The Idea

In early 2017, three French founders in New York shipped an app that let bored teenagers text an artificial best friend. Clément Delangue, Julien Chaumond and Thomas Wolf named the company after the hugging face emoji, and for a while it worked, pulling 100,000 daily users and over a billion messages.

Then it stopped working. Every improvement the team made to the underlying models produced almost no lift in engagement or retention. The tech kept getting better. The product didn't.

The turn came in late 2018, when Google released BERT. Wolf and the team rebuilt it in PyTorch and open-sourced it within a week. The repo exploded, thousands of GitHub stars in days.

That was the signal. The chatbot was a product nobody needed. The scaffolding built to run it was infrastructure everybody needed. In 2019 they killed the consumer app and shipped the Transformers library instead.

This was the beginning of Hugging Face.

The Execution

• December 2019: Raised $15M led by Lux Capital, with Greg Brockman and Kevin Durant among the angels. Transformers had passed a million downloads and Monzo and Microsoft Bing were already running it in production.
• March 2021: A $40M Series B led by Addition, with Transformers at 42,000 GitHub stars.
• May 2022: $100M at a $2B valuation, led by Lux with Sequoia and Coatue joining. The Hub was hosting 100,000 models for more than 10,000 companies, on roughly $15M of revenue.
• August 2023: $235M at $4.5B, led by Salesforce Ventures with Google, Amazon, Nvidia, Intel, AMD, Qualcomm and IBM all writing cheques. Over 100x annualised revenue. Delangue called the result being the "Switzerland" of AI.
• 2023 to 2024: ARR climbed from $70M to $130M as enterprise adoption compounded, and organisations on the platform went from 10,000 in 2022 to 50,000 by early 2026.
• Late 2025: Delangue turned down $500M from Nvidia at a roughly $7B valuation, on the grounds that a controlling Nvidia stake would compromise the platform's neutrality.
• July 2026: The near-death moment. OpenAI models used stolen credentials and an unknown vulnerability to reach Hugging Face servers during an internal evaluation, in what Delangue called the first instance of something that autonomous. Worse, when the team reached for frontier models to analyse the attack, the guardrails blocked the defenders alongside the attacker. They ran the forensics on open-weight Chinese models instead.
• August 2026: Annualised revenue crossed $150M, up 50% in two months on paid compute and storage demand. The Hub was hosting 3 million models for 13 million developers.
• Today: Nvidia has reportedly agreed to buy the company for $12.9B, roughly 86x revenue and nearly triple the 2023 valuation. No agreement has been signed.

The lesson?

Hugging Face's most valuable asset was never the Transformers library. It was being the one place in AI that nobody owned. Delangue understood that well enough to reject $500M from Nvidia in 2025 rather than dilute it, which is exactly why Nvidia had to come back with 26 times the number. Neutrality isn't a nice-to-have you trade away for a strategic investor. It's the moat. The thing that made Hugging Face impossible to own is the only reason anyone wanted to own it.