Nvidia has agreed to buy Hugging Face for $12.9 billion, in a deal valuing the company at more than $13 billion. It would be, by far, the largest acquisition in Nvidia's history. And as this was written, no agreement had been signed — according to The Information's reporting, the talks could still fall apart.
That is worth stating first, because the headline that travelled through several newsletters was “Nvidia bought”. It did not. It agreed on a price.
What was reported, and by whom
Business Insider broke the story over the weekend and The Information detailed it on Wednesday, with the number. CNBC, Forbes and others picked it up the next day. There is no official statement from Nvidia — which means the terms, and the existence of the deal itself, can still change. This is a story about a negotiation, not a completed fact, and that difference is exactly what tends to disappear on the way to the feed.
Three moves, not one
The detail the headline hides is that this is not the first approach. There are three, and read in sequence they tell a different story:
2023. Hugging Face raised $235 million in a round valuing it at $4.5 billion, led by Salesforce Ventures, with GV (Alphabet), IBM Ventures — and Nvidia itself. In other words: Nvidia was already a shareholder.
Late 2025. Nvidia offered $500 million for a stake at a $7 billion valuation. Hugging Face turned it down.
August 2026. The agreed price is $12.9 billion.
In under a year the value nearly doubled after a refusal. One number helps explain it: Hugging Face's revenue was running at about $150 million a year, up from roughly $100 million just two months earlier, with its president saying the company was close to profitability. This is not a company sold under pressure.
Why a chipmaker buys a model repository
The reported rationale is defensive and offensive at once. Defensive: the big closed labs are building their own chips to depend less on Nvidia, and keeping the open ecosystem strong keeps an entire market buying its cards. Offensive: Hugging Face is a way back into the cloud business Nvidia had scaled down, and a channel to sell spare compute to customers who are already there.
On figures published in 2025, the platform held more than 2 million models and more than 13,000 companies using its services. Nobody pays $12.9 billion for code that is already open. You pay for distribution — for the path along which models reach the people who put them into production.
We have cited Hugging Face four times — always as scenery
That is what made us write this. Looking back at our own blog:
When we reported that the world's most downloaded AI model runs on a used graphics card, the download count came from there. When OpenAI paused its own training, the model that had escaped the lab was hosted there. When 1,200 AI agents found each other and nobody noticed for two months, the target of the attack was Hugging Face — and it was the victim, not the owner of the agents, who raised the alarm. And when AI started operating laboratory instruments, the reference parts were, again, there.
In all four, Hugging Face shows up as a place: neutral, nobody's and everybody's, the ground the thing happens on. It is that assumption that changes once the ground has an owner — and an owner with a declared interest in selling hardware.
What actually changes for an ordinary company
Almost nothing today; something later. The two halves are worth separating, because they usually get mixed up.
What does not change: an open model stays open. The licence of an already published model is not revoked because the shelf changed hands, and whoever downloaded it still has the file. If your company uses AI through third-party APIs — ChatGPT, Claude, Gemini — this deal does not touch your day-to-day. For most small and mid-sized companies, that is exactly the case.
What does change: the layer through which open models reach production stops being neutral. That is the point Forbes raised, and it is what matters to anyone running their own software: what was treated as common infrastructure becomes the infrastructure of a supplier with a hardware preference. It is not a catastrophe. It is a change in the nature of a dependency almost nobody had written down as one.
Three questions to answer this week
1. Do you know where the model running in your process comes from? In most companies the honest answer is no — it arrives embedded in a library, a plugin or a third-party system, and nobody ever needed to know. Building that list takes hours, and it is the only one of the three questions you cannot answer from memory.
2. What happens if that download fails tomorrow? If any of your processes pulls a model in real time, it has a single point of failure that is not on your premises. The fix is cheap and old: keep a local copy of what is critical, and do not design a routine that depends on fetching the file on demand.
3. What would switching cost? Nothing needs switching now. What you need is the price of switching — in hours, in rework, in risk. That number is the real size of your dependency, and it applies to any supplier, not just this one.
The takeaway
The move itself is consolidation, and it will continue: AI infrastructure is getting too expensive to stay scattered. But the practical lesson does not depend on the deal being signed, and it is the same one that has come up in other stories we have told here: the infrastructure everybody treats as landscape always belongs to somebody. Finding out to whom, and what it would cost to leave, is half a week of work — and it is far cheaper done before than after.
If your company is starting to put AI into real processes and those three questions came back unanswered, that is exactly where our AI implementation for business comes in: mapping what is already running, where the dependencies sit, and what is worth bringing in-house before you build on top of someone else's ground.
Frequently asked questions
Has Nvidia actually bought Hugging Face?
Not yet. What was reported is an agreement on price — $12.9 billion, valuing the company at more than $13 billion — broken by Business Insider and detailed by The Information. The reporting itself notes that no agreement had been signed and that the talks could still fall apart. There has been no official statement from Nvidia. In other words: a reported negotiation, not a closed deal.
What is Hugging Face?
It is the largest public platform for sharing artificial intelligence models and datasets — in practice, the place where the community publishes and downloads open models. Founded in 2016, by 2025 it held more than 2 million models and more than 13,000 companies using its services. It is the route through which most open models reach the people who put them into production.
Will open models stop being free?
Not because of an acquisition. The licence of an already published model is not revoked when the platform changes hands, and whoever downloaded it keeps the file. What a change of control can affect over time are the platform's own terms — hosting, usage limits, integrations and product priorities — not the licence of models already published.
My company uses ChatGPT or Claude through an API. Does this affect me?
Not directly. Anyone consuming AI through third-party APIs never touches Hugging Face, and this deal does not change their day-to-day. It is different for anyone running their own software that downloads, hosts or fine-tunes open models: there the dependency is real, and it is worth knowing what it is.
How do I find out which AI models my company uses?
Start with the systems that already have some AI function and ask the vendor or the development team three things: which model it is, where it is obtained from, and whether a local copy is kept. Models usually arrive embedded in a library, a plugin or a third-party system, so this is rarely documented — but the inventory takes hours, not weeks, and is useful well beyond this case.
Why would Nvidia pay nearly twice what it offered last year?
In late 2025 Nvidia offered $500 million for a stake at a $7 billion valuation and was turned down. In the meantime Hugging Face's revenue went from roughly $100 million to roughly $150 million a year in two months, with the company close to profitability. On top of that, keeping the open ecosystem strong protects demand for Nvidia cards at precisely the moment the big labs are building their own chips.