Sanders Wants the Public to Own Half of AI. Half of What, Exactly?

The most interesting AI proposal of the summer didn't come from a lab. It came from an 84-year-old senator from Vermont.

At the beginning of June, Bernie Sanders published an op-ed in The New York Times proposing that the American public take ownership of half of the frontier AI companies — OpenAI, Anthropic, xAI and the rest. The mechanism: a one-time 50% tax on their stock, paid in shares instead of cash, with the shares going into a sovereign wealth fund that would pay dividends to every American. On June 18 he turned it into an actual bill, the American AI Sovereign Wealth Fund Act. The coverage split along the usual lines within hours: half the headlines said "socialism", the other half said "finally".

I read it from a different seat. I'm not a policy person; I run engineering teams that build on top of these models every day, and I've spent the last two years arguing — in production, with invoices to show for it — that the model is a commodity and the value lives in the layer built around it. So my first reaction to the op-ed wasn't political. It was an engineer's question: half of what, exactly?

And I want to be upfront about what this post is. I usually come to this blog with a thesis and defend it. This time I don't have one — not out of prudence, but because I've turned the proposal over for weeks and I keep finding good arguments pointing in opposite directions. What I can offer is the builder's read of what's actually on the table, and the questions I think we should be arguing about instead of the ones the headlines picked. Consider this an invitation, not a verdict.

The proposal is more concrete than the headlines suggest

The op-ed version is simple: tax half the equity of the biggest AI companies, once, and put it in a public fund. The bill version is more interesting, and almost nobody read past the headline to notice.

The June 18 text doesn't only target the labs. It defines "systemically important AI activity" to include AI data centers, computing infrastructure, AI services and advanced robotics, and applies the 50% equity tax to companies in those businesses with gross receipts above $200 million. It even forces structural separation: within 90 days, the AI operations have to be split from the non-AI business — no shared directors, no cross-holdings. The fund would hold voting shares, take board seats, and pay annual direct distributions to Americans.

Sanders' core argument is one sentence long, and it's aimed straight at people like me and probably you: "A.I. is built on our collective intelligence: our books, songs, artwork, journalism, computer code, scientific research, videos, conversations, images and ideas spanning generations." The stated goal is that the trillions AI may generate get "used to improve the lives of all of us — not simply to make the richest people in the world even richer."

Whatever you think of the remedy, the diagnosis deserves engagement rather than a reflexive eye-roll. So let me engage with it.

«Built on our collective intelligence» is the argument I can't wave away

I've been writing code since the late nineties. Some of it sits in public repositories and old forum threads. Statistically, fragments of it — and of yours — are inside the weights of every frontier model, alongside the journalism, the books and the Stack Overflow answers of a few hundred million other people. Nobody asked us. Nobody paid us. And the companies that did the scraping are now valued in the hundreds of billions.

None of that is contested, and I notice that the people quickest to dismiss Sanders rarely address it. Even Sam Altman didn't: his response, reported by the Associated Press, was that he also wants the public to hold equity in AI companies and would advocate for the general idea alongside Sanders — his objection was to the 50% figure, not the principle. When the CEO of the main company being targeted concedes the premise, the debate has moved further than the "socialism vs. innovation" framing admits.

I have to concede something in the other direction too, because honesty cuts both ways. The Cato Institute's counter-argument is solid: Norway and Alaska built their sovereign funds on resources the state already owned — oil under public land. Sanders proposes transferring half of privately built companies, which is a different legal and moral animal, and the constitutional challenges under the takings clause would be immediate and serious. Both things can be true: the training-data grievance is real, and a forced 50% transfer may be the wrong instrument for it. That tension is exactly what I'd like to see debated seriously.

Half of a frontier lab is not half of AI

This is the part where my seat gives me something the political coverage missed, so let me spell out the mechanism.

What does a frontier lab actually own? Four things: model weights, talent, compute contracts and distribution. Look at each with a builder's eyes. The weights depreciate brutally — the models we were paying premium prices for eighteen months ago are today's discount tier, leapfrogged two or three generations over. The talent walks out the door whenever a competitor offers more. The compute mostly belongs to the cloud providers. What holds value is the distribution — the brand, the default position in a hundred million workflows — and even that is being fought over quarter by quarter.

