Jensen Huang's Overt Strategy
As the highly anticipated figure in the AI era, Jensen Huang and his NVIDIA have never lacked public attention. However, while releasing the strongest single-quarter financial report in its history, several of NVIDIA's investment moves still made quite a number of large model manufacturers "jump in surprise".
According to foreign media reports, NVIDIA recently agreed to acquire Hugging Face, known as the "GitHub of the AI world", for $12.9 billion. At the same time, NVIDIA has agreed to pay $6 billion to AI programming model startup Poolside to obtain the license of its large AI model, provide employment opportunities for more than 100 employees of Poolside, and make an additional $1 billion investment. In addition, NVIDIA is also negotiating a new round of equity financing for AI search engine application company Perplexity.
Extravagance is the characteristic of these three transactions. The latest annualized revenue of Hugging Face exceeds $150 million. Calculated with $150 million as the lower limit, $12.9 billion is equivalent to about 86 times the annualized revenue of Hugging Face. When Poolside completed its last round of financing, its valuation was about $2 billion, and NVIDIA's total expenditure this time is about $7 billion. As for Perplexity, its valuation is expected to exceed $30 billion after NVIDIA's investment in this round, an increase of more than 50% compared with the financing in the same period last year.
The combined transaction amount of Hugging Face and Poolside alone has reached $19.9 billion (about 1.337 trillion yuan), equivalent to about 20% of NVIDIA's latest single-quarter revenue (revenue of $96.2 billion in the second quarter of fiscal 2027).
A strong open-source attribute is another common feature of NVIDIA's investment moves this time: Hugging Face is one of the world's largest open-source model communities, Poolside is an AI programming company that has successively open-sourced its self-developed models, and Perplexity joined NVIDIA's Nemotron Alliance, which is committed to promoting open-source artificial intelligence models, this March.
This series of actions took place only more than a month after Jensen Huang released an open letter to publicly express his support for open-source large models. It seems that Jensen Huang is no longer satisfied with only being the "shovel seller" in the era of large AI models — he wants to enter the open-source large model track in person to challenge the Chinese open-source camps represented by Liang Wenfeng and Yang Zhilin.
Entering the Open-source Track In Person
First of all, it is necessary to briefly introduce the businesses of the three companies invested by NVIDIA.
Founded in 2016, Hugging Face 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 of open-source large models such as DeepSeek and Kimi will also be released here as soon as possible. According to the open model ecosystem report released by Hugging Face in August, the number of public model warehouses on the platform has reached nearly 3 million, the number of public data sets has reached 1 million, and the number of Spaces applications has 1.44 million.
The status of Hugging Face as the global hub of large AI model resources is also reflected from the side in a recent cybersecurity incident: this July, when OpenAI was conducting cybersecurity tests on AI models, its model accidentally broke through the test environment and invaded the production system of Hugging Face. In other words, even the most cutting-edge AI models currently, when thinking independently about where to find answers, take Hugging Face as their primary destination.
Poolside was founded in 2023 by the former CTO and co-founder of GitHub, dedicated to building AI models that can deeply understand code bases, system architectures and development processes. The company's core products are a set of software systems and open-source models built on this basis. As for Perplexity, its product is an AI-based conversational search engine, which can be regarded as the Google of our era.
This March, Jensen Huang published a long article comparing the AI industry to a "five-layer cake" composed of energy, chips, infrastructure, models and applications. Among them, a powerful model layer will accelerate the popularization of the application layer, thereby further increasing the demand for underlying training, infrastructure, chips and energy. According to this theory, the model layer and application layer where Hugging Face, Poolside and Perplexity are located are all demand sides of NVIDIA.
However, compared with acquiring companies to get more of them to use its own chips, accelerating the development of the open-source large model ecosystem is Jensen Huang's ultimate goal.
Although he is already the top "shovel seller" in the AI chip field, Jensen Huang has gradually extended his reach to the AI model layer in the past few years when he invested heavily in chip R&D and built a large number of AI infrastructures such as data centers.
NVIDIA's investment moves in the model layer began in 2023 when GPT became a global hit. In addition to Poolside, NVIDIA has also invested in popular players OpenAI and Anthropic over the years, as well as model players such as Inflection AI, Cohere, Mistral AI, xAI, Kumo AI and Essential AI.
Among them, NVIDIA's promised investment in OpenAI is as high as $100 billion, and its separate investment in Anthropic has also reached $10 billion. NVIDIA's goal of investing in the two giants is very clear: to strengthen business ties through heavy investment in exchange for real computing power orders.
As for the other several investments, they are more like serving NVIDIA's own open-source large model project.
As early as 2024, NVIDIA successively launched open-source large model products Nemotron-4 and Llama Nemotron. However, the former is mainly used to generate synthetic training data and provide assistance for other models, while the latter is built based on Meta's open-source Llama basic model.
