In late 2025, Hugging Face turned down a $500M investment from Nvidia. Less than a year later, Nvidia agreed to buy the whole company for $12.93B.

Hugging Face is where the world's open AI models live: 3 million models, half a million datasets and more than 18 million developers, according to Nvidia.

It gives almost everything away and charges companies for privacy, security and compute, which adds up to about $150M in annualized revenue, growing 50% a year.

It raised $395M over six rounds, and the story it told investors changed completely three times. That is the part most teardowns skip.

Hugging Face never published a pitch deck. But every round left a public trail: the metric it chose to announce, who it let onto the cap table, and how it described itself. Put together, the trail reads like the deck. Below is the full breakdown.

Inside you'll find:

Most people assume Hugging Face makes money from AI models. It doesn't. It makes money from everything around them.

The business has three layers:

→ The product developers love is not the product companies pay for.

A developer downloads a model for free. Then her company needs that model private, behind single sign-on, audited, and running on a GPU. That is the invoice. By June 2025, more than 2,000 enterprises were paying, including Intel, Pfizer, Bloomberg and eBay (Contrary Research).

Revenue came slowly, then fast: about $10M in 2021, the first year of paid features (Forbes), about $70M in ARR by the end of 2023 (Sacra estimate) and $150M annualized by August 2026 (The Information).

The most telling detail is where the 2023 jump came from. Sacra estimated that much of it came from large contracts with big AI companies, not from developers upgrading one by one. The free community created the demand, and a few large deals turned it into revenue. If you are building open source, plan for both motions from day one: self-serve for the long tail, a sales team for the top 50 accounts.

Here is every round, with the story and the number it was announced with:

Read the metrics from top to bottom: user engagement, then developer adoption, then company adoption, then the size of the ecosystem. Each shift describes a different business, and each attracted a different kind of investor.

→ Your lead metric should describe the company you are now, not the company you started.

In 2019, 1 million messages a day was a bigger-sounding number than 19,000 GitHub stars. Leading with it would still have been a mistake. Lux did not back a chatbot with traction. It backed a library that researchers at Google, Microsoft and Facebook were already experimenting with, for free.

By late 2018, the chatbot had raised $5.2M and handled a million messages a day. By consumer-app standards, it was working.

Then, in November 2018, Google open-sourced BERT, the strongest language model of the moment. It ran on Google's TensorFlow, while many researchers worked in Facebook's PyTorch. Within days, Thomas Wolf and two teammates posted a PyTorch version that matched Google's results, and gave it away. It was not a strategy. The team worked in PyTorch and wanted BERT for its own work.

About a year later the library had passed 1 million downloads. The founders then made the call most founders avoid: they wound down the product they had raised money on and pitched the side project instead.

Slack was born the same way, as the internal chat tool of Tiny Speck, the studio behind a game called Glitch that failed. The pattern repeats because users, not founders, pick the product.

Run these three questions on your own company:

Two yeses mean a pivot conversation. Three mean you have found your next pitch.

The Series D is the most unusual round in this story. Most of the $235M came from Google, Amazon, Nvidia, Salesforce, Intel, AMD, Qualcomm and IBM, companies that compete with each other on chips, clouds and models.

Most founders avoid strategic investors because one of them can scare off the rest. Hugging Face took eight at once, so that no single one could steer the company.

→ One strategic investor is a boss. Eight are a market.

The neutrality had commercial value. Every one of those giants wanted its models and chips where developers already were. By 2026, Nvidia alone had published more than 500 models and 250 datasets on Hugging Face (Jensen Huang, via TechCrunch).

If your customers compete with each other, belonging to none of them is part of your product. Treat it as a fundraising constraint, not just a roadmap value: decide in advance who you will never let lead a round.

In late 2025, Nvidia offered to invest $500M at a $7B valuation, more than Hugging Face had raised in its entire life. The company declined, saying it did not want a single dominant investor that could sway its decisions (Financial Times).

In September 2026, Nvidia confirmed it would buy Hugging Face for $12.93B: about $11.9B for shareholders plus up to $1B in equity awards for employees who join. The deal is expected to close in the first half of 2027. Clem Delangue said the company approached Jensen Huang because scaling further required "more compute, more support, more collaboration, and more visibility" (TechCrunch).

The numbers behind the exit:

Look at what Nvidia had to promise in public: Hugging Face "will remain an open platform for the entire AI ecosystem," and "Nvidia compute will not be required to build on or deploy through Hugging Face" (Jensen Huang). Nvidia was paying for the neutrality, so it had to commit to keeping it.

Two habits made the earlier "no" possible:

A company that needs the money takes the first term sheet. A company that doesn't can wait for the right price.

This 10-slide outline is reconstructed from Hugging Face's public pitch trail. It works for any open-source, developer-first or community-led startup.