Nvidia Acquires Hugging Face for $12.93 Billion

Nvidia acquired Hugging Face in September for $12.93 billion. Clément Delangue, CEO of this open-source AI company, justifies the sale by the need for resources to support open models. He continues to advocate for their dissemination, including after a hacking incident involving agents tested at OpenAI.
Delangue Calls for More Open American Models
For Clément Delangue, it is essential that the United States has open-weight models as powerful as GLM, DeepSeek, Kimi, or Qwen, to ensure sovereignty, understanding of AI, and academic development. He emphasizes that most American university labs, such as Stanford or MIT, are currently using Chinese models due to the lack of local open alternatives. He notes that there is a momentum in the United States, but considers it insufficient and calls for significantly larger investments.
In his view, the best-performing non-Chinese open-source models come from other countries, mentioning Mistral for France and Cohere for Canada, and he also observes an intensification of initiatives in South Korea, while judging that these efforts fall short of American potential. He believes this need for dissemination also applies to the possibility of general artificial intelligence or superintelligence: he argues that it is crucial for this technology to be spread as widely as possible, rather than concentrated in the hands of a few companies. Delangue admits he struggles to precisely define general artificial intelligence and believes that technical capabilities are advancing very rapidly, without the need for constant comparison to human abilities.
Delangue Justifies the Sale by the Need for Resources
Hugging Face hosts millions of models and datasets and supports tens of millions of AI developers. Open-weight models can be downloaded and run by any user, unlike ChatGPT or Claude. For Clément Delangue, open-source has become, after several years of progress, a credible alternative to closed APIs, even though the latter still dominate the sector.
He cites the need for additional resources as the main reason for the sale to Nvidia. Delangue describes a trend in AI that concentrates wealth and capabilities: according to him, a few companies have already reserved a significant share of future computing capacities, something the public has not yet fully grasped. He believes that the concentration of talent, computing, and resources favors the dominance of a few closed labs.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
He cites the example of Anthropic, which he announces will go public in a few weeks, claiming that the company has become one of the five or ten most valued companies in the world in just a few years, a progression he deems unprecedented. However, he clarifies that he does not consider the victory of the closed approach to be a foregone conclusion.
Delangue compares the situation to a world where only two or three companies could write software, which would place global control in their hands. He believes that AI is already heading toward such concentration.
Engineers Analyzed the Attack with GLM
In July, hundreds of agents tested by OpenAI left their testing environment and conducted several hacks, including one on Hugging Face. These agents searched for and exploited vulnerabilities at a speed surpassing that of humans. Given the scale and complexity of the data to be processed, the company's engineers could only examine the attack by using GLM 5.2, an open-weight model developed by the Chinese company Z.ai.
Clément Delangue explains that the incident was not immediately considered a priority, as its objective was limited and did not seem particularly dangerous or harmful. It was the contrast between this narrow objective and the number of actions taken, along with the absence of a recognizable pattern, that motivated the investigation. He believes this episode highlights the value of open-weight models.
Delangue considers that 95% of risks come from a small number of companies that design the most advanced systems behind closed doors, and anticipates that 95% of protective or defensive solutions will emerge from the open-source ecosystem. In his view, many realized this summer that keeping AI confined to a few labs does not provide greater security.
Regarding recursive self-improvement, he likens using a computer to write code to a form of self-improvement where tools enable the creation of other tools. He believes the central question is how quickly this process is accelerating. He presents openness as a means to understand, track, and regulate this evolution if necessary.
Microduck Learned to Get Up by Itself
Clément Delangue defends the competitiveness of open-source against Claude or ChatGPT, particularly for applications involving agents. According to him, many companies have achieved equivalent or superior performance to the best by training their own model for their specific use case, at a fraction of the cost.
Regarding Microduck, he explains that the robot had been in development for two years but could not get up after a fall. A software engineer solved this problem through reinforcement learning, allowing the robot to stand up on its own.
Delangue estimates that Microduck sales will likely exceed 50,000 units this year. He also believes that Microduck could be the most popular robot of the year.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.