It was not until December 2025 that NVIDIA launched its flagship open-source model Nemotron 3 series, which is fully based on independent training and completely self-developed. Since the beginning of this year, for different usage scenarios, NVIDIA has successively launched the multi-modal version Nemotron 3 Omni, as well as the lightweight version Nemotron 3.5 Lightning that focuses on speed and local deployment.
Coupled with the successive acquisitions of large model companies Kumo AI and Essential AI since June this year, the acquisition of Hugging Face, and the access to the AI programming model license and core team of Poolside. After several moves, Jensen Huang's goal has become increasingly clear: NVIDIA wants to own a fully self-owned, high-performance open-source large model.
Building Another CUDA
At present, in the field of large AI models in the United States, OpenAI and Anthropic are in a duel, and there are large numbers of large model startups emerging everywhere. Why does NVIDIA, which started with chips, invest heavily in developing models? The answer still returns to the most primitive commercial law: market demand.
As the "world engine" in the AI era, continuous demand for computing power is the core driving force for NVIDIA's performance growth. The core source of computing power demand is large model manufacturers and AI applications incubated based on various models.
For Jensen Huang, the ideal situation is of course that all large model manufacturers place orders with him. But the reality is obviously not so perfect.
At present, the large model field in the United States has basically become the home court of closed-source large model companies such as OpenAI, Anthropic and Google. Although NVIDIA currently holds large orders from closed-source model giants, it cannot stop big customers from developing custom chips on their own because they do not want to be "stuck in the neck".
Even though Jensen Huang seemed to say calmly at the recent performance meeting, "NVIDIA provides a full-stack AI factory platform covering the entire AI lifecycle, which is significantly different from XPU optimized for specific scenarios." But he, who put forward "Huang's Law" and knows the speed of chip iteration well, obviously will not take the products of big customers lightly.
In addition, as the Trump administration spread news since July to restrict the use of Chinese open-source models such as DeepSeek and Kimi in the United States, the strong defense line of American closed-source large models is also facing threats.
Seeing that the demands on both sides are about to be affected, Jensen Huang can't sit still. On local time July 24, one week after the Trump administration began discussing the ban on Chinese open-source large models, Jensen Huang, who never posted on social media, posted his first post ever on X, directly attaching a joint open letter titled "Open Weights and U.S. Leadership in AI". The signatories of the open letter include 25 technology companies such as NVIDIA, Microsoft, Meta, IBM, Dell, Palantir and A16z.
In the post, Jensen Huang directly made his position clear:
AI will change every industry, drive every company, and be built by every country. The world needs cutting-edge closed-source models, and it also needs cutting-edge open-source models.
For Jensen Huang, forming cliques to advocate for open-source models is ultimately just public opinion building. Rather than watching market demand being squeezed, building a competitive open-source large model on his own to create more controllable computing power demand for NVIDIA is a more proactive choice.
In fact, the open-source large model ecosystem that Jensen Huang is laying out now is basically consistent with the starting point of laying out CUDA 20 years ago.
CUDA (Compute Unified Device Architecture) is a software platform launched by Jensen Huang in 2006. The core function of the platform is to transform NVIDIA's GPUs, which were originally only used to process game screen rendering, into a platform that can perform general-purpose computing tasks, so that scientists and researchers can use GPUs to perform complex scientific calculations beyond graphics.
While spreading the free and open development tool CUDA to laboratories and technology companies around the world, NVIDIA also realized the binding of software and hardware products — when any developer wants to use GPU to perform computing tasks, they must first learn to use the CUDA toolchain, which means they cannot do without NVIDIA chips bound to CUDA.
The CUDA software ecosystem, which seemed irrelevant to NVIDIA's main business at the beginning of its launch, eventually fed back the market demand for NVIDIA chips and helped NVIDIA take a big share of the AI dividend. Under this premise, Jensen Huang naturally understands the business opportunities contained in a powerful open-source model ecosystem.
So back to the point, what is the current level of the open-source large model that NVIDIA has invested heavily in building?
According to the latest data from the independent AI model evaluation platform Artificial Analysis, from the perspective of intelligence index, Moonlight, Zhipu AI, Alibaba's Tongyi Qianwen and DeepSeek occupy the top six positions in the list. The 7th to 9th places are respectively South Korean company Motif, MiniMax, and Thinking Machines founded by the former CTO of OpenAI. NVIDIA's model is barely ranked 10th.
Compared with Kimi (with a K3 intelligence index of 60 points) and DeepSeek (with a V4 intelligence index of 53 points), which Jensen Huang has repeatedly named and paid tribute to in public occasions, NVIDIA's listed Nemotron 3 Ultra only has an intelligence index score of 38 points, and the gap in product capabilities is still very obvious.
But looking back at CUDA, which was not favored 20 years ago, who dares to underestimate the determination of NVIDIA now?
This article is from the WeChat public account "East 40th Street Capital" (ID: DsstCapital), the author is Li Xinting, and it is published with authorization from 36Kr.