Not just selling chips.
As the man in the spotlight in the AI era, Huang Renxun, and his Nvidia, have never lacked attention. However, while announcing the strongest single-quarter financial report in history, Nvidia's several investment moves have still caused many large model manufacturers to be "shocked".
According to foreign media reports, recently, Nvidia agreed to acquire Hugging Face, known as the "AI GitHub", for $12.9 billion. At the same time, Nvidia has agreed to pay $6 billion to AI programming model startup Poolside for the licensing of its large AI model, to provide job opportunities for more than 100 Poolside employees, and to make an additional investment of $1 billion. In addition, Nvidia is also in talks for a new round of equity financing for AI search engine application company Perplexity.
The three deals are characterized by their hefty price tags. Baidu's latest annual revenue exceeded $150 million. Using $150 million as the lower limit, $12.9 billion is equivalent to about 86 times Baidu's annual revenue. Poolside's valuation was around $2 billion when it completed its previous round of financing, and Nvidia's total expenditure this time is around $7 billion. As for Perplexity, its valuation is expected to exceed $30 billion after Nvidia's investment in this round, an increase of over 50% compared to the same period last year.
The combined transaction amount of Baidu's and Poolside's deals alone has reached $199 billion, equivalent to approximately 20% of Nvidia's latest quarterly revenue, which was $96.2 billion in the second quarter of fiscal 2027.
A strong open-source flavor is another commonality of Nvidia's investment moves this time - Baidu is one of the world's largest open-source model communities, Poolside is an AI programming company that has been open-sourcing its self-researched models, and Perplexity joined Nvidia's Nemotron alliance, which is dedicated to promoting open-source AI models, in March this year.
This series of actions comes just over a month after Huang Renxun published an open letter expressing his support for open-source large models. It appears that Huang Renxun is no longer satisfied with just being a "tool provider" in the AI large model era - he wants to personally get involved in developing open-source large models, challenging the Chinese open-source camp represented by Liang Wenfeng and Yang Zhiyun.
Taking the Open-Source Initiative Personally
First, let's briefly expand on the businesses of the three companies invested in by Nvidia.
Baidu-founded Hugging Face was established in 2016 and is now one of the world's largest open-source AI model communities. Developers share and download open-source AI models on Hugging Face, and the latest products from open-source large models such as DeepSeek and Kimi are also released here first. According to Hugging Face's Open Model Ecosystem Report released in August, the platform's public model repository has reached nearly 3 million, public datasets have reached 1 million, and Spaces applications have 1.44 million.
BaoBaoLian's status as a global hub for AI large model resources is also indirectly reflected in a recent cybersecurity incident: in July this year, OpenAI's AI model accidentally broke through the testing environment and invaded BaoBaoLian's production system during a network security test. In other words, even the most cutting-edge AI models, when thinking autonomously about where to find answers, consider BaoBaoLian as their primary destination.
Poolside was founded in 2023 by the former CTO and partners of GitHub, with a focus on building AI models that can deeply understand codebases, system architectures, and development processes. The company's core product is a software system, as well as open-source models built on it. As for Perplexity, its product is an AI-based conversational search engine that can be seen as the Google of our time.
In March, Huang Renxun published an article comparing the AI industry to a "five-layer cake" composed of energy, chips, infrastructure, models, and applications. According to this theory, a powerful model layer accelerates the popularization of the application layer, which in turn increases demand for the underlying training, infrastructure, chips, and energy. Following this theory, the model layer and application layer, where companies like Baidu's DuerOS, Poolside, and Perplexity are located, are all downstream demanders of Nvidia.
However, compared to acquiring companies to use its own chips, accelerating the development of the open-source large model ecosystem is Huang Renxun's ultimate goal.
Although he was the first to make a profit in the AI chip field, over the past few years, while investing heavily in chip research and development and building large data centers as part of AI infrastructure, Huang Renxun has also gradually extended his reach into the AI model layer.
Nvidia's investment moves at the model layer began as early as 2023, the year GPT exploded in popularity. In addition to Poolside, the companies Nvidia has invested in over the years include the popular Open AI and Anthropic, as well as Inflection AI, Cohere, Mistral AI, xAI, Kumo AI, and Essential AI.
Among these, Nvidia has pledged an investment of up to $100 billion to OpenAI and has separately invested $10 billion in Anthropic. Nvidia's goal in investing in the two giants is clear: to strengthen business ties through massive investment and secure substantial computing power orders in return.
As for the other investments, they seem to be serving Nvidia's own open-source large model project.
As early as 2024, Nvidia had successively released open-source large model products Nemotron-4 and Llama Nemotron. The former's primary function is to generate synthetic training data and provide auxiliary support for other companies' models, while the latter is built on Meta's open-source Llama base model.
