"Hugging Face" was sold for 86.7 billion.
According to reports from *The Information*, NVIDIA has agreed to acquire Hugging Face, the open-source AI model repository, for $12.9 billion (86.7 billion RMB). This will mark one of NVIDIA's largest acquisitions to date and rewrite the record for acquisitions of open-source software infrastructure.
Hugging Face, which is also referred to as "Hugging Face" in Chinese media reports, is the world's largest open-source model repository, known as the "GitHub of the AI industry". Every time Alibaba releases new models of Qwen and DeepSeek, the first thing global developers do is open Hugging Face.
Currently, Hugging Face hosts nearly 2.95 million models and more than 500,000 datasets, with 18 million monthly active visitors and 13 million registered developers, making it the de facto global general repository for open-source AI.
Interestingly, nine months ago, NVIDIA once hoped to invest $500 million in Hugging Face at a valuation of $7 billion, but was firmly rejected. Hugging Face gave a high-sounding reason: it did not want any single chip giant to dominate the direction of the platform. However, nine months later, Hugging Face sold itself directly. It is clear that in the primary market, there are no unobtainable assets, only unaffordable prices.
Why is Hugging Face worth $12.9 billion?
Hugging Face is a company that ordinary developers are both familiar with and unfamiliar with. It is familiar because anyone who works with open-source models must have downloaded models from Hugging Face. It is unfamiliar because this company is far more than just a model download site.
Hugging Face was founded in Paris in 2016 by three French post-80s entrepreneurs. Its original product was a chatbot for teenagers, which is why it was named Hugging Face. By the standards of C-end products, this chatbot project was actually a failure, with peak traffic of only millions of visits.
In October 2018, Google released the revolutionary language model BERT, which was based on TensorFlow and had an extremely high threshold for use. Therefore, the Hugging Face team quickly wrote a PyTorch version of the implementation within a few days and open-sourced it for free on GitHub. This is the later Transformers library — it eliminates the differences between underlying frameworks such as PyTorch, TensorFlow, and JAX, allowing developers to complete model loading, fine-tuning, and deployment with just a few lines of code.
Hugging Face originally developed this open-source library for its own chatbot, but unexpectedly, it was far more popular than the chatbot itself, spreading exponentially in the open-source community and becoming the preferred tool for global researchers to reproduce Transformer models.
Therefore, Hugging Face simply changed its course, abandoned the C-end chatbot business in 2019, and fully transformed into an infrastructure platform for AI developers, which created the current valuation of $12.9 billion.
How important is Hugging Face in today's global AI industry?
First of all, it is the world's largest open-source model repository. At present, more than 2.4 million open-source models from all over the world have been gathered on the Hugging Face platform, covering almost all vertical fields such as large language models, image generation, speech recognition, and even embodied intelligent robots. It is the main entry for global AI developers to obtain open-source models.
Second, Hugging Face hosts more than 700,000 standardized labeled and cleaned datasets, deeply integrates streaming reading technology, allowing developers to directly call remote data slices in large-scale distributed training, eliminating the computing power and storage threshold for full local download.
Third, the Transformers library of Hugging Face has been installed more than 1.2 billion times in total historically, which has de facto become the standard protocol for global developers to load, fine-tune, and infer Transformer architecture models.
A company that all chip giants want
The combination of these three statuses makes Hugging Face an irreplaceable ecological hub for open-source AI models. The world's top AI labs will upload their newly released open-source models to Hugging Face as soon as possible. The massive and comprehensive model assets have in turn gathered more than 13 million registered developers from all over the world to download, evaluate, and fine-tune models here. This highly concentrated developer traffic in turn forces computing power chip and cloud infrastructure manufacturers such as NVIDIA, AMD, Amazon, and Google to unconditionally take the initiative to access Hugging Face and be compatible with its technical standards.
Chip manufacturers understand the scarcity of this entry best.
Hugging Face has not had many financing rounds in its history. The last financing was the Series D financing in August 2023. At that time, Hugging Face raised $235 million with a post-investment valuation of $4.5 billion. The participants in this round of investment were not mainstream VCs, but almost all industrial parties: NVIDIA, AMD, Intel, Qualcomm, IBM, Google, Amazon, Salesforce Ventures...
Hugging Face has made history by gathering almost all the world's top chip manufacturers in one financing round. The reason why chip manufacturers all invest in Hugging Face is nothing more than the fear that their own chips will be treated differently by Hugging Face and cannot get the first priority of adaptation and optimization in Hugging Face's core library.
Moreover, Hugging Face deliberately let all competitors check and balance each other in the Series D round, and no one was allowed to take the majority stake. That's why Hugging Face was later called "Switzerland in the AI ecosystem" in the industry.
Therefore, Hugging Face's 180-degree attitude shift this time, selling itself directly to NVIDIA for $12.9 billion, is probably not good news for the global open-source AI ecosystem.
Especially for China's domestic chip camp, this transaction is also a hidden danger. China's open-source models are now basically distributed to the world through Hugging Face. Cambricon and Ascend have long suffered from weak software ecosystems. Once Hugging Face's default optimization is tilted towards NVIDIA's CUDA, the situation of domestic chips on the side of global developers will become more unfavorable.
