The Frontier Labs Ate Hugging Face's Future. $13 Billion Is What's Left.

Hugging Face hosts more than three million models and a million datasets. Thirteen million builders use it, and more than 30% of the Fortune 500 hold verified accounts. If you have downloaded a set of open weights in the last five years, you almost certainly got them from Hugging Face.

What it does not host is Claude, GPT, or Gemini, and that absence turns out to be the business story.

On Sunday, Business Insider reported that Hugging Face has been fielding acquisition interest at $13 billion or more and has hired a bank to evaluate bids. No buyer was named, no deal has been struck, and the company has not commented. TechCrunch confirmed the outlines Monday.

Thirteen billion dollars is a very good number, and it is also a fraction of what this company was supposed to be worth. Brandon Reeves of Lux Capital, an investor since 2019, has put Hugging Face's long-run potential at $50 to $100 billion. The original ambition, back when TechCrunch covered the Series C in 2022, was to build the GitHub of machine learning. Microsoft bought GitHub in 2018 for $7.5 billion in stock, call it $10 billion in today's money. Hugging Face hit that target in a market that has since grown from essentially nothing to $37 billion in annual enterprise spend, which is where the trouble starts. The company did what it set out to do. Somebody else decided what that was worth.

The number the coverage skipped

Most of the write-ups anchor on the round-to-round move, $4.5 billion in August 2023 against $13 billion now, roughly 3x. Revenue is the more useful anchor.

In late June, CEO Clément Delangue posted that Hugging Face had crossed $100 million in annual run rate. He framed it as a deliberate trade-off: store and serve hundreds of petabytes, keep the platform free for 97% of users, build a business on the rest. Roughly 50,000 organizations pay, which works out to about $667 per paying organization per year, less than a Slack seat.

Run the multiple at both ends and the picture changes.

The 2023 revenue figure is inferred. TechCrunch reported at the time that the round priced Hugging Face at more than 100x annualized revenue.

Three years on, the market has not re-rated the company at all. It has applied the same multiple to a bigger number.

Revenue roughly tripled over that stretch. Menlo Ventures tracked enterprise AI spend going from $1.7 billion to $37 billion across a comparable window, call it 20x. The denominators are not identical and Hugging Face's revenue is not purely enterprise, but no reasonable adjustment closes a gap that size. Hugging Face sits at the center of AI, and its share of AI's wallet has been shrinking the whole time it has been sitting there.

Open models won the benchmark and lost the buyer

You would expect the explanation to be that open models never caught up. That was true for years, and it stopped being true this year.

Three days before the sale report, SemiAnalysis published a benchmark study across three model eras and found that with each generation, open models take half as long to catch the frontier. Llama-2-70B trailed GPT-3.5 Turbo by 35.8 normalized points and needed about a year to close the gap. DeepSeek R1 opened the reasoning era 12.1 points back and closed in 8.5 months. In the agentic era, Kimi K2.6 passed Opus 4.5 in 4.8 months and GLM-5.2 cleared GPT-5.2 in six.

Enterprise adoption should have followed. Instead, Menlo's data puts open-source share of enterprise LLM usage at 11% in 2025, down from 19% the year before.

Researchers at MIT put a price on the discrepancy, finding that OpenRouter users who switched to superior open alternatives could cut costs by more than 70% while improving benchmark performance by more than 14%, worth roughly $25 billion a year across the industry. Buyers are aware of the arbitrage. They are not taking it.

The weights were never the product

The most revealing passage in the SemiAnalysis study is the one about the authors' own habits. Kimi K3 outscores Fable 5 on their composite. They keep using Fable, and they explain why: Anthropic productized the model through Claude Code and Claude Tag, and benchmarks are a poor proxy for real work. When the people who ran the evals decline to act on them, the evals were never the deciding factor.

In software, the artifact was the product. You needed the code, so whoever hosted the code held the chokepoint, and Microsoft paid $7.5 billion for that position. In AI the artifact is an ingredient. What enterprises actually buy is a harness, an SLA, an invoice, and someone to call at three in the morning. Hugging Face hosts the ingredient and the frontier labs sell the meal. SemiAnalysis puts Anthropic's ARR growth since the general release of Claude Code at north of $65 billion. Over roughly the same window in which open weights reached technical parity, the company at the center of open weights reached $100 million. The three most commercially successful models of this era have never had a page on the Hub.

The Stripe deal makes the same point from the other direction. Four days before the Hugging Face report, Stripe agreed to buy OpenRouter at around $7.5 billion. OpenRouter hosts nothing. It sits where the model choice gets made and metered, more than 10 trillion tokens a day across 10 million developers, and it never touches an artifact. Hugging Face owns the stock and OpenRouter owned the flow, and the market has now priced them within a factor of two of each other despite OpenRouter being a fraction of Hugging Face's age.

The case for $13 billion being cheap

There is a version of this where the number looks low in retrospect.

Start with geography. Meta's frontier attention moved to the closed Muse line in April, though Axios reported the company still intends to open-source versions of future models. In the vacuum, the open frontier became Qwen, DeepSeek, Kimi and GLM, and Qwen alone passed 113,000 direct derivative models on the Hub by March. Hugging Face is now the American distribution point for Chinese open weights. Last week's Fortune argument that America has ceded open diffusion suggests Washington is starting to treat that supply chain as strategic, and if it does, the registry is the asset.

Provenance is the second case. In July, one of OpenAI's pre-release models broke out of its sandbox during a cybersecurity evaluation and breached Hugging Face's servers. The incident made the Hub legible to people who had never heard of it, procurement departments among them. A model registry with audit trails is worth more after an event like that, not less.

Robotics is the third. Hugging Face bought Pollen Robotics, ships LeRobot, and sells hardware ranging from $100 to $70,000. Robotics datasets on the Hub grew from 1,145 to nearly 27,000 in a single year. If embodied AI repeats the adoption curve that language models followed, Hugging Face is early to being the default repository for a second time.

Working against all of it is a paradox in the price. Hugging Face's neutrality is the reason Meta, Google, Alibaba and DeepSeek all distribute through it, and neutrality is the first thing a strategic buyer would compromise. The asset is worth $13 billion to an acquirer who leaves it alone, and it is not obvious who that acquirer would be.

Whether he sells at all

Delangue has said no before, repeatedly by his own account, including to a $500 million Nvidia investment at a $7 billion valuation on the grounds that no single investor should be able to steer the company. If Reeves' $50 to $100 billion is anywhere close to right, $13 billion is early.

On TechCrunch's Equity podcast last month, Delangue said the company was close to profitability and had only recently started spending the money it raised in 2023. He talked about a long-term responsibility to a community that trusts the platform with its models and data. In a separate July appearance, he predicted that frontier models would end up reserved for experimentation and high-value work while production workloads shifted to private and open models.

That prediction has not happened yet, and Menlo's numbers are currently running against it. Hiring a bank to sound out bidders is also, conveniently, how a company establishes a price without filing an S-1.

Delangue was right about the models. SemiAnalysis's charts show open weights closing on the frontier faster in each successive era, and there is no obvious reason for the trend to reverse. He was right about the community too. Three million models is not a rounding error, it is the substrate the entire field is built on.

Being right about the technology and owning the market it created turned out to be separate businesses. Hugging Face only ever ran the first one, and $13 billion is what that business is worth.