Three Founders Spent a Decade Building an AI Community. Nvidia May Pay $12.9 Billion for It

Three French founders spent nearly three years building an AI companion for teenagers.

Ten years later, Nvidia is reportedly prepared to pay $12.9 billion for what they built next.

The Information reported on August 26 that Nvidia had agreed to acquire Hugging Face. Business Insider was more cautious: the companies had not signed an agreement, and the talks could still fall apart. Neither Nvidia nor Hugging Face has publicly confirmed the deal. TechCrunch put both accounts side by side.

So it is too early to say Hugging Face has been sold.

But the reported price already poses a useful question. Hugging Face is said to be generating about $150 million in annualized revenue. Why would Nvidia offer roughly 86 times that number?

The answer begins with a startup that did not work as planned.

They wanted to build a friend who lived in your phone

Clément Delangue, Julien Chaumond, and Thomas Wolf started Hugging Face in Paris in 2016.

Chaumond and Wolf knew each other from engineering school. They had even played together in a short-lived band that covered Alanis Morissette. Chaumond already knew Delangue. The three wanted to build something scientifically difficult but also fun.

Their answer was a chatbot for teenagers.

This was years before ChatGPT. Siri and Alexa could set timers and check the weather, but the founders thought assistant-style conversations were boring. They imagined Hugging Face as an AI Tamagotchi: less a tool than a virtual friend.

The idea was hard to finance in France. Chaumond later said they were pitching a consumer product with no monetization plan—a virtual companion for teenagers. Betaworks in the United States gave them an initial $200,000 check and a place in its New York chatbot accelerator, and the team crossed the Atlantic.

They worked on the product for almost three years. Users exchanged billions of messages with it. It was not unused, but it did not give the company a convincing future.

The decisive turn did not come from an offsite or a grand strategy deck.

It came on a Friday night.

One weekend changed the company

In late 2018, Google released BERT, a language model that set a new standard for natural-language processing. Google’s implementation used TensorFlow, while much of the research community preferred PyTorch.

Thomas Wolf told his co-founders that he wanted to spend the weekend porting BERT to PyTorch.

Their response was essentially: sure, have fun.

On Monday, Wolf published the code on GitHub and tweeted about it. The highly specific technical post received around 1,000 likes.

That would be a small signal for Hugging Face today. For three relatively unknown founders still making a consumer chatbot, it was impossible to miss.

Developers began fixing bugs and contributing other models. They had not arrived because Hugging Face announced a pivot. They came because the repository made the newest AI research easier to use.

The founders followed the signal. They added models, improved the tools, and supported the emerging community. In 2019, Hugging Face formally moved away from the consumer chatbot and toward open machine-learning infrastructure. The repository became Transformers; models, datasets, and interactive demos followed.

Delangue has said the company did not first decide to become “the GitHub of AI” and then execute a master plan. It watched the community vote with its behavior and followed.

That is the most valuable part of the founding story: the original product consumed almost three years; the direction that may now be worth $12.9 billion first appeared as a founder’s weekend side project.

How free models became a multibillion-dollar platform

Developers now use Hugging Face to publish and find models, inspect model cards, download datasets, and turn demos into interactive Spaces. Much of that activity is free.

But it creates something difficult to reproduce: a map of which models matter, who maintains them, what developers are trying to build, and what compute they will need next.

Hugging Face makes money from enterprise features, private hosting, storage, and compute. In 2023, it raised $235 million at a $4.5 billion valuation. The investors included companies that compete with one another across the AI stack: Google, Amazon, Nvidia, Intel, AMD, Qualcomm, IBM, and Salesforce.

Delangue described that structure as a kind of Switzerland for AI. No single giant received privileged access, which helped the platform serve competing models, clouds, and chips.

That neutrality is commercially valuable.

When developers choose a model, they rarely begin on a chip company’s sales page. They search Hugging Face, compare downloads, read documentation, try a demo, and only then decide where to run it.

The company at the beginning of that path is close to the next compute purchase.

Nvidia is already inside the loop

Nvidia is an investor, a customer, and a partner of Hugging Face.

In March 2025, Delangue welcomed Nvidia as an enterprise customer and said nearly 2,000 Nvidia employees were using Hugging Face. Nvidia’s organization had close to 20,000 followers on the platform.

Three months later, the companies launched Training Cluster as a Service. Organizations on Hugging Face could request GPU clusters for training, with Nvidia’s DGX Cloud Lepton underneath. Hugging Face said 250,000 organizations on the platform could request the service.

The loop is straightforward:

1. A developer finds a model on Hugging Face.
2. Hugging Face helps the developer obtain training or inference compute.
3. Nvidia hardware powers that compute.
4. More open models are produced, drawing the next group of developers back to Hugging Face.

As an investor, Nvidia can participate in that loop. As the owner, it would control the loop itself.

That is how an 86-times-revenue price begins to make sense. The reported $12.9 billion is not a conventional valuation of current subscriptions. It is a price for the default entrance to open AI.

From rejecting a dominant investor to considering a sale

There is an uncomfortable contradiction in the reported deal.

The Financial Times previously reported that Hugging Face rejected a $500 million Nvidia investment in 2025 at a proposed $7 billion valuation. Hugging Face reportedly did not want one dominant shareholder influencing its decisions.

Less than a year later, the number on the table is $12.9 billion, and the proposal is no longer an investment but a full acquisition.

Those offers create different kinds of pressure. A dominant minority investor can reduce a founder’s control while leaving the company responsible for years of independent growth. A sale exchanges all control for a certain outcome.

No outsider can make that decision for the founders. Hugging Face’s annualized revenue reportedly rose from roughly $100 million to $150 million in two months, and Delangue has said the business is close to profitability. It is not obviously selling out of desperation.

That is precisely what makes the offer difficult. When someone puts a price far above current revenue on a decade of work, independence stops being a slogan and becomes a measurable opportunity cost.

The models will not vanish overnight. Neutrality is the real question

If the acquisition happens, free models on Hugging Face will not all become paid products the next morning.

Models are uploaded by different developers and organizations under their own licenses. Buying the platform does not give Nvidia ownership of every model. Locking everything down would also destroy the developer network Nvidia is reportedly willing to pay $12.9 billion to obtain.

The important changes would be subtler. Will search and recommendations remain neutral? Will AMD hardware and Google TPUs receive equal treatment? Will enterprise data and models become more deeply tied to Nvidia’s ecosystem? Can a chip company own the platform and resist turning the easiest choice into the choice most favorable to its chips?

There are no answers yet.

But Hugging Face once invited competing giants into the same funding round to demonstrate its neutrality. If one of them becomes the sole owner, the phrase “Switzerland of AI” will need a new explanation.

Nvidia wants the habit created after that weekend

The three founders of Hugging Face did not train GPT or manufacture GPUs.

They did something less dramatic: they turned models scattered across papers, labs, and companies into things developers could find, download, and run.

It began as Thomas Wolf’s weekend port. Someone liked it, someone fixed a bug, and someone uploaded the next model. Ten years of those actions became an entrance.

The reported $12.9 billion price is for the habit behind that entrance: where a developer goes first when they want to try a new model.

Chips change generations. Popular models change names. Habits are much harder to rebuild with money and code.

Three founders set out to make an AI friend for teenagers. That product did not become the company they imagined.

The tool they made for developers on the side may become one of the most expensive entrances Nvidia has ever bought.