A trio of French "nerds" arrived in New York to create a cheeky chatbot for teenagers. A decade later, his startup has become an artificial intelligence platform that is a pillar of the technological ecosystem and a key piece in the strategy of the most valuable company in the world.
Nvidia on Thursday agreed to buy its AI company, called Hugging Face, for about $13 billion, doubling down on the chipmaker's effort to challenge the dominance of proprietary AI models from OpenAI and Anthropic. With the deal, Nvidia is helping foster the ecosystem that Hugging Face built for so-called open-weight AI tools, which anyone can download and modify.
But when Clément Delangue and Julien Chaumond first arrived at the Manhattan offices of venture capital fund Betaworks for its “BotCamp” startup program in the sweltering summer of 2016, their goal was more modest. They wanted to create a general-purpose chatbot that wasn't boring like Apple's Siri. The release of OpenAI's ChatGPT was more than seven years away.
“It won't always give you the right answer, but it will always give you a fun one,” Delangue, the chief executive, said during a demo that summer, in which the app exchanged selfies with Chaumond, the chief technology officer. "And of course, he's always very cheeky."
The name of the app - Hugging Face - was a provisional name they had put on their application for the program, based on the name of their favorite emoji. They intended to change it later, but never did.
The following spring, Thomas Wolf, a physicist turned patent attorney, joined the founders as chief scientific officer.
"They're deeply nerdy. They're deeply into science," John Borthwick, Betaworks managing partner and current investor, said in an interview.
Raised in northern France, Delangue has long been involved in digital collaboration and open systems. As a teenager, he sold motorcycles on eBay and later worked for the company, according to a profile on the website of Sequoia Capital, one of Hugging Face's investors.
As a university student, he developed a platform for students to share notes and knowledge from their classes with others, according to a TEDx talk he gave in Paris in 2012.
Delangue was already working in product and marketing for startups in New York when he and Chaumond – a mathematician who had worked for startups and for the French government – joined forces. They pursued a shared interest in natural language processing, a type of computer program they thought could be used to create a chatbot, and Chaumond flew from Paris to New York.
"It was the first time we worked together in real life (IRL) for more than a day in a row," Chaumond wrote.
They took an online class at Stanford with Richard Socher, a proponent of natural language processing, and formed a study group on the topic. Wolf was also in the class, Socher said.
"I was very impressed by them. I thought they asked intelligent questions," said Socher, who is now founder and CEO of AI startup Recursive. He sometimes goes tree climbing on walks in the Bay Area with Delangue.
A major turning point for Hugging Face came in 2018, when Google launched a new version of its transformer, the neural network technology at the core of modern generative AI models.
Wolf told his co-founders that he would take the weekend to make an open source version in the most popular language accessible to AI researchers, Delangue recounted in a 2024 podcast episode. By Monday, the version of Hugging Face had gone viral, and six months later, it had become a new business model.
"We think we need to pivot," Delangue recalled telling investors.
As AI models grew in size, developers began to need a new platform to host them. Hugging Face built one. In practice, it created an AI model sharing platform similar to what Microsoft's GitHub is for software developers' code, a role that has grown as the AI revolution has taken off.
Today, there are millions of open weight models on Hugging Face.
The company made headlines in July when it shared that it had been hacked by an autonomous swarm of AI agents, which it later discovered had escaped an internal cybersecurity test at OpenAI. The company attempted to analyze the attack logs using proprietary AI models, but was unable to due to security barriers, so instead it turned to an open weights model from a Chinese developer.
Since the incident, Delangue has been very outspoken about the need to preserve and promote access to high-quality open source AI models so that potential targets of cyber attacks can be effectively defended. Nvidia has done it too.
Borthwick said he believes Nvidia's investment makes sense for Hugging Face's goal of democratizing agentic AI beyond the 18 million developers currently on the platform. "Nvidia is making that growth possible through this acquisition," he said.
Delangue said Thursday on CNBC that he approached Nvidia over the summer to seek a deal. He saw the world facing a choice between a path in which proprietary models like those from Anthropic or OpenAI dominate the field, and another in which open weight models proliferate.
"There is a path where open source AI is available to everyone, and everyone can become an owner, a builder of AI, and not just rent it or use it from other people," Delangue said. "We needed to double down on open source AI to really distribute the technology as much as we can around the world."
