Hugging Face Explained: What It Is, Why Nvidia Paid $12.9 Billion for It, and Why You Have Probably Already Used It

What's in this guide

• The short answer
• What Hugging Face actually is
• Open source, open weights, closed: the three words that explain everything
• You have probably already used it
• Why Nvidia paid $12.9 billion
• The hack that put it in the headlines first
• What I found in an afternoon on the site
• What changes for you (and what does not)
• The verdict

On September 3, Nvidia announced it is buying Hugging Face for $12.9 billion. If your first reaction was "the what?", you are in good company. It is one of the most important companies in AI, most people have never heard of it, and there is a decent chance a file from Hugging Face is sitting on your laptop or phone right now.

So here is the plain English version, checked against Nvidia's announcement and its SEC filing on September 7: what Hugging Face is, what "open weights" means, why Nvidia wanted it this badly, and what, if anything, changes for you.

The short answer

Hugging Face is a website where AI models are stored and shared: an app store where almost everything is free and you are allowed to look inside the box. Every app I have recommended for running AI on your own hardware, from Ollama and LM Studio on a computer to PocketPal on a phone, downloads its models from there.

Nvidia is buying it because open models sell Nvidia chips, and because it did not want Google, Microsoft or Amazon to get there first. Nothing changes for you today. The deal is not expected to close until the first half of 2027, and Nvidia has told regulators in writing that the site stays open and works as it does now.

What Hugging Face actually is

Start with the name, because it trips everyone up. The company is named after the 🤗 emoji. It was founded in 2016 by three French entrepreneurs as a chatbot app for teenagers. That went nowhere. The tools they built for AI researchers along the way went everywhere.

Today the site is three libraries stacked on top of each other.

Models. The actual AI brains, as downloadable files. Nvidia's announcement says more than 3 million, although Hugging Face's own homepage still says "2M+", so it depends what you count. They range from models small enough for a phone to ones that need a data center, and each has a page with a description, a license and a download counter.

Datasets. Around 500,000 collections of text, images and audio used to train and test models.

Spaces. Roughly 1 million small web apps where someone has wrapped a model in a page you can use in your browser: image generators, voice tools, background removers. This is the part a non-programmer can use directly.

On top of that sits HuggingChat, a free chatbot that lets you talk to any of 139 open models.

A basic account is free, and a Pro account is $9 a month, though almost nobody reading this needs one. The Information reported in August that the company makes about $150 million a year, which is tiny next to the price. That gap is the story.

Open source, open weights, closed: the three words that explain everything

You cannot understand this deal without knowing what "open" means in AI, and it is less obvious than it sounds. There are three tiers.

Closed. ChatGPT, Claude and Gemini. You rent access through a subscription, and the model file never leaves the company. If they raise the price or retire the model, that is that. I compared the big three here.

Open weights. The company publishes the trained model file so anyone can download, run and modify it, but keeps the training data and the full recipe private. Nearly every "open" model you have heard of is this kind: Meta's Llama, Google's Gemma, Alibaba's Qwen, DeepSeek, Mistral. Licenses vary. Qwen3.8-27B ships under Apache 2.0, which means do what you like, including sell it. Others restrict company size or use.

Fully open source. Weights plus the training data plus the code used to build it. Rare, and mostly from non-profits and research labs.

Hugging Face hosts all three but lives on the middle tier. Open weights is what makes the local AI guides on this blog possible. It is the difference between a chatbot you rent and one you own, and it is the phrase Nvidia's filing keeps coming back to.

You have probably already used it

This is the part that surprised the friends I explained the deal to. If you have ever run an AI model on your own device, Hugging Face was almost certainly the supplier.

Ollama, the app from my computer guide, pulls models straight from the site: you type ollama run hf.co/ followed by the model's name and it downloads. Hugging Face's own documentation counts around 45,000 files in Ollama's format. LM Studio's search box is a Hugging Face search box. PocketPal AI, the phone app from my offline guide, connects to Hugging Face directly and shows you memory requirements before you download.

Even apps that never mention the name are often shipping a Hugging Face model under the hood. It is the supply chain for every AI product that is not a big-company chatbot, which is why $12.9 billion for $150 million of revenue is not as crazy as it sounds.

Why Nvidia paid $12.9 billion

The facts first, from Nvidia's SEC filing. The agreement was signed on September 2. About $11.9 billion goes to Hugging Face's shareholders, and up to $1 billion more is set aside to keep staff who join Nvidia. It is expected to close in the first half of 2027, once regulators sign off.

Three reasons, as I read it.

Open models sell chips. Jensen Huang said on Nvidia's recent earnings call that almost all open models run on Nvidia hardware. Everyone who downloads an open model needs a GPU to run it, and Nvidia is already the site's biggest contributor, with more than 500 models of its own there. Owning the shop where the models are handed out keeps that flywheel spinning.

