The Nvidia Poolside deal, as reported by Eric Newcomer's newsletter in August 2026, is a $6 billion non-exclusive license plus roughly $1 billion in equity, not an acquisition. Nvidia did not buy a model. It bought the process that produced one and hired 109 of the 150 people who ran it.
What the Nvidia Poolside deal reportedly covers
The Nvidia Poolside deal, first reported by Eric Newcomer in August 2026, is a $6 billion payment for a non-exclusive license to Poolside's model-development platform, plus about $1 billion in equity, with 109 of the startup's roughly 150 employees moving to Nvidia. It is not an acquisition: Poolside keeps its founders and can license the same technology to others.
Nvidia has not confirmed the terms and Poolside has not either, so the deal rests on one piece of reporting. What can be checked independently is the context around it: Nvidia already runs the largest open-weight model program of any American company, and Poolside, led by former GitHub CTO Jason Warner, published benchmark results in July 2026 showing its smaller model beating Nvidia's flagship. Those two facts frame what the money is actually for.
Three things reportedly changed hands, and one thing did not:
- A non-exclusive license to Poolside's internal training and development platform, which Poolside calls the model factory.
- About $1 billion in fresh equity for Poolside, which stays independent.
- 109 employees, the team that operated that platform.
- No model weights were purchased. Poolside's Laguna models remain Poolside's.
Why Nvidia already gives models away for free
Nvidia went open years before this deal, so the question is not why it publishes open weights. Its Nemotron family ships weights, training recipes, and in most cases training data for anyone to download, with no key and no invoice. In March 2026 the company said Nemotron models had been downloaded from Hugging Face more than 45 million times, making Nvidia the largest single organization publishing there, with close to 4,000 models on the hub.
The range kept growing through 2026. In June, Nvidia released a flagship with 550 billion total parameters, 55 billion active per token, and a one-million-token context window, free for commercial use. In August it added a small model at the other end of the range: 30 billion total parameters with 3 billion active, sized for a single desktop.
So the models exist, the researchers exist, and the compute obviously exists because Nvidia manufactures the compute. That is exactly what makes a $6 billion license confusing until you look at the benchmarks the models lose.
How Poolside's smaller model beat Nvidia's flagship on Terminal Bench
Poolside's published benchmarks are the strongest public evidence that Nvidia's own models lagged. On 21 July 2026, Poolside published results for Laguna S 2.1, a 118-billion-parameter model with 8 billion active, against rivals including Nvidia's flagship on Terminal Bench, an agentic coding test where the model drives a real shell and must finish the job.
Because the table was published by Poolside about itself, read it as a vendor benchmark. Poolside says it took the highest available score for every rival, whether the vendor's own claim, the benchmark author's board, or a third-party leaderboard, and published the full trial trajectories so anyone can inspect what the models did.
| Model | Total / active parameters | Terminal Bench score |
|---|---|---|
| Laguna S 2.1 (Poolside) | 118B / 8B | 70.2 |
| Nvidia flagship (550B class) | 550B / 55B | 56.4 |
| Multilingual bug fixing | Laguna 78.5 | Nvidia 67.7 |
| Tool use test | Laguna 49.7 | Nvidia 34.3 |
The pattern repeats across tests: Laguna has about one-fifth of the total parameters and one-seventh of the active ones, yet leads by 13.8 points on Terminal Bench and by double digits elsewhere. That gap, on a benchmark the larger model should dominate, is the number that explains the check.
The model factory: what $6 billion actually bought
The $6 billion bought a process, not weights. Laguna S 2.1 started pre-training on 22 May 2026 on 4,096 Nvidia chips and shipped publicly in under nine weeks, about 60 days from the first training token to public weights. In the twelve weeks before that, the same small team had already shipped two other models.
Poolside calls its internal platform the model factory and describes it as the system that lets the company, in its own words, industrialize the model development process, maximize research iteration speed, and minimize the attention researchers pay to bookkeeping and infrastructure. Four thousand chips is a rounding error next to the clusters Nvidia's largest customers rent. The advantage was organizational speed, and the license plus 109 staff is Nvidia buying that advantage rather than rebuilding it.
The uncomfortable question this raises applies to everyone buying compute: if Nvidia, the company that manufactures the compute, could not close the gap with chips alone, then silicon was never the bottleneck in this part of the race. That is an inference from the reported terms and the published benchmarks, not something either company has stated.
Why Nvidia cares: open weights decide whose hardware wins
Nvidia does not sell models; it sells the machine underneath every token anybody generates anywhere. Jensen Huang put the commercial logic plainly in July 2026: free AI should be great for hardware and great for chips. Making the thing adjacent to your product free is a distribution strategy, and Nvidia is the one company in the market whose business model makes giving away the best model a dominant move.
The urgency comes from usage data. OpenRouter, a marketplace that routes developer traffic to whichever model a developer picks, is one of the cleanest public reads on what people actually run. In June 2025, American labs combined took roughly 70 percent of tokens flowing through it; a year later that share sits near 30 percent, with Chinese open-weight models taking most of the rest depending on where the boundary is drawn. Hugging Face reports about 41 percent of downloads over the trailing year went to Chinese models, and Alibaba's Qwen family alone has passed a billion cumulative downloads.
