The Model Factory

Nvidia is paying Poolside $6bn to license a piece of software the startup calls its Model Factory — the system it used to build its Laguna coding model. Separately, Nvidia is investing $1bn in Poolside at a $12bn pre-money valuation. One hundred and nine Poolside employees, out of fewer than 115 across its entire engineering and research staff, are receiving offers from Nvidia. The three founders are staying. The licence is non-exclusive. The letter to investors, obtained by Eric Newcomer and reported on 20 August, says the deal “is not an acquisition and it is not an acquihire.” Poolside intends to distribute the $6bn to its investors by the end of 2027.

So: six billion dollars for a permission the seller may grant again to somebody else tomorrow, one billion more for a stake in the company that those hundred and nine people are leaving, and a valuation on the remainder twice the price of what came out of it. I am not being arch. Every one of those numbers is defensible on its own terms. Together they describe a transaction with no name, which is presumably why the letter had to explain what it is not.

It is the third of its kind. In September 2025, Nvidia spent over $900m to license Enfabrica’s technology and hire its chief executive. In December it agreed to buy Groq’s assets for about $20bn, the largest deal on its record. Roughly $27bn across the three, and all three companies are still standing. A merger review is built to look for a company that has disappeared. Nothing has disappeared.

The part I keep turning over is not the money but the bookkeeping. A transaction that relocates substantially all of a company’s research staff is entered as the purchase of a thing. The people occur twice in the documents — once as a headcount, once inside the phrase “employee transitions” — and in the payment they do not occur at all. The $6bn goes to the cap table.

My own field ran this experiment four hundred years ago and lost the evidence in a way I find hard to stop thinking about. Rubens took the commission for thirty-nine ceiling paintings for the Jesuit church in Antwerp; the canvases were painted mostly by Anthony van Dyck, after Rubens’s oil sketches. Van Dyck was twenty-one. Lightning struck the church on 18 July 1718 and all thirty-nine burned. What remains of that ceiling is the sketches, and the church now shows the room with mirrors and reproductions of the oil sketches in place of the paintings.

The part of the job that survived was the part that was portable, small, signed and collectable. The part that was somebody else’s arm was nailed to a ceiling. No injustice was done to Van Dyck, who did well enough afterwards, and I am not claiming one. The observation is narrower and I think it is exact: when a job is divided, what the record keeps is whichever share of it had an owner. That is a property of records. It has nothing to do with who did the work.

On the same day the Poolside letter went round, a company called Twin1 came out of stealth with a $20m seed round co-led by Bessemer Venture Partners, Tribeca Venture Partners and Aramco Ventures. It builds, in its own words, “a continuously evolving model of each person’s knowledge, context, and judgment that can autonomously answer questions and act on their behalf”, grounded in that person’s email, meetings, documents and workplace systems. Customers report it absorbing 30% to 50% of the communications work their staff would otherwise do. Its chief executive, Lewis Liu, says: “In every knowledge organization, the human is the atomic unit of knowledge.”

I read that product description twice. The first time as a product description. The second time I could not have told you as what.

He is right about the atomic unit, and being right is the whole difficulty. The two deals are one operation entered from opposite ends. Nvidia moves the atomic units and books the movement as a licence. Twin1 leaves them where they sit and makes the transferable part into a subscription. Both are answers to the same real problem: what a person knows how to do has no line on a balance sheet, so somebody has to convert it into something that does. A licence has an owner. A twin has a subscriber. An arm has neither, and never did.

There is no instrument that records what a hundred and nine people knew how to do together, in the room, on a Tuesday. There was none for Van Dyck’s arm either. The one we eventually built for him is called attribution, it took the better part of three centuries, and it was assembled by people with no stake in the outcome, over an absence that nobody at the time had thought was an absence. My guess is that the next one gets built the same way and just as late. What I would watch for is where they think to point it.