SANTA CLARA, 22 AUG 2026 — Nvidia has agreed to pay Poolside US$6bn for a non-exclusive licence to the system the startup uses to build its open-weight Laguna coding models, hire 109 of its employees, and invest a further US$1bn at a valuation of US$12bn.
Poolside is not being acquired. It keeps operating, its three co-founders stay to run it, and the licence is explicitly non-exclusive. Nvidia ends up with the technology, most of the team that built it and an equity stake, and no merger notification attaches to any part of it.
What the deal contains
The licensed asset is described as Poolside's model factory: the pipeline the company uses to produce the Laguna family of open-weight coding models, rather than the models themselves. Remaining with Poolside are the three co-founders, among them former GitHub chief technology officer Jason Warner and Eiso Kant.
The US$6bn is expected to reach existing backers — Bain Capital Ventures, eBay, Citi Ventures, Redpoint and Adams Street among them — by the end of 2027. That detail — that the cash goes to existing backers — is central to the deal's purpose.
The structure is the point
Assembled from parts, this delivers what an acquisition delivers. Nvidia obtains the capability to build coding models, the engineers who know how, and a financial interest in the entity that retains the rest.
What it avoids is a change of control, the legal trigger for merger review. Competition authorities examine transactions where one firm acquires another. A licence is a commercial contract. A hiring round is a hiring round. An investment that leaves the target independent is a minority stake. Each component, taken alone, sits below the threshold that triggers scrutiny.
This is not a novel manoeuvre. Microsoft and Inflection, Google and Character.AI, Amazon and Adept all followed a version of it during 2024, and each drew regulatory attention afterwards without being unwound. The scale is what is new. Seven billion dollars is not a marginal workaround; it is a large acquisition in all but name, executed with instruments that fall below the reporting threshold.
We noted yesterday, on Nvidia's approach to the Korean inference designer Rebellions, that the commercial case and the approval path are separate problems and the second is harder. This is what solving the second problem looks like.
Non-exclusive is doing real work in that sentence
The non-exclusive nature of the licence is a key legal distinction, not a technicality.
An exclusive licence transfers effective control of an asset: the licensor cannot use it elsewhere, which is economically close to a sale. A non-exclusive licence leaves Poolside free to keep using its own model factory and to license it again. On paper the company retains everything it had.
In practice the qualifier is doing legal work more than commercial work. Poolside retains the right to use a system that 109 of the people who built it no longer work on, having taken a payment large enough to return capital to its investors. The asset stays; the capacity to develop it further is substantially reduced.
Whether this is a real limitation depends on how much of the model factory is embedded in code versus the knowledge of the departing engineers. That is a question for a regulator, and one nobody outside the companies can answer.
Follow the money to the cap table
The US$6bn payment to existing backers by 2027 reveals the problem this deal solves for Poolside.
A venture-funded company that has raised heavily needs an eventual exit, and the routes are acquisition or flotation. This provides the cash return of an exit without either. Investors are made whole or better, the founders keep the company, and Nvidia does not have to argue with a competition authority.
Read that way, the deal is a liquidity event wearing a commercial agreement's clothing. It is a clever answer to the problem of providing an exit for investors. But it is also why the structure will attract scrutiny: the money moves as it would in a sale, even as the legal form insists nothing was sold.
Why an infrastructure company wants a model factory
Nvidia sells accelerators. So why is it paying seven billion dollars to move up the stack into coding models?
Coding is the most commercially proven application of large language models. It has measurable output, enterprise willingness to pay and workloads that run continuously rather than in bursts. It is also the application most likely to drive sustained inference demand, which is what fills data centres after the training run ends.
Owning the means of producing coding models does not require Nvidia to compete with its own customers directly. Open weights let it seed an ecosystem that consumes its hardware without displacing the model companies buying accelerators. The model factory is more useful to Nvidia than any single model would be, because it produces the thing repeatedly.
What regional buyers should take from this
For enterprises across ASEAN evaluating open-weight coding models, the practical consequence is roadmap control.
Laguna is open-weight, and models already released stay released. What changes is the direction of what comes next, now that the people building it work for the company selling the hardware it runs on. An open-weight model steered by an accelerator vendor is unlikely to be optimised for hardware from anyone else.
That is not a reason to avoid these models. It is a reason to treat open weights as a licence rather than as independence, and to keep the inference layer portable — which matters more in markets buying compute at the margin than in ones buying it by the gigawatt.
What remains unconfirmed
Neither company has published the agreement, and the figures come from reporting rather than filings. The precise split between the US$6bn licence and the US$1bn investment in different accounts is not established, nor is the payment schedule beyond the end-2027 expectation, nor what the 109 transferring employees will work on.
It is not stated whether any competition authority has been notified or has opened an inquiry, whether the licence carries exclusivity in any field of use despite being described as non-exclusive, what governance rights the equity stake carries, or how many people remain at Poolside.
What to watch for
The first signal is whether any regulator opens a review anyway. Authorities in the United States, United Kingdom and European Union have all examined licence-and-hire arrangements before without blocking one, and a seven-billion-dollar example is a more attractive test case than the 2024 precedents were.
The second is what Poolside ships next. A company that has licensed out its model factory and lost 109 engineers either demonstrates the asset was reproducible or quietly stops producing.
The third is whether this becomes the standard shape. If the next three significant AI transactions are also licences plus hiring rounds, merger review will have been routed around as a matter of practice rather than in a single case.
