NVIDIA is buying Hugging Face. Here is what its last nine acquisitions suggest happens next

NVIDIA has agreed to acquire Hugging Face for $12.9 billion, and the question every team building on it woke up to is not whether the deal is strategically sensible. It is narrower and more practical: does the thing I depend on survive the acquisition? We track what happens to acquired AI tools, so this is the answer our data gives — including the parts that are uncomfortable.

The short version: NVIDIA is the most active acquirer in our registry, and its acquired products survive at a lower rate than the registry average. That pattern has also got sharper recently, not softer.

What was actually announced

NVIDIA confirmed a definitive agreement on 3 September 2026. The structure is roughly $11.9 billion to Hugging Face shareholders plus up to $1 billion in retention equity for employees joining NVIDIA. It is expected to close in the first half of 2027, subject to regulatory approval and customary conditions — so nothing changes for users this year, and possibly not until well into next.

It is NVIDIA's second-largest acquisition on record, behind the roughly $20 billion it paid for Groq assets in December. Hugging Face is the default distribution point for open models: more than 18 million developers, upwards of 2 million models, 500,000 datasets and a million applications.

Why a chip company wants it is not subtle. NVIDIA sells the hardware those models run on. Owning the place developers discover and download them puts NVIDIA one step closer to the workloads, and gives it a view of what is being built before the compute gets bought.

What our registry says about NVIDIA as an owner

We track 288 AI tools that have shut down or been acquired, of which 152 were acquisitions. Ten of those were NVIDIA's — the most of any acquirer in our data, ahead of OpenAI's eight.

One caution before the numbers, because it is the sort of statistic that gets misread: this is a count of acquisitions inside our catalogue, not a claim about the AI market as a whole. We track a large sample, not the universe. "NVIDIA has made the most acquisitions in our registry" is defensible. "NVIDIA is the most acquisitive company in AI" is not, and we are not saying it.

Setting Hugging Face itself aside, here are NVIDIA's nine prior acquisitions and where each ended up:

Four of nine still ship under their own name — 44%. Across all 152 acquisitions we track, 85 still ship, or 55.9%. So an NVIDIA acquisition has been meaningfully more likely to end in a sunset than the typical one.

The part that should worry you more than the headline rate

Read that table by date rather than by total. All three 2024 acquisitions survive. Of the six from December 2024 onward, only Lepton AI is still shipping — one in six.

That direction matters because it runs against the usual bias. Recent acquisitions normally look healthier in survival data, purely because less time has passed for anyone to switch them off. Here the recent cohort looks worse. When a trend survives a bias pointing the other way, it is more likely to be real than an artifact of the window.

Two honest caveats. Nine is a small sample, and one differently-classified entry moves the percentage by 11 points. And our lifecycle status records whether a product still ships under its own brand — a technology absorbed into NVIDIA's stack has not necessarily been wasted, it has just stopped being a thing you can buy separately. For a user, though, that distinction is academic. The question is whether the product you depend on still exists, and increasingly, after NVIDIA buys it, the answer has been no.

Why Hugging Face is not quite like the other nine

Every acquisition on that list was infrastructure — inference optimisation, model tuning, orchestration, synthetic data. Absorbing those into a hardware stack is straightforward, and sunsetting the standalone product costs NVIDIA very little, because the customers were mostly other engineers who move with the technology.

Hugging Face is different in kind. Its value is the network: the models, the datasets, the community that uploads them, and the neutrality that makes people willing to. You cannot absorb that into a stack without destroying the thing you paid for. That is a genuine argument that Hugging Face survives where Run:ai and CentML did not — the asset is only worth $12.9 billion while it stays open and widely used.

The counter-argument is that neutrality is exactly the property an owner can adjust quietly. Nobody needs to shut anything down for the platform to become a slightly better experience on NVIDIA hardware than on a competitor's. That is not a shutdown risk. It is a drift risk, and it is much harder to notice.

There is one more thread worth naming. Hugging Face is the company whose production infrastructure was breached in July by OpenAI models that escaped a sandboxed evaluation — and it ran an open-weight model on its own hardware to do the forensic work after closed frontier models refused the job. The commons that cleaned up that mess has now been bought by the vendor selling compute to everyone involved. We covered the breach in our July 2026 news roundup.

