Jose Romero
← All letters
Prefer email? Read this on Substack

Nvidia is buying Hugging Face. It could have been worse

Hugging Face is where I download every model I run. The Qwen quants from the last three videos, the GLM file I spent a night on, Nvidia's own Nemotron models: all of it comes off that site. So when the reports landed that Nvidia has agreed to buy it for $12.9 billion, I went through the articles on camera instead of reacting to the headline.

First, the status, because it matters: this is reported, not announced. The Information says the deal is agreed at $12.9 billion. Business Insider, the same night, said Nvidia was in talks above $13 billion with no signed agreement and that it could still fall apart. Neither company has confirmed. Everything below can still change.

The numbers

Hugging Face does not publish its revenue. The number everyone is working from is what its CEO told TechCrunch: about $150 million a year, up from roughly $100 million two months earlier, and "close to profitability." Against a $12.9 billion price that is about 86 times revenue.

That multiple looks insane until you stop reading it as a revenue purchase. The last funding round, in August 2023, was $235 million at a $4.5 billion valuation, led by Salesforce with Google, Amazon, Nvidia, AMD, Intel, IBM and Qualcomm in the round. Revenue has gone from about $50 million in 2024 to $150 million now. And for Nvidia, which just reported a $96 billion quarter, $12.9 billion is about five weeks of sales. They are buying the on-ramp, not the P&L: four million models, the datasets, the libraries every local-AI tool calls into, and the place developers go to decide which model to run.

The closest precedent is GitHub. Microsoft paid $7.5 billion for it in 2018 on roughly $300 million of revenue, about 25 times. GitHub was doing a billion a year by 2022. Same shape, different multiple.

They said no once

The part of the story I find most interesting is not in the video, because I only pinned it down afterwards. Late last year Nvidia offered Hugging Face $500 million at a $7 billion valuation, and Hugging Face turned it down. The reason reported by the Financial Times: it did not want a single dominant investor.

Nine months later it hired a bank, shopped itself at $13 billion or more, and (reportedly) agreed to sell the whole company to that same investor. Two things changed in between. Revenue tripled, which gave them leverage. And Nvidia spent 2026 becoming the loudest open-weights company in the industry, which made the buyer they had been wary of in December look like the most aligned one by August.

The letter both of them signed

On July 24, five weeks before any of this, Jensen Huang posted his first-ever tweet: a letter called "Open Weights and American AI Leadership," hosted by Microsoft and signed by 25 companies on day one. Nvidia and Hugging Face were both on it, next to Meta, Microsoft, Mistral, IBM, Dell, Mozilla, the Linux Foundation and others. It asked Washington not to put "premature restrictions" on open models, at a moment when the administration was floating a ban on Chinese open-weight models after Kimi K3.

Who was missing on day one tells you as much: OpenAI, Anthropic, Google, Amazon, Apple, xAI. OpenAI and Google added their names within about a day, Amazon later; the list passed 270 signatures by early August. Anthropic, Apple and xAI never signed. Anthropic published its own position instead, in which Dario Amodei wrote that Anthropic "has never advocated for a ban on open-weights models" and called open models without dangerous capabilities "a public good," while disagreeing with the letter on whether openness helps defenders more than attackers. On August 5 the White House's review framework exempted open-weight models entirely. The letter's side won.

I bring this up because it is the sourced version of the thing I say on camera: it could have been worse. The buyer and the seller lobbied Washington together, for the exact thing the site exists for. The frontier lab that would not sign is the one I keep imagining as the alternative owner. That is my thought experiment, not a report; nobody has reported any other bidder.

Why Nvidia is the less-bad owner

Nvidia's whole posture this year is that it has a lot of rich friends. It invests in the labs ($30 billion in OpenAI, up to $10 billion in Anthropic, checks to xAI, Mistral, Cohere, Perplexity and a dozen more), and Jensen's line on a podcast in April was "we don't pick winners, we need to support everyone." If you are a frontier lab, they sell you GPUs. If you are a gamer who wants to run a model at home, they sell you a GPU. If you want to build a data center to host open weights for other people, they sell you the racks.

The financial logic for open weights is the same logic. Jensen's own number is that roughly one in four AI tokens generated today come from an open model, and he wants that share to climb. Every open model that runs on someone's own hardware is Nvidia hardware somewhere. A closed lab owning the open-weights hub would have an incentive to route you to its API. Nvidia has an incentive to keep the weights open and downloadable, because that is what sells chips.

What I am actually watching

Am I glad Hugging Face is getting bought? No. Am I worried? Not really, but not for nothing either. Nothing about the licenses on models already published changes; a file you downloaded under MIT is yours. What can change is quieter: subscription tiers, "recommended hardware" badges on model cards, which quant formats get first-class support, and whether the AMD, Intel and Apple optimization work that Hugging Face engineers ship today keeps shipping. My rig is AMD. I am in the middle of a ROCm-versus-Vulkan comparison on it. This lands on my desk.

There is also a regulator in the way. A full acquisition means a Hart-Scott-Rodino filing in the US and a likely look from the EU, and Nvidia is already under antitrust inquiries in both over how it allocates GPUs. AMD, Intel and the clouds building their own chips all have reasons to object. This may not close quickly.

GitHub cuts both ways as a precedent. Microsoft promised independence, kept it, and GitHub roughly tripled. But nobody runs GitHub on a competitor's platform. Hugging Face's value is that it runs on everyone's. What I am doing in the meantime is boring: mirroring the model files I actually depend on to my own storage, and keeping llama.cpp and Ollama as the layer I trust.

Sources

Get the letters

Every post goes out free on Substack. What I build, break, and fix, written up in plain English. No spam, unsubscribe anytime.

Subscribe on Substack