# Nvidia Commits About $7 Billion to Open-Weight Models Through a Poolside Licensing Deal

The chipmaker will pay $6 billion to license Poolside's training stack, invest $1 billion at a $12 billion pre-money valuation, and move more than 100 of the startup's staff onto its Nemotron open-weight team.

- Published: 2026-08-24T07:20:23.472Z
- Canonical: https://polylog.news/ai/2026-08-24/nvidia-commits-about-7-billion-to-open-weight-models-through
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [The Next Web](https://thenextweb.com/news/nvidia-poolside-6bn-model-factory-licence), [Seeking Alpha](https://seekingalpha.com/news/4636058-nvidia-compete-us-chinese-ai-models-under-6b-poolside-deal), [MarkTechPost](https://www.marktechpost.com/2026/08/11/nvidia-ai-releases-nemotron-3-5-lightning-and-nemo-switchyard/)

Nvidia has agreed to pay Poolside $6 billion for a non-exclusive licence to the startup's model-training technology, and to invest a further $1 billion at a $12 billion pre-money valuation. The deal was reported by The Wall Street Journal, summarized in [Russian-language technical channels](https://t.me/ai_machinelearning_big_data/10749) and confirmed in [English coverage](https://thenextweb.com/news/nvidia-poolside-6bn-model-factory-licence). About 109 Poolside engineers will move to Nvidia and join Nemotron, the company's family of downloadable open-weight models. Poolside's management team stays independent and continues its own research.

The stated purpose is competitive rather than scientific. Nvidia executives have framed the spending as an effort to build a United States open-weight stack that can compete with Chinese releases such as DeepSeek and Kimi. The aim, according to [Seeking Alpha](https://seekingalpha.com/news/4636058-nvidia-compete-us-chinese-ai-models-under-6b-poolside-deal), is also to give enterprises a cheaper and modifiable alternative to the closed application programming interfaces (APIs) sold by OpenAI and Anthropic.

The starting position is modest. Nvidia's most recent open release, [Nemotron 3.5 Lightning](https://www.marktechpost.com/2026/08/11/nvidia-ai-releases-nemotron-3-5-lightning-and-nemo-switchyard/), is a 30-billion-parameter mixture-of-experts model with about 3 billion active parameters. It ships with post-training data and recipes and targets the high-volume mechanical layer of agent work: tool calls, retries, formatting and validation. Nvidia reports about 86 percent accuracy on PinchBench, an agentic benchmark of its own construction, and roughly 30 percent faster completion of 10,000 tasks than a comparable Qwen model. Those are vendor numbers on a vendor benchmark. Independent comparisons over the past month have generally placed Kimi K3 and DeepSeek V4 Pro ahead of the American open-weight entries on task coverage.

There is a clear conflict of interest here. Nvidia sells the accelerators that any serious open-weight training run consumes, and a larger supply of downloadable models expands on-premises and sovereign inference demand, the segment where Nvidia captures the most value per unit sold. Critics have argued the company is effectively underwriting demand for its own hardware. Both can be true at once: the deal buys Nvidia a training organization it did not have, and it also buys future GPU consumption.

## What this means

Nvidia is converting balance-sheet capacity into a model-training organization. The channel it gains is distribution, not new technical capability. If Nemotron becomes the default American open-weight base, the buyers most exposed are closed-API vendors selling mid-tier reasoning and coding capacity, because an open model that is merely adequate at one tenth the marginal cost eliminates the mid-priced tier those vendors depend on. Poolside investors get a liquidity event at a $12 billion valuation without a sale, which sets a reference price for other pre-revenue model labs raising money this quarter. The open question is whether the acquired team ships a model that independent evaluators rank against Kimi K3 and DeepSeek V4 Pro within a year, or whether Nemotron remains a competent agent-execution family that never reaches frontier-level performance.

## What to watch

- Whether the first Nemotron release built with the Poolside team is evaluated by third parties such as Artificial Analysis rather than only on Nvidia's own PinchBench, which would show whether the capability claim holds up outside the vendor's own test setup.
- Whether other chip and cloud vendors follow with licensing deals for model labs instead of acquisitions, which would signal that talent-and-licence structures are now the standard way to buy an AI team without triggering merger review.
- Download and deployment share of American versus Chinese open-weight models on public hubs over the next two quarters, the clearest measure of whether the $7 billion deal actually shifts which model becomes the default choice.