Meanwhile, everything I see in production points the other way: the durable value of AI accrues in the layer built on top of the models, not in the models themselves. The workflows, the integration with real systems, the verification, the guardrails — the harness. That layer is distributed across thousands of companies, and no equity tax on a handful of labs captures it.

So the honest engineering question about the proposal is this: if the public takes half of the labs, is it taking half of AI's future value — or half of a treadmill, where every asset has to be rebuilt every eighteen months just to stay in place?

I can argue it both ways, and that's precisely why I'm not writing a verdict. Because the bill's fine print cuts back the other way: by including data centers and computing infrastructure, it reaches the one layer that behaves like durable capital. Weights depreciate; gigawatts don't. If AI consolidates the way electricity did — Andrew Ng's old comparison, which the Fortune coverage of this proposal picked up — then the utility layer is exactly where a public stake would make long-term sense, and Sanders' drafters were smarter than his headline. A Carnegie Mellon study cited in the same coverage projects data centers could raise US electricity bills 8% by 2030; the public is already an involuntary stakeholder in this build-out, on the cost side only.

Public stakes in tech stopped being a left-wing hypothetical this year

The context that makes this more than a Sanders story: the Trump White House currently holds equity stakes, warrants or golden shares in roughly twenty private companies — Intel among them. Elizabeth Warren has published her own AI equity-tax proposal. Altman says he's open to the general idea. Whatever your politics, the observable fact is that both ends of the American political spectrum now practice state ownership of strategic tech companies; they differ on who benefits and how the shares are acquired, not on whether the state should hold them.

For those of us reading from Europe, this is a strange inversion. State stakes in strategic industries are the least exotic thing in the world here — and yet the most instructive European precedent argues against the naive version of the plan: Norway's sovereign fund invests exclusively outside Norway, precisely so the state never sits as shareholder in the companies it regulates. Sanders' fund would do the opposite: the same government writing AI safety rules would hold half the equity of the companies subject to them. If you want a single question to take away from this post, take that one — can a state be both referee and shareholder of the same industry, and what does the answer do to regulation the day the dividends start funding the budget?

The questions I'd put on the table before picking a side

A diagnosis without action is just an opinion, so here is my action — except this time it's not three bets, it's five questions. These are the ones I'd want argued by people smarter than me in each dimension, and the ones I'll be testing my own position against:

1. Which layer holds AI's durable value — models, compute or applications? If it's the models, an equity stake in the labs captures it. If it's compute, the bill's data-center reach matters more than its headline. If it's the application layer — my two years of production experience say it is — then no lab-focused instrument captures much at all, and the wealth distribution question needs a different tool entirely.
2. Can the referee hold shares? Norway says no and invests abroad. The Cato objection about regulatory capture in reverse — a state that profits from the companies it polices — deserves a real answer, not a shrug.
3. If the claim is «trained on everyone's work», why does compensation stop at the border? The training corpus was global. My Catalan blog posts are in there next to your code. A US-only dividend built on the global commons rests on the same logic it denounces — extraction without compensation, one level up.
4. What do the incentives do at the margin? A 50% equity tax triggered at $200 million in gross receipts is also a message to every company approaching that line: stay small, restructure, or leave. Does frontier work move offshore, into open weights, or into corporate structures the bill can't reach?
5. What's the strongest alternative instrument? Antitrust, data dividends paid to creators, compute as regulated public infrastructure, or plain corporate taxation that actually collects. Ownership is one tool among several; the op-ed compares it to none of them.

I keep coming back to the question I started with: half of what. The op-ed answers "half of the AI companies". The bill quietly answers "half of the AI infrastructure". And the engineer in me suspects the real wealth will sit in a third place — the millions of harnesses built on top — which neither version touches. Whether that makes the proposal pointless or merely incomplete is the debate I want to have.

I wrote about the adjacent question — what sovereignty over AI infrastructure actually means, beyond where the data sleeps — in AI sovereignty is infrastructure, not data residency. This one ends differently than my posts usually do: not with my answer, but with a request for yours. If you read this from a seat I don't sit in — policy, law, a lab, a treasury — tell me where the engineer's read fails. I'd rather update than be quotable.