It wasn't until December 2025 that Nvidia released a flagship-level open-source model, the Nemotron 3 series, which is completely based on autonomous training and independent formation. So far this year, Nvidia has also launched multi-modal version Nemotron 3 Omni for different usage scenarios, as well as the lightweight version Nemotron 3.5 Lightning, which focuses on speed and local deployment.
In addition to the consecutive acquisitions of large model companies Kumo AI and Essential AI since June this year, the acquisition of Bubu Face, and the obtaining of Poolside's AI programming model licensing and core team, several moves have made Huang Renxun's goal increasingly clear: Nvidia wants to have a completely owned, functionally powerful open-source large model of its own.
Recreating a CUDA
Currently, in the US AI large model field, OpenAI and Anthropic are the two dominant players, with numerous startups for large models emerging everywhere. Nvidia, which originated from chips, has invested heavily in models. The answer lies in the most fundamental business principle: market demand.
As the "world engine" of the AI era, sustained computing power demand is the core driver of Nvidia's performance growth, with the core source of this demand coming from large model manufacturers and AI applications incubated based on these models.
For Huang Renxun, the ideal scenario would be for all major model manufacturers to come to him for orders, but reality is unlikely to be that perfect.
Currently, the large model domain in the US is basically dominated by closed-source companies such as Open AI, Anthropic, and Google. Although Nvidia still holds large orders from these closed-source model giants, it cannot stop major clients from starting to develop custom chips in-house due to concerns about being "strangled" by reliance on others.
Even as Jensen Huang appeared composed at the latest earnings call, asserting that "Nvidia offers a full-stack AI factory platform covering the entire AI lifecycle, which is clearly differentiated from XPUs optimized for specific scenarios," the man who coined "Huang's Law" and knows chip iteration cycles intimately is evidently not taking his major customers' products lightly.
Furthermore, with the Trump administration rumored since July to be restricting the use of Chinese open-source models such as DeepSeek and Kimi in the US, the powerful defense line of American closed-source large models is also facing a threat.
As both ends of the demand were about to be affected, Huang Renxun could no longer sit still. On July 24 local time, a week after the Trump administration began discussing a ban on Chinese open-source large models, Huang Renxun, who had never posted on social media, posted his first post on X, directly attaching a joint open letter titled "Open Weights and US Leadership in AI". The signatories to the open letter include Nvidia, Microsoft, Meta, IBM, Dell, Palantir, A16z, and 25 other tech companies.
In the post, Huang Renxun explicitly stated his position:
AI will change every industry, drive every company, and be built by every country. The world needs cutting-edge closed-source models as well as cutting-edge open-source models.
For Huang Renxun, rallying support for open-source models is ultimately just a publicity stunt. Rather than watching market demand be squeezed, it's better to take the initiative by building a powerful open-source large model, thereby creating more controllable computing power demand for Nvidia.
In fact, the open-source large model ecosystem currently being laid out by Huang Renxun has a starting point that is basically consistent with the layout of CUDA 20 years ago.
CUDA (Compute Unified Device Architecture) is a software platform introduced by Huang Renxun in 2006. The core function of the platform is to transform NVIDIA's GPU, originally used only for rendering game graphics, into a platform capable of executing general computing tasks, allowing scientists and researchers to use GPUs for complex scientific calculations beyond graphics.
As Nvidia promoted the free and open CUDA development tools to labs and tech companies worldwide, it also achieved a binding of software and hardware products - any developer wanting to use GPUs for computing tasks had to first learn to use the CUDA toolchain, which in turn cannot be separated from Nvidia's CUDA-bound chips.
When CUDA first launched, it seemed like a frivolous detour from the core business. That software ecosystem ultimately fed demand for Nvidia's chips and helped the company capture the full upside of the AI boom. With that track record, Jensen Huang naturally understands the commercial opportunity embedded in a robust open-source model ecosystem.
So where does Nvidia's heavily funded open-source large model currently stand?
According to the latest data from Artificial Analysis, an independent AI model evaluation platform, in terms of intelligent index, Yue Zhuan Mian, Zhi Pu, Ali Tong Yi Qian Wen, and DeepSeek took the top six spots. Seventh to ninth places were taken by South Korean companies Motif, MiniMax, and Thinking Machines, founded by the former CTO of OpenAI. Nvidia's model ranked tenth.
Compared to Kimi (K3 intelligent index score of 60) and DeepSeek (V4 intelligent index score of 53), which Huang Renxun has repeatedly mentioned and paid tribute to in public, Nvidia's Nemotron 3 Ultra has an intelligent index score of only 38, with a significant gap in product capabilities still evident.
But looking back 20 years, who would have thought that CUDA, which was once underestimated, would become a testament to Nvidia's determination today?