all of the above depending on the room.

AI politics is going to be full of morally flattering language.

Treat that language as packaging.

Then read the incentives.

What this changes

The practical takeaway is not "AI companies are bad because politics."

That is too easy.

Large industries participate in politics. AI was never going to be exempt. A technology that affects work, speech, education, science, defense, energy, privacy, children, software, and national competition was always going to become an election issue.

The useful takeaway is sharper:

AI governance is moving from principle statements into power-building.

That means the next phase of AI policy will be shaped by:

• money
• state races
• federal preemption fights
• local data-center politics
• model-release incidents
• open-weight geopolitics
• safety groups that need credibility
• pro-innovation groups that need public trust
• companies trying to turn their preferred compliance model into the default

For builders, investors, workers, and users, this means AI regulation is no longer something to check once a quarter.

It is part of the operating environment.

The question is not just:

What can the model do?

It is:

Who is trying to decide what the model is allowed to do?

The AI race has entered its super PAC era.

And that means the next benchmark may be a ballot.