NVIDIA Buys Hugging Face for $12.93 Billion
The number on NVIDIA’s blog is oddly specific: $12,930,300,000. Not “about $13 billion”, not a rounded headline figure, but twelve billion nine hundred thirty million three hundred thousand dollars. What it buys is Hugging Face — the platform where, in practice, the world puts its models, finds them, and downloads them.
This is NVIDIA’s second-largest acquisition ever, behind the roughly $20 billion Groq deal in December 2025. TechCrunch confirmed the transaction on September 3; CNBC reported that Hugging Face CEO Clem Delangue approached Jensen Huang weeks before the announcement — nobody knocked on his door, he knocked on theirs.
The two deals are also structurally different, which often gets glossed over. Groq was not strictly a company purchase: NVIDIA paid for a perpetual, non-exclusive licence to Groq’s patent portfolio and software, plus the core team including its CEO, while the Groq entity nominally continued to exist. Hugging Face is an outright acquisition of the company. This is the first time NVIDIA has bought the front door to a software ecosystem rather than technology or talent.
What was bought is not a company but the ecosystem’s junction
To see why this matters, start with the platform’s scale, as NVIDIA itself lists it:
- 18 million developers, researchers, and creators
- 3 million models
- 500,000 datasets
- 1 million applications
- 200,000+ companies discovering, evaluating, customising, and deploying AI there
The significance of these figures is not that they are large but that they are concentrated. If you want an open model today — the weights, the fine-tuning recipes, someone else’s variant — there is effectively one place to go. Hugging Face is not merely a website; it is the open AI ecosystem’s default infrastructure, the way npm is for JavaScript and PyPI is for Python.
That position is now owned by the world’s largest AI chip company.
Huang’s one sentence is the most important line in the announcement
NVIDIA clearly knew what the worry would be, and addressed it flatly:
“Developers will choose the models they want, the frameworks they want, the clouds and inference service providers they want and the computing platforms they want. NVIDIA compute will not be required to build on or deploy through Hugging Face.”
Huang added that “Hugging Face will remain an open platform for the entire AI ecosystem”, and coverage at VideoCardz notes the platform stays open to AMD and other hardware vendors.
The value of that sentence is that it rules out the most immediate fear: that after the purchase, models targeting AMD, Google TPUs, or in-house silicon would quietly slide down the rankings, or that the CUDA path would become conspicuously smoother than everything else.
What has to be said plainly, though: this is a promise, not a binding commitment. The announcement names no time limit, no governance mechanism, no third-party oversight. Whether it holds, and for how long, currently rests entirely with NVIDIA. Expect the open-source community to read every future platform change with a magnifying glass.
Deal structure: what the announcement leaves out
What is missing from the official post is worth noting. NVIDIA’s blog gives the headline figure and nothing else — no cash-versus-stock split, no closing conditions, no expected completion date.
Several financial outlets report the structure as roughly $11.9 billion in cash plus up to $1 billion in equity for staff retention — but that breakdown comes from secondary reporting and does not appear in NVIDIA’s own announcement, a distinction worth preserving when citing it. Anyone tracking the deal should watch for regulatory progress and a formal closing announcement.
Why Hugging Face wanted to sell
CNBC supplies a detail that is easy to skip past: Delangue went to Huang, weeks ahead of the announcement, framed as considering “the next chapter of Hugging Face”.
The direction of travel matters. A company sitting at the centre of an ecosystem, with 18 million developers, does not usually seek a buyer because the business is failing. It does so because it can see a next stage it cannot fund. Storage and bandwidth for model files scale non-linearly as models grow, and Hugging Face’s core service is free for the overwhelming majority of its users. Who pays that infrastructure bill has always been the hardest question in this business model.
NVIDIA’s stated intention — to “scale Hugging Face’s platform, strengthen its infrastructure” — is, in part, an answer to exactly that question.
There was an almost identical script eight years ago
If this all sounds familiar, that is because it has happened before.
On June 4, 2018, Microsoft announced it was buying GitHub for $7.5 billion in stock. The open-source reaction was fierce, and the concerns map almost line for line onto today’s: a large company with its own commercial interests acquiring the neutral infrastructure an entire ecosystem depends on. Would it favour its own products? Use the data for something else? Tighten the free tier? A wave of developers did move projects to GitLab.
Microsoft’s response was strikingly similar to NVIDIA’s now: GitHub would retain its “developer-first ethos” and “operate independently to provide an open platform for all developers in all industries”. The deal closed on October 25, 2018, after the European Commission cleared it on October 19 following a Phase I investigation, and open-source veteran Nat Friedman became GitHub’s CEO.
Eight years later it is widely regarded as one of the decade’s best tech acquisitions, and GitHub did not become a closed Microsoft tool. But there is an important caveat: the fears were not unfounded so much as misdirected. GitHub went on to become the training and distribution base for Copilot, and “training commercial models on the public code hosted here” became the centre of years of licensing disputes. Microsoft did not do the thing developers feared (lock-in, favouritism); it did a thing nobody had thought to fear.
The lesson for Hugging Face is direct: the real risk is usually not the one you are currently worried about. Huang’s promise about compute will most likely be kept. What deserves attention is the question that surfaces three years from now and does not yet have a name — what a company that owns 3 million models and 500,000 datasets can do with them.
Will regulators block it
Using GitHub as the reference point, the European Commission cleared that deal at Phase I without a deeper probe.
But 2018 and now are very different environments. NVIDIA’s position in AI accelerators, and regulators’ recent posture toward tech acquisitions generally, are both far less relaxed than they were. Scrutiny will likely converge on one question: does a company with a dominant hardware position, acquiring the primary channel through which developers discover and deploy models, gain leverage sufficient to exclude competitors?
Seen that way, Huang’s “NVIDIA compute will not be required” line is not merely public relations — it reads almost as a pre-written answer to the regulators. NVIDIA’s announcement gives no expected closing date, and review progress is likely to be what determines the timeline.
What this means for you
Very little changes in the short term — the platform still works, weights still download, APIs still respond. Three things are worth holding onto:
- Watch behaviour, not promises. If features appear that offer an extra optimisation path on NVIDIA hardware, that is not inherently bad. What matters is whether non-NVIDIA paths get the same level of support at the same time.
- Keep your own copies of weights that matter. This is not about NVIDIA; it is what you should always do with a critical single-platform dependency. The acquisition is just the reminder that you have not.
- Watch the commercial terms. Hugging Face’s free tier is currently generous. If the new owner intends to “strengthen infrastructure”, where the money comes from is a fair question — and the usual answer is tightening free allowances for heavy users.
A company that sells shovels has bought the marketplace where the prospectors gather. The new owner says you are still welcome to use someone else’s shovel. There is no particular reason to disbelieve that today — and nothing guaranteeing it either.
