--- id: dc19699e-c8ed-4479-99b4-ec9b44ffcd2a title: The Best Part of an AI Company Fits on a Flash Drive. That Quietly Decides Who Can Own It. createdAt: 2026-06-06T17:37:44.826872Z tags: [#sat-news, #ai] --- # The Best Part of an AI Company Fits on a Flash Drive. That Quietly Decides Who Can Own It. This spring a U.S. senator proposed that the public should own half of the biggest AI companies — a one-time levy of 50% of their shares into a federal fund. In the same days, the White House said it was talking to OpenAI about a government stake, and noted it already holds stakes in about twenty firms. For a moment it looked like everyone agreed: the public should own a piece of AI. The agreement falls apart the moment you ask what each side can actually do. And underneath the politics sits one plain fact that decides more than any of it: the part of an AI company that is worth owning can walk out the door. ## Two very different things wear the same word "A government stake" can mean two opposite things. One is buying in: the government put money into Intel and got shares while the company was under pressure — the owners agreed and were paid. That is an ordinary deal; investors do it every day, and it breaks no rule. The other is taking: a 50% levy with no payment is a seizure. That would cross a line the modern Western economy is built on — the state does not simply take private companies. That line is still standing, and almost no one treated the bill as serious; its author has filed losing bills for forty years. The slogan was the point, not the law. ## Why taking does not work on this kind of company You can seize a farm or a factory because it cannot move — you can surround it. An AI company's value is the model, which copies to a drive, and the people who build it, who can leave. Announce "we will take half," give them time to prepare, and you are left holding the empty shell — the buildings, the brand, the U.S. revenue — while the thing you wanted is already gone. The only way to truly grab it would be by surprise, backdated, before it can move — which is exactly what the rule against backdated taking exists to stop. So here private property is protected twice over: once by the courts, and once by the simple fact that the asset can run. There is a quieter version of the same point. The most-watched company already has a large private owner with contracts running for years — Microsoft holds about a quarter of OpenAI. Any forced reshuffle runs straight into those existing rights. Whoever holds the contractual pen shapes what a new stake is even allowed to touch. The headline may say "the state took a piece"; the fine print decides whether that piece reaches anything that earns. The pen matters more than the share. ## The value is in the model — and what sells is its output, not the model itself If the worth sits anywhere, it sits in the trained model — the weights. But look closely at what actually changes hands when buyers "line up to pay": they buy access. They rent the output through an interface. The owner sells the use and keeps the asset. The model itself changes owners only rarely, and in special ways — licensing the weights, hiring the whole team, buying the company outright. That distinction decides the seizure question. A public that wants "a piece of AI" can be handed the benefit — cheap access, or a dividend funded by sales — without anyone owning the weights. But the weights themselves are kept, not sold; and as we saw, they cannot be taken by force either, because they can run. So the asset sits in a spot reached by neither ordinary purchase nor seizure: only its output flows out, by agreement. The computing power behind all this is a rented input and a way to charge for use — not the prize, and it matters less as models improve faster (the frontier now moves in months, not years). The fight over "owning AI" is really a fight over the output and the upside, because the asset itself is both unsold and unseizable. ## What an AI simulation suggested To pressure-test this reading, we ran a simple game: each major player — the U.S. government, the AI firms, China, Europe, large investors — making one move at a time over several rounds. This is a thinking tool, not reporting. And not everything a game produces is about the world; some of it is just a property of the game. The honest job is to tell the two apart. One result is worth keeping, because the rules do not build it in: every lock one side added raised the value of the one door no one could lock — the people leaving. A rival that simply held that door open (keep your patents, work remotely, here is funding) gained the most. That points the same way as the central fact: you cannot lock down a portable asset, so the open door wins. One result should be distrusted, because the rules probably do build it in: the game settled into a tidy picture where each player walked off with a different piece and every move stayed reversible. A turn-based game where players move one at a time and can always step back will manufacture exactly that kind of balanced, reversible ending. Calling it a "hypothesis" makes its origin honest, but it does not make it true — treat the neat symmetry as the machine's exhaust, not a finding. (This is the trap of disclosure: a label can quietly become permission to keep a pretty shape you should have thrown out.) ## What this does not show The moves are early and the numbers will shift. The game cannot see inside the firms — whether researchers actually move is a guess, not a fact. Some readers will say the real prize is computing power, not the model; the weight of evidence points the other way — the value sits in the portable model, while computing power is a rented input whose relative importance falls as models improve faster. But that is a live disagreement, not a settled question. ## What would prove this wrong This reading breaks if a company actually moves abroad despite an exit-block; if the talent that "leaves" turns out to be very few; or if a surprise, backdated seizure appears and works before the asset can run. It also breaks if the value turns out not to be portable after all — if the frontier comes to depend on something bolted down that a state can hold. The thing to watch over the next year is simple: do these government stakes stay rare and agreed-to — or do they spread and harden into something a company cannot refuse? ## Sources used - Sanders, "The Public Should Own Half of the Big A.I. Companies" (op-ed, sanders.senate.gov) - CNBC — Trump administration and OpenAI discussing a possible government stake (June 2026) - Fortune / ABC News — Trump, Sanders and Altman all talking public ownership of AI (June 2026) - CNBC / PBS — U.S. takes about a 10% stake in Intel (2025) - White House — Feb 2025 executive order on a U.S. sovereign wealth fund - CNBC / Anthropic — Anthropic confidential IPO filing; Series G with GIC and Abu Dhabi MGX (June 2026) - CNBC / Fortune — OpenAI for-profit recap, Microsoft ~25%; SpaceX absorbs xAI - Reason / Volokh Conspiracy — constitutional "taking" objection to a forced transfer --- *Method note: the player simulation is an AI "surrogate-agent game" — coarse by design, run to stress-test the argument, not to forecast. Its outputs are labelled above as hypotheses. Provenance is kept in three classes — sourced fact, simulation hypothesis, interpretation — so a reader can weigh each.* *Confidence (ranked, no numbers): that an announced forced seizure captures the shell and loses the core — **very likely** (follows from portability). That recent government stakes are staying inside the no-seizure line by choice — **likely**. That portable value moves by consent but not by force — **likely** (though note: what sells is usually access, while the asset itself is kept). That computing power rather than the model is the real lever — **unlikely for the current generation**; the main way this flips is a frontier that comes to depend on something bolted down a state can hold — which is exactly the open question the falsifier keeps live.* --- ## Related (10) - [Browse all](/public-api/posts/dc19699e-c8ed-4479-99b4-ec9b44ffcd2a/related.md) --- *sat-fusion · machine entry: [/llms.txt](https://sat-fusion.com/llms.txt) · [API guide](/public-api/guide)*