The sharper risk is hardware lineage. Weights are just numbers in a file; the surrounding stack of kernels, quantization formats, serving code, and fine-tuning recipes gets written for the hardware the model was born on. In February 2026, a Chinese lab shipped a 745-billion-parameter frontier open model trained end to end on Huawei Ascend chips with no Nvidia silicon anywhere in the pipeline. If the default open model on earth is tuned for Huawei, every enterprise that adopts it inherits a stack that does not need CUDA. The Nvidia Poolside deal is a bet that owning the default is worth about eight days of revenue.
The cost and the deal structure regulators will notice
Money is the least of it. The $6 billion license is roughly eight and a half days of Nvidia's revenue, about ten days including the equity, and around a third of what Nvidia spent on research and development across its last fiscal year. For a company of Nvidia's scale, the price is not the issue. The structure is.
Nvidia is not buying Poolside. Poolside stays independent, keeps its founders, keeps the right to license the same technology to anyone, and walks away with about $1 billion in fresh equity. Nvidia takes the process and 109 of the 150 people who ran it. This mirrors two earlier deals in Nvidia's recent history: in September 2025 it paid over $900 million for a networking startup's technology and staff, and three months later it licensed technology from an inference-chip rival for a reported $20 billion while hiring the team. Three licensing agreements, unlike one merger, avoid a merger review, and one analysis of the reported terms called the price for a non-exclusive license deranged before making the sharper point that regulators may look at substance rather than labels.
There is also a customer-relations cost. OpenAI and Anthropic buy Nvidia's chips at enormous scale, and Nvidia is now openly building models pitched as the cheap, customizable alternative to theirs. A supplier can climb into its buyer's business only so many times before those buyers start funding alternative silicon. That consequence is speculative, but the tension is structural and visible in Nvidia's own product strategy.
The Anthropic correction: no open-weights ban lobby
A circulating version of this story says Nvidia goes open while Anthropic lobbies to ban open weights. That version is wrong. On 24 July 2026, Jensen Huang made his first post on X, an open letter warning Washington against premature restrictions on open-weight models. It launched with 25 signatories and passed 150 within days. Anthropic did not sign it.
Three days later, Dario Amodei published Anthropic position, and its opening claim is that Anthropic never advocated a ban on open-weight models. What the company wants is chip export controls, a crackdown on industrial-scale distillation, and mandatory safety testing for capable open and closed models alike. Amodei also called open models without dangerous capabilities a public good.
You can disagree with Anthropic position, and plenty of serious people do. But declining to sign a letter is not lobbying for a ban, and arguments should be aimed at positions people actually hold.
Frequently asked questions
- Did Nvidia buy Poolside?
No. According to reporting by Eric Newcomer in August 2026, Nvidia paid $6 billion for a non-exclusive license plus about $1 billion in equity and hired 109 employees. Poolside remains independent and can license the same technology to anyone else. Neither company has confirmed the terms.
- What is the model factory Nvidia licensed?
It is Poolside's internal platform for industrializing model development, built to maximize research iteration speed and minimize researcher bookkeeping. Using it, Poolside shipped a 118-billion-parameter model from first training token to public weights in under nine weeks on 4,096 Nvidia chips.
- Did Poolside's model really beat Nvidia's flagship?
On the benchmarks Poolside published on 21 July 2026, yes: Laguna S 2.1 scored 70.2 on Terminal Bench against 56.4 for Nvidia's 550-billion-parameter flagship. The table was published by Poolside, though it says it used the highest available scores for rivals and released full trial trajectories.
- Why does Nvidia give open models away?
Nvidia sells the hardware underneath every token generated anywhere, so free models are a distribution strategy, not charity. Huang said in July 2026 that free AI should be great for hardware and chips. The reported deal protects that strategy as Chinese open-weight models trained on Huawei Ascend hardware gain adoption.
- Is the Nvidia Poolside deal a merger?
Structurally, no, and that distinction matters for regulators. A merger triggers a merger review; a licensing agreement with a hiring component does not. This is the third such technology-and-team deal attributed to Nvidia in under a year, which is why some analysts argue regulators should examine substance rather than labels.
Turn a video like this into your own written analysis
This story only holds together because someone connected a leaked letter, a benchmark table, and a year of download statistics into one argument. If you have the same kind of material sitting in a YouTube video, an explanation, an interview, a commentary you already recorded, the reasoning is trapped in a format nobody can search or quote.
Skala blog turns a YouTube video into a structured written article: paste the URL, the video is transcribed, and the transcript becomes a publishable draft you can review and edit. If you cover technical topics the way this piece does, whether it is deals like the Nvidia Poolside agreement, tooling write-ups in the spirit of Dev doido, or hands-on guides like the Crazystack typescript tutorials at crazystack.com.br, your spoken analysis deserves a written form that search engines and answer engines can actually find.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits