Poolside wrote its investors a letter on August 21, and Newcomer got it first. The letter lays out a transaction in three parts. Nvidia pays $6B for a non-exclusive licence to the Model Factory, the system Poolside used to train its Laguna family of open-weight coding models. Nvidia invests a further $1B at a $12B pre-money valuation. And Nvidia extends offers to 109 Poolside employees who worked on Laguna.
The letter states the framing directly: this is “not an acquisition and it is not an acquihire.” The three founders stay. The company keeps operating. The $6B goes out to investors by the end of next year (TNW).
Every word of that is legally accurate. It is also the day 109 people changed employers.
What $6B bought
Poolside CEO Eiso Kant sized the team on the Latent Space podcast last month, in passing. “Less than 70 people built this model,” he said. “Less than 115 between engineering and researchers, like, together did this effort.”
Set 109 against those two numbers and the shape of the deal resolves. The offer list covers essentially the entire technical core. The licence names software. What moved is the group that wrote it.
Divide the fee by the headcount and you get roughly $55M a head. The arithmetic is crude, because the fee nominally buys the Model Factory rather than labor. But The Information’s Amir Efrati made the same observation from the other direction: it is not clear why a licence is worth $6B. Nvidia has not commented publicly.
The third deal of this shape
Nvidia has run this structure twice before.
In December it paid Groq $20B for its inference technology and took its top engineers. Groq stayed independent, raised $650M for what was left, and this month closed $350M at a $3.5B valuation with Nvidia participating. Put those two numbers side by side. The technology and the people priced at $20B. The going concern priced at $3.5B eight months later.
Enfabrica, a hardware startup, went the same way for about $900M.
Three deals, roughly $27B committed, all licence-plus-hiring, none of them a purchase. Every licence non-exclusive, which leaves each seller free to license the same asset again. The structure also skips the regulatory review that attaches to buying a company.
Three times in nine months at that scale is not opportunism. It is a template.
Why the team was available at all
The letter explains it, and the explanation has nothing to do with model quality.
“At the end of last year, we had a 6 week window in which to raise $2 billion dollars to pay for a 40,000 GB300 cluster coming online in January. We didn’t close it in time, and we lost the cluster.”
The letter goes on: 10,000 to 20,000 of those chips would have produced a frontier-competitive model, while next year’s frontier needs a cluster “far more than an order of magnitude larger.” The binding constraint, it says, “is not only capital, it is physical data center space and contracted compute.”
That is a clean causal chain. The team did not fail. The company around the team could not contract compute, and a team without compute is inventory.
Where the residual company points
The people who did not get offers are not unemployed. They work somewhere with a balance sheet about to receive $1B and a stated intention to distribute $6B to investors. The question is what that somewhere does.
Poolside Infrastructure Company spun out in January and is building a 1.2GW data center in Texas. It hired a CEO two months ago and a CFO this week. Both appointments predate Friday’s announcement, which suggests the split was planned rather than improvised. The letter’s forward thesis points somewhere else again: it argues human-level capability will be “fully commoditized by open source models,” that intelligence-bound problems like software and accounting become “a low margin commodity,” and that the value sits in experiment-bound problems where no amount of intelligence substitutes for running the experiment.
Read as a career document rather than a strategy document, that paragraph says the remaining staff at an AI coding company have just been told, by their own founders, that AI coding is heading toward low margin. The founders add that they are “not ready to share the updated vision.”
What this means if you are one of the 109, or one of the people who were not
Three things, in order of urgency.
A new category of employment event exists, and no HR system has a name for it. It is not a layoff, because nobody was terminated. It is not an acquisition, because the legal entity did not change hands and employees did not transfer as a block. It is not an acquihire, because the company and its founders remain. Latent Space called it a reverse-execuhire. The label has no attached severance convention, no notice period, no non-compete norm, and no line in any unemployment statistic. Anyone caught inside one of these is negotiating without precedent to cite.
The list is the entire outcome. In an acquisition everyone transfers. In a layoff the people who stay hold a job with a defined purpose. This structure does neither. 109 people hold Nvidia offers. Everyone else stays at a company whose founders say they are “not ready to share the updated vision.” Same day, same company, two career paths that can no longer see each other.
The list was not drawn by level, tenure, or performance rating. It was drawn by an acquirer’s diligence answering one question: who actually built this. That makes the defensive move concrete. Keep the evidence of your contribution somewhere an outsider can verify it. Commit history, model cards, paper authorship, the author field on design docs, review records with your name attached. Being on the team and being on the list are different things, and only one of them is legible from outside the building.
Compute contracts now sit upstream of employer selection. The old checklist for picking an AI company was mission, model quality, comp, founders. Poolside adds a question that ranks above all of them: how much compute has this company contracted, through what year, and does its financing cadence match the payment windows on the next cluster generation. A team of under 70 people shipped an open-weight model pitched as the West’s answer to DeepSeek and Qwen. What decided their employment was a six-week fundraising window.
On May 2 we covered Meta buying Pinto and Wang’s ARI into Superintelligence Labs on the same day 8,000 people got the layoff memo. The read then was that the market will pay more for one named team than for tens of thousands of generalists. Three and a half months later the price has gone up and the structure has gotten more precise. You no longer have to buy the company. Buying the list is enough.