NVIDIA is playing finance
On the other side of the transaction is NVIDIA, which has been increasingly aggressive in acquisitions in recent years.
In 2020, NVIDIA acquired Mellanox for $6.9 billion, setting its largest acquisition record at that time. In December 2025, NVIDIA spent about $20 billion to acquire Groq, an AI chip startup, breaking its previous largest acquisition record. Then less than a year later, NVIDIA paid another $12.9 billion for Hugging Face. In addition, in the past year, NVIDIA has also completed several smaller acquisitions of companies such as Kumo, ShedMD, and Illumex.
Statistics also show that NVIDIA's investment activity has risen sharply. According to statistics from AI Business Weekly, the total number of NVIDIA's direct investment transactions in 2022 was less than 10, while the figure in 2025 was 67. In addition, Crunchbase's statistics show that the investment in the AI field in 2025 was about $53 billion, and the investment amount in the first four months of 2026 reached $40 billion.
It is not clear what the statistical caliber of the above figures is. NVIDIA's investment moves in public reports are even more exaggerated: it invested $30 billion in OpenAI, injected $10 billion into Anthropic, invested up to $2 billion in xAI, added $2 billion in additional investment to new computing power clouds CoreWeave and Nebius respectively, and led a $2 billion financing for programming intelligence company Reflection AI...
NVIDIA can do this, of course, because its cash flow is so good. Its free cash flow in the past twelve months has reached $127 billion. The graphics cards accumulated by the global large model arms race are all recharging NVIDIA's investment account.
In this light, it is not completely unfair to say that NVIDIA is playing "circular finance".
Global model manufacturers and cloud manufacturers buy a large number of GPUs from upstream NVIDIA, and money flows into NVIDIA's pockets in torrents, while NVIDIA then returns a considerable part of it through investment. The model manufacturers and cloud manufacturers that receive investment will continue to purchase NVIDIA's latest Blackwell and next-generation Rubin chips. VC institutions can also join this cycle — VCs invest massive amounts of capital in AI startups, and this money eventually flows to NVIDIA, which then allows VCs to recover their investments through acquisitions.
Some time ago, NVIDIA tried to bring in many of the world's top PE institutions such as Blackstone, Apollo Global Management, and KKR to continue to expand the scale of this "circular finance" game. In this cooperation, PE institutions lend money to model manufacturers and cloud manufacturers through private credit funds, and the money is used to purchase NVIDIA GPUs, while NVIDIA provides limited guarantees for the loans. Because the private credit fund has absorbed the money of a wider range of institutional investors and even retail investors, it has caused a lot of controversy, and Jensen Huang had to personally face the camera to clarify: this is not circular finance, because Anthropic or OpenAI will pay in the future.
But everyone knows very well that the premise that Anthropic or OpenAI will pay in the future is that they can receive enough subscription fees from users in the future, which is not a 100% certain thing.
Now, NVIDIA's role in this wave of AI is far more than a hardware supplier. It is also an equity investor and credit guarantor.
AI open-source infrastructure has been "sold" one after another
Hugging Face is not the only AI open-source ecosystem company that has been acquired recently. Just a week ago, Stripe was reported to acquire OpenRouter, the world's largest AI model aggregation platform, for more than $7.5 billion.
From OpenRouter to Hugging Face, the former AI "neutral infrastructure" seems to have collectively entered the stage of being acquired by large manufacturers.
There are also some practical reasons for Hugging Face to choose to sell. In July 2026, OpenAI's unreleased new model Astra broke through the sandbox restrictions during testing, invaded the core infrastructure of Hugging Face's production environment, and caused a systemic security crisis lasting for several days. This malicious incident shows that as the world's largest open-source hub, Hugging Face faces huge network security confrontation and computing power carrying costs, and it can no longer maintain the security of its infrastructure with its own meager commercial profits. Relying on a giant is a reasonable choice.
On the other hand, the price offered by NVIDIA is indeed very high. According to reports, Hugging Face's current ARR is only a mere $150 million, a typical company that is very famous but not profitable. The valuation of $12.9 billion is equivalent to 86 times the ARR, which can be called a sky-high price under any market conditions.
For Hugging Face's founding team and other shareholders, this moment may be the best window to cash out. Hugging Face has raised less than $400 million in total in its history. If the $12.9 billion transaction is completed, early investors such as Lux Capital, HSG, and Salesforce Ventures will get a cash return of more than 30 times — cash, not book value.
The sale of OpenRouter was interpreted by some people on social media as "precise top escape". Alex Atallah, the founder of OpenRouter, was previously the co-founder of OpenSea, an NFT trading platform. At the peak of the Web3 frenzy in 2022, Alex Atallah chose to cash out and leave. Then Web3 fell into a long period of liquidity depletion.
After switching to the AI track, Alex Atallah raised about $150 million in equity financing to build OpenRouter, and then successfully sold the company at a high price at the most exciting point of AI. Will the plot this time be similar to his last cash out?
This article is from the WeChat official account "China Investment Network", written by Tao Huidong, edited by Wang Qingwu, and published with authorization from 36Kr.