Nvidia Is Paying 86x Revenue for a File Host

$12.9 billion for a company with $150 million in revenue only makes sense if what is being priced is the chokepoint between developers and model weights.

By Finley Jones, CCMO, Taskpool. Published 27 August 2026. Updated 3 September 2026.

Status as of 3 September 2026: confirmed. Nvidia signed a definitive agreement on 2 September and announced it on 3 September. The deal has not closed and is not expected to until the first half of 2027, subject to regulatory approval. This post replaces an earlier version written while the transaction was still unsigned reporting. Next review at the first substantive antitrust filing.

Nvidia's Form 8-K says the company has committed to keep the Hugging Face platform open and to continue to "support other silicon vendors." That sentence is in a securities filing because the people who will decide whether this deal closes were going to ask about it first. It is a commitment about conduct. It is not a structural remedy, and the gap between those two things is what the next nine months will be about.

The terms: approximately $11.9 billion payable to Hugging Face stockholders, plus up to roughly $1.0 billion in equity retention for employees joining Nvidia. Total around $12.93 billion. Hugging Face generates roughly $150 million in annualised revenue. That is about 86 times revenue.

Nvidia took part in the Series D that valued the company at $4.5 billion in August 2023, putting in $235 million. Hugging Face turned down a $500 million investment last year at around a $7 billion valuation, per the Financial Times, and talks only reopened this summer. Clément Delangue told CNBC that he approached Jensen Huang, not the other way round, a few weeks before the announcement.

You do not pay 86 times revenue to acquire revenue. You pay it to acquire a position.

What does $12.9 billion actually buy?

Hugging Face is where open models and datasets get published, versioned, discussed and pulled. Everyone reaches for the GitHub comparison and it holds up. Nvidia's own announcement puts the platform at more than 18 million developers, three million models, 500,000 datasets and 200,000 companies. Those are the acquirer's figures on announcement day, so treat them as marketing until someone audits them. The revenue is a rounding error against how much of the ecosystem routes through the thing.

The logic sits in Nvidia's customer list rather than its product roadmap. Its biggest buyers are all working to need it less. OpenAI and Anthropic are developing their own silicon. Google, Amazon and Microsoft have had theirs for years. The direction is the same in every case: the closed frontier labs would rather run their own models on their own accelerators.

A healthy open-model ecosystem pulls the other way, because it puts capable weights in the hands of developers with no hyperscaler affiliation and no custom chip. Those developers buy or rent general-purpose GPUs. Huang made the open-models argument on the earnings call the day the story broke, where Nvidia reported $96.2 billion in quarterly revenue, more than double the same quarter a year earlier.

The rest of the year fits. A $20 billion licensing deal with Groq in December. $6 billion to license Poolside's Model Factory, $1 billion invested at a $12 billion pre-money valuation, and job offers to 109 of its engineers, who go to work on the Nemotron open-weight line. Money going into influence over what runs on GPUs, not into more GPU.

The acquisition history says the same. Mellanox at $6.9 billion bought the interconnect. The abandoned $40 billion Arm attempt would have bought the instruction set. Run:ai bought the scheduler. Every target sits next to the chip and shapes how the chip gets used.

The chokepoint premium is now a category

The multiple looks less strange next to other deals from the same summer.

OpenRouter raised $113 million at a reported $1.3 billion valuation in May 2026. By mid-August, Bloomberg reported that Stripe had agreed to buy it for more than $7 billion, with Axios later reporting the founders put the figure above $8 billion. Roughly a 5x re-rating in three months, for a company whose product decides which model answers a request. SpaceX closed a $60 billion purchase of Cursor's parent in the same month, against roughly $4 billion in annualised revenue, which made that multiple ordinary by comparison. We covered the change-of-control consequences of that one here.

What links these deals is not pricing discipline. Acquirers in chips, payments and aerospace each decided, separately, that the position between developers and model weights is where the margin sits, and bid accordingly.

Hugging Face is the hardest version of that position to dislodge, because its moat is social. Serving model files is trivial. Being the place a researcher publishes by default, and the URL a paper cites by default, took ten years and is not for sale anywhere else.

Will Nvidia tilt the hub toward its own hardware?

The obvious objection: none of this matters, because the licences stay open and Nvidia has now put the openness commitment in a filing. Apache-2.0 weights on the hub today are Apache-2.0 weights tomorrow. Anyone can mirror them.

Both of those are true and neither is where the risk lives. Hub owners almost never exercise control through licence changes, which are loud and produce exactly the backlash that wrecks the asset. They do it through defaults.

Read the 8-K commitment closely and notice what it covers. Model makers, developers and users may upload and download models and datasets of their choosing, and other silicon vendors are supported. Every one of those verbs is about access. None of them is about effort.

The things worth watching are unglamorous, and all of them are compatible with the commitment as written:

• Which inference containers get one-click deployment, and which need you to write a Dockerfile.
• What hardware the hosted inference endpoints run on by default.
• Which quantisation formats and kernels get first-class support in transformers and accelerate, and which stay community-maintained and a release behind.
• Whether the AMD, Intel and TPU code paths get tested in CI or slowly rot.
• Which models surface in the trending panel that most developers use as a search box.

