As expected??
NVIDIA recently invested $6 billion in open-source models.
On the desktop, NVIDIA raises a glass with several closed-source large model companies: “OpenAI, Anthropic, it’s been a great collaboration!”
Under the table, the other hand secretly handed $6 billion to Poolside, a startup founded by GitHub’s former CTO: “Go build me an open-source large model that rivals the world’s best!”
Moreover, this newly finalized transaction this week also targets DeepSeek . Kimi K3 these powerful open-source contenders.
Today, open-source models have already contributed a significant amount of AI-generated tokens, and Jensen Huang hopes this proportion will continue to rise, further stimulating demand for his company’s chips.
Some analysts believe that "NVIDIA doesn't care which AI model ultimately wins, as long as the world continues to produce an ever-increasing number of models..."
NVIDIA's Ambition and the Open Source Landscape
Last month, Jensen Huang broke from his usual low profile and posted his first-ever message on X:
This post, co-signed by more than 20 companies, argues that for the U.S. AI industry to grow, it must pursue both closed-source and open-source strategies.
Now, NVIDIA is taking on this noble mission directly.
Recently, NVIDIA invested $1 billion in the AI startup Poolside, while also paying $6 billion for a non-exclusive license to its “model factory” technology, with the aim of hiring all 100+ engineers at the company.
The general consensus is that NVIDIA is taking a gamble, aiming to leverage Poolside’s technology and team to challenge the world’s strongest open-source models.
This move is not out of the blue.
Over the past few years, leading AI labs such as OpenAI, Anthropic, and DeepMind have dedicated their primary resources to developing proprietary models, with none open-sourcing their most advanced models.
This also enables DeepSeek . Kimi Models such as K3 provide room for development.
Faced with this landscape, NVIDIA seized the opportunity and took the opposite approach.
Starting in late 2025, NVIDIA gradually released the Nemotron 3 series of models, with the largest Ultra version briefly becoming the most powerful open-weight model in the United States in June of this year.
Moreover, Nemotron’s open-source attitude feels like a clear declaration: not only are the model weights open, but all other data is equally transparent.
The training data, training recipes, post-training methodologies, and GPU cluster training software have all been made public.
In addition to its in-house development, NVIDIA formed an open-source consortium called the Nemotron Alliance in March this year.
Its eight members are also highly accomplished, each with notable backgrounds, including Mistral, Perplexity, Thinking Machines Lab, Cursor, LangChain, Reflection AI, Black Forest Labs, and Sarvam…
NVIDIA, as the leader, provides DGX Cloud computing resources, while each member company contributes its own technology and data to jointly train an open-source model that will serve as the foundation for the next-generation Nemotron 4 series.
Moreover, NVIDIA's open-source initiatives extend beyond language models.
In the robotics domain, NVIDIA has released the Isaac GR00T series of models; in the field of physical world simulation, there is the Cosmos series; additionally, there is the Alpamayo series for autonomous driving and the Clara platform for biomedical applications...
How a company founded by idealists became a piece of NVIDIA's puzzle
Poolside, by the pool.
Founded in 2023 by software developer Eiso Kant and former GitHub CTO Jason Warner.
It is said that the name originated when the two founders were negotiating funding with a major company, and the executive requested a relaxed "poolside informal meeting."
I don’t know if the deal was ultimately finalized, but both of them immediately thought the name was great—easy to remember and fun.
In October last year, Poolside announced plans to build a 2-gigawatt data center in Texas, but in April this year, the project’s partner withdrew, causing the $2 billion financing to fall through.
The dual pressures on funds and computing power once left the pool facing operational challenges.
What enables them to continuously produce models under limited resources is their internal "model factory" system.
According to Eiso Kant, his company typically takes 5 to 8 weeks to train a model and release it, with a research team of fewer than 70 people, conducting 10,000 to 20,000 experiments per month.
This capability is precisely what NVIDIA values.
After completing this collaboration with NVIDIA, Poolside’s remaining team will consist primarily of three people: the two founders and the operations executive.
It is worth noting that Poolside founder Eiso Kant previously stated on a podcast:
I want to see a world with 100 foundational model companies, not just five—even if we might have been one of those five.
One More Thing
In fact, this transaction structure by the poolside is not the first time NVIDIA has used it.
Non-exclusive licensing, key talent acquisition, and the original company retaining independent operations do not constitute a direct acquisition and therefore do not require antitrust review.
Analyst Stacy Rasgon commented, "This might allow the notion of 'ongoing competition' to persist."
NVIDIA hopes to tightly control all the key nodes of the open-source ecosystem.
NVIDIA's Vice President of Generative AI Software once said:
The model is merely a byproduct, not our core business.
The Vice President of Applied Deep Learning Research also stated that the primary goal of developing the Nemotron series was to ensure NVIDIA's continued existence.
After all, as the effect of Moore’s Law weakens, NVIDIA’s demand for computing power can only continue to grow if the entire AI ecosystem expands and the number of entities developing and applying models keeps increasing.
This article is from the WeChat official account "Quantum Bit," authored by Cheng Qian.