NVIDIA places a $40 billion bet on open source, selling shovels while mining.

As expected??

Nvidia was recently exposed to have bet 6 billion US dollars on open-source models.

On the table, Nvidia toasts with several closed-source large model companies with one hand: "OpenAI, Anthropic, we have a great cooperation!"

Under the card table, the other hand quietly took out 6 billion US dollars and gave it to Poolside, a startup founded by the former CTO of GitHub: "Go and build a world-class open-source large model for me as soon as possible!"

Moreover, this deal, which was just finalized this week, also targets strong open-source players such as DeepSeek and Kimi K3.

Nowadays, open-source models have generated a considerable amount of AI-generated tokens, and Jensen Huang hopes this proportion will continue to rise to continuously stimulate the demand for his own chips.

Some analysts believe that "Nvidia does not care which AI model wins in the end, as long as the world keeps producing more and more models continuously..."

Nvidia's Ambition and Open-Source Layout

Last month, Jensen Huang, who used to be low-key, posted his first post on X in his life:

This post was jointly signed with more than 20 companies, advocating that if the US AI industry wants to achieve development, it must attach equal importance to both closed-source and open-source models.

Now, Nvidia is personally stepping in to fulfill this glorious mission.

Recently, Nvidia invested 1 billion US dollars in AI startup Poolside, and spent another 6 billion US dollars to buy the non-exclusive right to use its "model factory" technology, hoping to bring all more than 100 engineers of the company under its wing.

The general public believes that Nvidia's move is a bet. It wants to rely on the support of Poolside's technology and team to challenge the world's most powerful open-source models.

This move is not unexpected.

In the past few years, leading AI laboratories such as OpenAI, Anthropic, and DeepMind have invested main resources in the research and development of closed-source models, and none of them have open-sourced their most advanced models.

This also provides development space for models such as DeepSeek and Kimi K3.

Facing this landscape, Nvidia seized the opportunity to take a completely opposite approach.

Since the end of 2025, Nvidia has successively released the Nemotron 3 series models, among which the largest Ultra version once became the most powerful open-weight model in the United States in June this year.

Moreover, Nemotron's open-source attitude is completely like a kind of declaration: in addition to the model weights, other data are not hidden.

Training data, training recipes, post-training methodologies, and GPU cluster training software are all made public.

In addition to independent R&D, Nvidia also formed a small open-source group in March this year — the Nemotron Alliance.

Its 8 members are all highly capable with impressive backgrounds, including Mistral, Perplexity, Thinking Machines Lab, Cursor, LangChain, Reflection AI, Black Forest Labs, Sarvam……

As the leader, Nvidia provides DGX Cloud computing resources, and each member company contributes its own technology and data to jointly train an open-source model, which will serve as the foundation of the next-generation Nemotron 4 series.

Furthermore, Nvidia's open-source layout is not limited to language models.

In the field of robotics, Nvidia released the Isaac GR00T series of models; in the field of physical world simulation, there is the Cosmos series; there are also the Alpamayo series for autonomous driving and the Clara platform for the biomedical field……

How a Company Founded by Idealists Became a Piece of Nvidia's Puzzle

Poolside, as its name suggests.

It was founded by software developer Eiso Kant and Jason Warner, former CTO of GitHub, in 2023.

It is said that the name came from a financing negotiation with a large manufacturer, whose senior executive proposed a relaxing "poolside informal meeting".

It is not known whether the negotiation was successfully concluded, but the two founders thought the name was excellent, easy to remember and interesting as soon as they heard it.

Last October, Poolside announced that it would build a 2GW data center in Texas. Unexpectedly, in April this year, the project partner withdrew, and the subsequent 2 billion US dollar financing also fell through.

The dual pressure of capital and computing power once put Poolside in a difficult business situation.

What supported them to continuously produce models with limited resources is their internal "model factory" system.

According to Eiso Kant, his company usually takes 5 to 8 weeks to train a model before its release, the R&D team is no larger than 70 people, and can run 10,000 to 20,000 experiments per month.

This capability is exactly what Nvidia values.

After completing this cooperation with Nvidia, the remaining team of Poolside will mainly consist of three people: two founders and an operations executive.

It is worth mentioning that Eiso Kant, the founder of Poolside, once said in a podcast:

I want to see a world with 100 foundation model companies, not a world with only 5, even though we could have been one of those 5.

One More Thing

In fact, the deal structure for Poolside this time is not the first time Nvidia has used it.

Non-exclusive licensing, recruitment of core talents, and the original company retaining the name of independent operation, none of which constitutes a direct acquisition, so it does not need to undergo antitrust review.

Analyst Stacy Rasgon commented that "doing this may allow the narrative that 'competition still exists' to continue to be maintained."

Nvidia hopes to firmly grasp all the key nodes of the entire open-source ecosystem in its own hands.

Nvidia's Vice President of Generative AI Software once said:

Models are just by-products, not our core business.

The Vice President of Applied Deep Learning Research also said that the primary purpose of developing the Nemotron series is to ensure the continued existence of Nvidia.

After all, as the effect of Moore's Law weakens, only when the entire AI ecosystem continues to expand and the number of entities developing and applying models continues to increase, can Nvidia's demand for computing power continue to grow.

This article is from the WeChat official account "QbitAI", author: Cheng Qian, published with authorization from 36Kr.