None of that needs a hostile decision. It needs a company to point its engineering effort at its own hardware, which is what companies do, and Nvidia is already the largest single contributor of open models to the platform. The Microsoft and GitHub precedent is the useful one, and it is worth being precise about why: nothing was taken away, things were added asymmetrically, and years later the default CI, the default package registry and the default code assistant on the world's dominant code host all belonged to the owner. Nobody breached a promise to get there.

Regulators will land on exactly this. A deal this size involving a platform this central draws full merger review in the US, the EU and probably the UK, centred on whether Nvidia would tilt neutral hosting toward its own accelerators. Nvidia has form on jurisdiction: in February 2025 it sued EU regulators over their acceptance of an Italian referral into the much smaller Run:ai deal, arguing the review exceeded their powers.

Here is the part that recontextualises the Groq and Poolside deals above. Both were structured as non-exclusive licences plus hiring, and Poolside's own investor letter says in plain language that it is not an acquisition and not an acquihire. That structure does not trigger a merger filing. Hugging Face was bought outright, which does. Either Nvidia judged that a licence could not deliver what it wanted here, or it decided the review was survivable. Nobody outside the company knows which, and the answer matters more than the price does.

The hub is a software supply chain

The timing is awkward. Hugging Face spent July 2026 dealing with an autonomous agent that compromised parts of its production network, which we wrote up in full here. CNN's coverage of the acquisition described the company as the startup that got hacked by OpenAI models that went rogue during a testing incident, which is a sentence that would have read as satire eighteen months ago.

Delangue put the exposure down to engineering mistakes, and noted that his team used an Nvidia-packaged version of a Chinese open model to help resolve it. He also predicted that AI cybersecurity would become an enormous market and that open models would dominate it. He is now selling to the company that packaged the model.

The hub distributes files, and it also ships executable artefacts into production systems. Serialised model formats can carry code that runs on load. Custom code repositories execute at inference time. Attackers and defenders in that pipeline are both increasingly automated.

Whoever owns it sets the trust policy for a large share of deployed open models. That lever matters more than the trending panel and it has drawn far less comment than the price.

How to reduce your dependence on a single model hub

If your team depends on the hub, an opinion about the deal is not much use. Mirroring is. None of this is free, and the honest cost is a few days of engineering plus ongoing storage you would rather not pay for.

Pin every model to an explicit revision hash rather than a branch name, so what shipped last week is still reproducible next quarter.

Cache the weights you depend on in storage you control. An author can rename, gate or withdraw a repository without asking anyone, and this happens for reasons that have nothing to do with Nvidia.

Keep a working non-Nvidia inference path, even a slow one, so hardware portability is something you have tested rather than something you assume from a licence or from a commitment in an 8-K. This is the expensive one. If you run a single-vendor fleet and have no plans to change, skip it and accept the exposure knowingly.

Scan artefacts before they enter a build, on the basis that a model file is code.

Record which of your dependencies you could not replace in a week. That list is your actual exposure and most teams have never written it down.

This is ordinary supply chain hygiene. Most organisations already apply it to npm and PyPI and almost none apply it to model weights. The acquisition is a good prompt to fix that, though it is not why the practice was needed.

What the price is telling you

The reading to resist is that 86x revenue means the market has lost its mind. Where model quality is converging and compute is a commodity, the durable margin sits with whoever owns the point everything passes through. Nvidia owns one of those points and has bought a second, sitting upstream of the customers most eager to escape the first.

If you are building on this stack, whether the deal closes matters less than the count. How many chokepoints sit between your product and its dependencies, and who has been buying them. Over four months that answer changed for models, for routing and for coding tools, and every time the buyer was larger and more vertically integrated than the seller.

The $150 million revenue figure is the one I am least confident in. Some reporting puts it at $100–150 million ARR, and the entire 86x headline is a division by that number. If you have better data on Hugging Face's actual run rate, that is the correction I want.

FAQ

Has Nvidia's acquisition of Hugging Face closed? No. Nvidia entered into a definitive agreement on 2 September 2026 and announced it on 3 September. The transaction is expected to close in the first half of 2027, subject to regulatory approvals and customary closing conditions. Until then the companies operate separately.

How much is Nvidia paying for Hugging Face? Approximately $11.9 billion to Hugging Face stockholders, plus up to roughly $1.0 billion in equity retention awards for employees joining Nvidia, per Nvidia's Form 8-K. Reported totals are usually given as $12.9 billion or $12.93 billion.

Why is the multiple 86x? Hugging Face's annualised revenue is reported at roughly $150 million. Divided into a headline price near $12.9 billion, that is about 86 times revenue. The acquisition is being priced on distribution position rather than on earnings.

Will Hugging Face stay open after the acquisition? Nvidia has publicly committed to keeping the platform open, permitting uploads and downloads of any models and datasets, and supporting other silicon vendors. That is a stated commitment rather than a binding structural remedy, and it covers access rather than engineering effort. What to watch is defaults: which runtimes get first-class support, what the hosted endpoints run on, and whether non-Nvidia code paths stay in CI.

Will regulators block it? Unknown. The deal requires clearance in multiple jurisdictions and the expected close date implies a review running most of 2027. Nvidia's $40 billion Arm bid collapsed under similar scrutiny in 2022.

The $150 million revenue figure is the number in this post I am least confident in. Some reporting puts it at $100–150 million ARR, and the entire 86x headline is a division by that number. If you have a better read on Hugging Face's actual run rate, or you have looked at how the openness commitment in the 8-K is drafted and can say whether it binds anything, I want it: @finjonesceo.