Nvidia's Hugging Face

The decision

On September 3, Nvidia agreed to buy Hugging Face for $12.9 billion. About $11.9 billion is cash, and up to $1 billion goes to keeping staff.

If you haven’t used it, Hugging Face is the public library of AI. It hosts more than 3 million models and 500,000 datasets. 18 million developers and 200,000 companies use it to find, download and share models. When Meta, Mistral, DeepSeek or Alibaba release an open model, it almost always lands on Hugging Face. The service is free for most people.

Its revenue is about $150 million a year. Nvidia paid roughly 86x that.

The deal has a backstory. Hugging Face reportedly turned down a $500 million Nvidia investment earlier because it worried about one investor holding too much influence. Then this summer, CEO Clément Delangue went to Jensen Huang himself. The company that refused a minority stake to protect its independence sold all of it to the same buyer.

Huang made one promise on the record: “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.”

Nvidia now owns the front door to open AI. That is the decision.

The strategic why

To see why Huang wants this, you have to see why open models cut both ways for Nvidia.

Open models are Nvidia’s best customers. Nvidia’s biggest buyers are a handful of giants: Microsoft, Amazon, Google, Meta and the labs they fund. Every one of them is building its own chips. Google has TPUs. Amazon has Trainium. That concentration is Nvidia’s quiet weakness. Open models fix it. When a bank, a hospital system or a country can download a strong model for free, it needs its own hardware to run it, and most of the time that hardware is Nvidia’s. Open models turn a handful of huge customers into tens of thousands of smaller ones.

Huang knows this. In July he published an open letter, “Open Weights and American AI Leadership,” and recruited more than 150 organizations to sign it. Hugging Face was one of the first 25. Six weeks later he bought it.

Open models are also Nvidia’s exit door. Nvidia’s real moat isn’t the chip. It’s CUDA, the software that roughly 4 million developers use to program its chips. Twenty years of tools, libraries and habits are built on it. Closed models stay inside a few labs. Open models travel, and anything that travels can be rewritten to run on someone else’s silicon.

DeepSeek showed how on October 1. It open-sourced TileLang, a programming language it pitches as a simpler alternative to CUDA, plus 6 libraries built for Huawei’s Ascend chips. Huawei helped build them. DeepSeek plans to run at least 160,000 Huawei chips in one Inner Mongolia data center. US export controls keep Nvidia’s best chips out of China, so China is building the road around CUDA. The road is open source, which means anyone can use it.

So the question for Huang was never whether open models would win. It was who would sit where they’re distributed. Hugging Face is that place. Every model card, every download button and every “deploy” menu is a chance to show what runs best where. Nvidia doesn’t need to block anyone. It only needs Nvidia to be the default answer.

There’s also a quieter prize: information. Owning Hugging Face shows Nvidia which models developers are downloading, which ones are rising and which hardware they’re asking for, months before any of it shows up in a purchase order. For a company that has to plan chip production years ahead, that view alone is worth a lot.

The price was small. Nvidia says it will sell twice as many chips in 2027 as in 2026. On September 28, it added $150 billion to its buyback, the largest in history. Hugging Face cost less than 9% of that.

Huang didn’t pay to stop open models. He paid to stand where they arrive.

What it defends against

The CUDA escape route. DeepSeek’s TileLang is the clearest threat to Nvidia’s software lock-in so far, because it’s free, it’s open and it comes with frontier models attached. The best defense against a portable model ecosystem is owning the place that ecosystem lives. If Nvidia-optimized versions of every popular model are one click away on Hugging Face, the escape route still exists. It’s just longer.

The giants’ custom chips. Google, Amazon and the labs want to run their own models on their own silicon. Nvidia can’t stop that. What it can do is make sure the rest of the market, everyone who doesn’t build chips, grows faster than the giants do. Open models are how that market grows, and Hugging Face is where it shops.

Being a parts supplier. Before this deal, Nvidia reached developers through clouds and labs that other companies owned. That’s why its distribution control scored a 6 on our index. Now it owns a direct relationship with 18 million developers. For the first time, Nvidia doesn’t have to rent its way to them.

The risk

The asset is neutrality, and Nvidia can’t own neutrality. It can only borrow it.

Hugging Face is valuable because AMD, Intel, Google and Amazon all trust it. Its libraries support their chips. Its developers pick hardware freely. The moment a model page ranks Nvidia first, a new backend ships late for AMD, or the free tier quietly favors CUDA, that trust starts leaking. There’s a precedent. Microsoft bought GitHub in 2018 and mostly kept it open, and GitHub survived. Docker Hub tightened its terms after it was commercialized, and developers started looking elsewhere.

Second, regulators. Nvidia’s earlier deals with Groq, Enfabrica and Poolside were structured as licenses plus hires, which kept them out of formal merger review. This one is a straight acquisition, so it needs an FTC and DOJ filing in the US and a review in Europe. That could run into 2027. Nvidia’s argument is creative. Its executive Justin Boitano called Hugging Face “almost structurally by definition kind of like a deconcentration platform.” Regulators may not buy that line from the most concentrated company in AI.

Third, the deal could speed up the very thing it defends against. Developers who were happy on a neutral Hugging Face now have a reason to set up a mirror outside Nvidia’s reach. DeepSeek and Huawei already have the tools. They just got a better pitch.

Huang bought the town square. Now he has to stay out of it.