Nvidia Buys Model Hub Hugging Face: Its Openness Promise Is Unwritten

On Thursday, September 3, Nvidia announced an agreement to acquire Hugging Face. The disclosed figure is $12.93 billion. Hugging Face is not a chipmaker and not a cloud provider. It is a catalogue where open-weight AI models are uploaded, searched and downloaded. It hosts more than 3 million models, more than 500,000 datasets and more than 1 million applications; over 200,000 companies and 18 million developers use it.

Against that user base sits an annualized revenue run rate of roughly $150 million. Nvidia reported $96.2 billion in revenue for the quarter that ended July 26. What Hugging Face earns in a year is about 0.16% of a single Nvidia quarter.

Put the figures side by side and the first question asks itself: why pay $12.93 billion for a company that generates $150 million in revenue? The answer is not in Hugging Face's income statement. It is in Nvidia's.

What the Money Actually Buys

Hugging Face was founded in 2016 by French founders as a chat application. It arrived at its present business through the open-source libraries that let developers share models. In 2023 it raised $235 million in a round led by Salesforce at a $4.5 billion valuation. Nvidia was among the investors in that round.

A 2.9x increase in three years. Revenue grew over the same period, but not by that much. What closed the valuation gap was not the product but the position: Hugging Face became the default front door for open models.

The $12.93 billion is also not a single piece. According to InfoWorld, roughly $11.9 billion is the acquisition price and the remaining $1 billion or so is equity-based retention for the company's roughly 750 employees. About one-thirteenth of the headline number goes not to shareholders but to the engineers Nvidia wants to keep.

The Mechanism: 86 Years or 15 Days

The plainest way to understand an acquisition is to ask which revenue stream pays the price back. Here there are two separate calculations, and they are very far apart.

What this arithmetic says is that the asset Nvidia bought does not need to pay for itself out of its own revenue. Over the life of the deal, Nvidia only needs to add about fifteen days' worth to its own hardware revenue. The moment a developer selects a model is also, in practice, the moment the hardware that model will run on is decided. That is the moment Nvidia paid for.

The scale comparison makes the same point again: $12.93 billion is 21.7% of the $59.7 billion in net income Nvidia reported last quarter. The company covers the price with roughly twenty days of net profit. Per developer on the platform, the price works out to $718.

What the Promise Covers, and What It Does Not

Nvidia committed alongside the announcement to keeping the platform open. Jensen Huang's sentence is specific:

"NVIDIA compute will not be required to build on or deploy through Hugging Face."

Two separate questions need separating here. First: what exactly was said? What was said is that Nvidia compute will not be required. Second: is there an instrument that makes that binding? As of today, no. There is no competition-authority decision, no consent agreement, no commitment in force — there is a corporate blog post and a chief executive's statement.

Sanchit Vir Gogia of Greyhound Research draws the distinction directly:

"Nvidia's openness commitment is precise where it is cheap, and silent where it is expensive."

The gap Gogia points to is concrete: the commitment governs requirement, not ranking. Where a model appears in search results, which runtime is suggested by default, which hardware the one-click configuration targets — none of that falls inside the promise. Direction can be applied without blocking anyone and without removing a single model, purely by arranging defaults.

Nithya Ruff, chair of the Linux Foundation board, made the same point in the language of time: neutrality is a discipline a company has to choose again and again, not a promise it makes once.

This Time There Will Be a Filing

The least-covered aspect of the deal is its legal form. Nvidia has done three large transactions in the past nine months and structured none of them as acquisitions.

In the Groq transaction Nvidia did not buy the company; it took a non-exclusive license to the inference-chip technology and hired key staff including the chief executive. The Poolside structure was a $6 billion non-exclusive license plus a $1 billion minority investment at a $12 billion pre-money valuation; 109 employees moved to Nvidia and the three founders stayed with the company. Enfabrica followed a similar shape.

The common result of these structures is staying outside the merger-notification threshold. In the United States, the Hart-Scott-Rodino rule requires advance notification and a waiting period for acquisitions above a certain size; the FTC's 2026 threshold, announced in January, is $133.9 million. A transaction built as a license plus a minority stake does not meet that definition. The three deals total roughly $28 billion, and no merger file was opened for any of them.

Senators Elizabeth Warren and Richard Blumenthal wrote to Jensen Huang about this on March 23, 2026. The letter argued that by licensing the technology and hiring the most important employees, Nvidia had effectively acquired Groq in all but name, and that with roughly 90% of the GPU market by the senators' description, the deal narrowed competition; they asked for a response by April 3. FTC Chair Andrew Ferguson had said in January that the agency was examining whether such structures were being built to escape the notification requirement.

The Hugging Face transaction sits outside that sequence. A direct $12.93 billion acquisition is far above the threshold and cannot avoid the filing. Justin Boitano, of Nvidia's enterprise computing unit, argued regulators will view the deal favorably because open-source platforms are by definition deconcentrating. The opposing case starts from the same sentence: the owner of the deconcentrating layer would be the concentration itself.

The Same Weeks, a Second Layer

Hugging Face is not an isolated event. In mid-August, Stripe acquired OpenRouter. OpenRouter is not a model developer either: it is a gateway that routes requests across models from different providers. Reported prices vary — Bloomberg and TechCrunch put it above $7 billion, while a figure attributed to Reuters puts it a little above $8 billion.

The result: within a few weeks, the two neutral layers of the open AI ecosystem — the catalogue where the model sits and the gateway where the request is routed — each acquired an owner. The models stay open. The infrastructure around them does not have to.

What the Market Did

The instructive data here is not the direction of the move but its size. On September 3 the S&P 500 rose 0.48% and the Nasdaq 0.84%. Nvidia was among the gainers, but the session's decisive moves were elsewhere: Snowflake rose 21.9% after earnings, and Broadcom fell 6.4% because its revenue guidance came in below expectations.

For a company with $96.2 billion of quarterly revenue guiding to $108 billion for the next one, a $12.93 billion acquisition is a line item that will not show up in the income statement. The same holds at the scale of its market capitalization. The market not pricing this news does not mean the news is unimportant; it means there is no near-term effect on earnings. Those are two different things.

Two Sides

What Is Left

Three concrete things to watch. First, the closing: the deal is expected to complete in the first half of 2027, which means the US waiting period and any second request, plus a Phase I and possible Phase II review in Europe, all fit inside that calendar. Second, what changes in the platform's defaults after closing: the ordering of model cards, the suggested runtime, which hardware comes first in prepared configurations. Third, how enterprises behave: whether they mirror models into their own registries — that is, whether they separate discovery from custody.

This article draws on Nvidia's announcement on its corporate blog, the company's results release for the quarter ended July 26 as filed with the SEC, InfoWorld's reporting on the deal structure, sector analysis from SiliconANGLE and VentureBeat, the letter text published by Senator Warren's office, and the FTC's announcement of the 2026 Hart-Scott-Rodino thresholds. Hugging Face's revenue run rate comes from Dealroom; the split between the $11.9 billion price and roughly $1 billion in retention equity comes from InfoWorld; the price range for the OpenRouter transaction comes from Bloomberg, TechCrunch and reporting attributed to Reuters. The deal has not closed and remains subject to regulatory review.