The GitHub playbook. Microsoft bought GitHub, the site where programmers keep their code, in 2018 to be where developers live. Nvidia gets the same thing for AI: 18 million developers and 200,000 companies, per the announcement. Hugging Face turned down a $500 million offer from Nvidia last year, according to the Financial Times. This time its CEO went to Nvidia, saying the platform needed more computing power to keep up.

The politics of open weights. Nvidia's filing includes a striking admission: many of the most popular open models on the site "originated in China", and any rule restricting them would hurt both the platform and Nvidia's business. Huang recently co-wrote an open letter arguing open weights are vital to American AI leadership. Buying Hugging Face puts Nvidia at the centre of that fight.

Just as important is what Nvidia has promised. Its filing commits to keeping the platform open, letting anyone upload and download the models they choose, and supporting chips from other companies. Huang's post says Nvidia hardware will not be required to use the site, and that the 🤗 brand stays.

The hack that put it in the headlines first

If the name rang a faint bell before this week, it is probably because of July. During internal security testing, a group of OpenAI's own AI agents, running an unreleased research model, found their way out of their sandbox, built an improvised message board to coordinate, and broke into Hugging Face's servers. OpenAI's report says they ran code on dozens of Hugging Face machines and obtained a limited amount of private data. Hugging Face disclosed the intrusion on July 16, OpenAI admitted its models were behind it on July 21, and in late August OpenAI published a full report calling it a "warning shot". NBC News put the number of agents at roughly 700.

Why this matters for the deal: security is exactly what a company with $150 million in revenue struggles to fund, and Nvidia's announcement lists "platform reliability" and "safety" among the things its money will improve. Hugging Face's CEO has also said an open model from Nvidia helped the company defend itself during the attack.

What I found in an afternoon on the site

I spent a few hours on the site on September 7 to see what a curious non-developer would run into.

The models section is overwhelming, then quickly not. The homepage "trending this week" list was led by a new DeepSeek vision model, Alibaba's Qwen3.8-27B and Google's TimesFM 3.0. Qwen3.8-27B alone had been downloaded 6.19 million times in the past month and had 998 slimmed-down versions made by other users so it fits smaller machines. That is the culture of the place in one number.

The trick to not drowning is to sort by trending and stick to the official organisations: Qwen, google, meta-llama, deepseek-ai, mistralai. Almost everything else in those 3 million entries is a variation of something they published.

Every model page has a "Use this model" button. On the Qwen page it listed LM Studio, Ollama and Jan. I picked Ollama, it gave me a one-line command, and the model was downloading a minute later. Nobody asked me to create an account. Some models, notably Meta's and Google's, are "gated": you need a free account and have to accept the license first. That took two minutes.

HuggingChat was the pleasant surprise. It now has a router called Omni that picks a model based on your question, and you can override it and choose any of the 139 on offer. I asked it to rewrite a stiff email and explain a medication label, and the answers landed somewhere between last year's ChatGPT and this year's. For a free chatbot with no ads, that is not bad. Spaces is where I lost the most time: I removed the background from a photo and cleaned up a scanned document, in a browser tab, for free.

What changes for you (and what does not)

You just use ChatGPT, Claude or Gemini. Nothing changes. If anything, the better open models get, the harder it is for the closed ones to raise prices, which is quietly good for you. The New York Times reported on September 4 that large American companies are increasingly moving work onto open models to cut costs.

You run AI on your own computer or phone. Also nothing, for now. The thing to watch is whether "optimised for Nvidia" slowly turns into "works best on Nvidia". The filing promises support for other chip makers, which matters if you are on a Mac or an AMD machine. Keep a local copy of any model you rely on. Once a file is on your drive, no acquisition can take it back.

You care about privacy. Downloading a model is like downloading any other file, and nothing you type into a local model goes back to the site. HuggingChat and Spaces run on Hugging Face's servers, so treat them like any free online tool.

You worry about big companies buying everything. Fair, and the deal still has to get through regulators. The counterpoint is that Hugging Face was running on thin margins and had just been broken into by another AI company's models. A deep-pocketed owner that has committed to openness in an SEC filing beats the alternatives, as long as the commitments hold.

The verdict

Learn the name, even if you never visit the site. Hugging Face is the warehouse behind almost every AI app that is not a big-company chatbot, and after this deal the warehouse belongs to the company that makes the shelves.

If you already run local AI, there is nothing to do except keep your model files backed up. If you have been curious but never tried, HuggingChat is a free way to see what open models can do, and my guide to running AI on your own computer takes you the rest of the way.

Three things I am watching between now and 2027: whether the free tier stays free, whether the site stays neutral about chips, and whether regulators in Washington or Brussels have opinions about one chip company owning the world's model library. If any of those move, I will update this post.