This week’s Hacker News thread — “If the models are so good, why doesn’t OpenAI just start 100 companies?” — sits at 538 points and 587 comments, and nearly every one of those comments answers a question nobody asked. The consensus in the replies is that the models are commodities and the moats are distribution, sales, support, integration, trust. All true. All beside the point, because every item on that list is a problem capital and headcount can solve, and OpenAI has both in quantities that would make a mid-sized sovereign state nervous.
Here is the answer the thread skips: a company is not a bundle of tasks. It is a legal person. A thing that signs, borrows, hires, indemnifies, gets audited, gets fined, and — this is the part the thread cannot metabolize — gets sued.
A hundred companies means a hundred defendants
Look at how OpenAI has arranged its own affairs. The 2019 restructuring put a capped-profit arm under a nonprofit board, with investor returns ceilinged at 100-fold. Last October’s reorganization kept the foundation on top. Its consumer terms disclaim warranties in the usual fine print; when enterprise customers wanted real protection, the indemnity arrived through OpenAI’s Copyright Shield, announced at its 2023 developer day, and through Microsoft’s Copilot Copyright Commitment that September. Read those documents together and you learn exactly what the lab thinks about liability: it is something a partner absorbs.
Now imagine a hundred subsidiaries. Each needs an entity, a registered agent, a directors-and-officers policy, an errors-and-omissions policy, a bank that will open an account, and a counterparty willing to sign an MSA with a company whose entire workforce is an inference endpoint. There is no mature insurance market for autonomous business errors at that scale, and Delaware’s Court of Chancery is unmoved by benchmark scores.
A paralegal at a registered-agent shop in Wilmington, reached on a call from the parking garage because her office had lost power, put the whole thing more efficiently than I can: “Nobody calls us because they have a great idea. They call because a client won’t wire money to a name.”
The rent beats the P&L
There is a second reason, and it is the one that makes the first one permanent. OpenAI does not want to be a hundred companies. It wants to be the layer a hundred thousand companies pay on the way through.
The Apps SDK, the agentic checkout work with Stripe, the whole architecture of the last two years points one direction: a toll booth, not a trucking fleet. A basis point on somebody else’s transaction is a claim on revenue that requires no payroll, no shift scheduling, no customer-success department, and no wrongful-termination exposure. Starting a hundred companies would convert that take rate into a hundred operating businesses, each competing with the customers currently paying the bill.
It would also destroy the thing OpenAI’s investors actually bought. A company is a durable claim on future cash flows and gets priced as a multiple of them. A model is a depreciating asset attached to a training bill. Anyone who has run a small business knows the spread between the multiple on recurring revenue and the multiple on a clever tool, and nobody who has run one would trade the first for a hundred of the second.
Deregulation cannot manufacture a defendant
The reflex on my side of the aisle is to blame the compliance burden — the EU AI Act, the state legislatures, the disclosure regimes — for why the labs don’t spin out a hundred ventures and let them run. This gets the causation backwards, and it is worth being blunt about why.
The binding constraint is private ordering. Insurers price risk against a named insured. Procurement departments at large buyers route contracts through legal review that terminates in a signature from an entity with assets. Banks run know-your-customer checks on people and companies, not on models. None of that is a regulation you can repeal. Repeal every AI statute on the books and a Fortune 500 general counsel still will not sign an indemnity running to a shell corporation operated by an agent.
Limited liability itself is a government product. The LLC is a state-conferred privilege, sold for a filing fee, and it comes with a process server attached. You cannot collect the benefits of corporate personhood while declining the part where the sheriff knows where to find you. That is not a bug in the arrangement; it is the consideration.
And the thread’s own premise deserves the same treatment. The interesting question was never why OpenAI won’t start a hundred companies. It is why a hundred companies can now plausibly exist with almost no employees — and who, precisely, is standing behind them when something goes wrong. Five hundred eighty-seven comments, and not one of them asks who gets served.