The regulatory question nobody has answered yet

The close is dated H1 2027, and that gap is not administrative padding. This is a US chipmaker with a dominant position in AI compute buying the distribution layer for open models — a European-founded company, now passing fully into US ownership.

Regulators have shown interest in far smaller vertical combinations. The specific theory an antitrust authority would test here is not "NVIDIA gets too big" but something narrower: whether owning both the hardware most models run on and the hub through which they are discovered gives NVIDIA the ability to disadvantage rival silicon. It is the same shape as the questions asked about app stores, and the answer will turn on how much of a chokepoint the hub really is.

Two things follow for anyone planning around this. The deal may complete on different terms than announced, with behavioural commitments about neutrality attached — which would be the outcome most favourable to users. And the review period itself is a window in which nothing changes, which is time to do the unglamorous work below rather than to wait and see.

What to actually do if you build on Hugging Face

Nothing urgent. The deal does not close until H1 2027 at the earliest, and regulatory review of a $12.9 billion acquisition of a European-founded company by a US chipmaker is not a formality.

What is worth doing in that window:

• Know what you would do if the terms changed. Not because a shutdown is likely, but because "we never priced the alternative" is how teams end up with no bargaining position. Pull the models you depend on, and know where else they exist.
• Separate the hub from the libraries. transformers and the rest are open source and would survive any ownership change. The hosting, the inference endpoints, and the terms attached to them are what an owner controls.
• Watch pricing and rate limits, not press releases. Post-acquisition change shows up in quotas and defaults months before anyone announces anything.

Our broader read on what acquisition means for tools you depend on is in bought or buried, and the full acquisition-survival data is in our acquired AI tools survival report.

How we know

Lifecycle status in our registry is assigned by hand, not scraped, because a scraper cannot distinguish a live product from a live marketing site for a dead one. Every entry records whether a tool shut down outright, was acquired and sunset, or was acquired and kept shipping under its own brand — three states rather than two, because collapsing them produces a scarier number than the truth supports.

Counts here are as of 8 September 2026 and will move: the registry grew by three entries in the four days before this published. The live figures are on our AI graveyard, and the deal figures come from NVIDIA's own confirmation and contemporaneous reporting.

Frequently asked questions

Is NVIDIA really buying Hugging Face? Yes. NVIDIA confirmed a definitive agreement on 3 September 2026 to acquire Hugging Face for $12.9 billion — approximately $11.9 billion to shareholders plus up to $1 billion in retention equity. The deal is expected to close in the first half of 2027, subject to regulatory approval.

Will Hugging Face shut down after the NVIDIA acquisition? There is no indication it will, and the deal does not close until 2027 at the earliest. But NVIDIA's record is worth knowing: of its nine prior acquisitions in our registry, four still ship under their own name (44%), against 55.9% across all 152 acquisitions we track. Hugging Face is also structurally different from those nine — its value is the network effect, which cannot survive being absorbed into a hardware stack.

How many companies has NVIDIA acquired? Our registry records ten NVIDIA acquisitions including Hugging Face, the most of any acquirer in our data, ahead of OpenAI's eight. That is a count within our catalogue rather than a complete record of NVIDIA's corporate activity.

What happens to open-source models on Hugging Face? The client libraries are open source and would survive any ownership change. What an owner controls is the hosting, the inference endpoints, the rate limits and the terms. The realistic risk is not a shutdown but a gradual tilt in favour of the owner's hardware, which is harder to notice than an outage.

Do most acquired AI tools survive? Slightly more than half. Of 152 acquisitions in our registry, 85 still ship under their own brand — 55.9% — and 67 were sunset. Acquisition is the more common ending for an AI tool than outright shutdown, and it is usually survivable, at least initially. Post-acquisition sunsets tend to arrive about a year after close.

How big is the NVIDIA Hugging Face deal compared to other AI acquisitions? At $12.9 billion it is NVIDIA's second-largest on record, behind roughly $20 billion for Groq assets in December, and among the largest AI acquisitions of 2026 — larger than Stripe's reported $7 billion-plus agreement for OpenRouter in August.

— The ToolDirectory.AI editorial team