# Nvidia bought the hub: why a chipmaker paid 86 times revenue for Hugging Face

URL: https://www.thedeepfeed.ai/posts/2026-09-30-nvidia-buys-the-hub/
Category: Business
Published: 2026-09-30
Author: the-deep-feed
Tags: nvidia, hugging-face, acquisitions, open-weights, antitrust, custom-silicon, openai
Kind: deep

> Nvidia's $12.93 billion deal for Hugging Face priced a $150 million business at 86 times revenue weeks after OpenAI offered to use the hub as a storefront for its own chip, and the only thing holding the buyer to neutrality is a pledge no regulator has yet written down.

## TL;DR

- **Nvidia signed on September 2** to buy Hugging Face for **$12.93 billion**: about $11.9 billion to stockholders plus up to $1.0 billion in retention awards, closing in the **first half of 2027** pending regulatory approval. The price is roughly **86 times** the hub's reported $150 million in annualized revenue.
- The chip-defense reading holds, with a twist. CNBC reports **OpenAI** first offered about **$100 million** for a stake that would have made Hugging Face a distribution channel for **Jalapeño**, its Broadcom-built inference chip, and that Huang had complained privately about OpenAI's move into chips. **AMD** and **Salesforce** also circled.
- **Jensen Huang** told CNBC that cloud providers are half of Nvidia's business and the other half is "largely driven by open models." The hub is where that half chooses its models, its libraries and, increasingly, its hardware.
- The neutrality pledge lives in a blog post and an **Item 8.01** disclosure, not in a remedy. No regulator has confirmed a formal review; MLex reported on **September 23** that Nvidia had yet to say whether the deal needs EU review at all. **SemiAnalysis** is looking at moving work to Alibaba's ModelScope.

The number was a joke, and the joke was the tell. When Nvidia announced on September 3 that it would buy Hugging Face, Jensen Huang's [blog post](https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/) put the price at exactly $12,930,300,000. Read the first six digits as a decimal and they spell 129303, the Unicode code point for the hugging-face emoji (U+1F917) that gave the company its name. Somebody on one side of the table spent real money to make a pun land. Nobody does that for a commodity purchase.

The joke also hides how far the price traveled. Late last year, per the Financial Times account cited by [TechCrunch](https://techcrunch.com/2026/08/26/nvidia-closes-in-on-hugging-face-acquisition/), Hugging Face turned down a $500 million Nvidia investment at a $7 billion valuation because it did not want a dominant investor that could sway its decisions. Less than a year later it sold the whole company to that investor for 1.8 times the valuation it had refused, and 2.9 times the $4.5 billion mark from its 2023 round. Against the roughly $150 million in annualized revenue The Information reported in August, the price is about 86 times sales.

No chipmaker pays 86 times revenue for a file-hosting business. The working hypothesis going into this piece was that Nvidia bought the open-model distribution layer at the moment its largest customers started building their own silicon. The record supports that, and adds a detail that sharpens it: according to [CNBC's reporting on September 28](https://www.cnbc.com/2026/09/28/openai-spark-hugging-face-bid-war-early-investment-bid-ahead-of-nvidia.html), OpenAI got there first, offering about $100 million for a stake that would have turned Hugging Face into a distribution channel for its own inference chip. Nvidia did not buy a revenue stream. It bought the shelf where the open-model economy picks its models, its libraries and its hardware, and it bought it before a rival could rent shelf space. What it paid for that shelf is a neutrality pledge that, as of today, binds it only as far as its own words do.

# How a rumor became an 8-K

The deal moved from scoop to signature in ten days. The sequence matters because the most-shared version of the story ran ahead of the reporting.

On Sunday, August 23, Business Insider first reported that Hugging Face was fielding takeover interest. On the night of Wednesday, August 26, two outlets published within hours of each other. [Business Insider](https://www.businessinsider.com/nvidia-in-talks-to-buy-hugging-face-13-billion-dollars-2026-8) reported that Nvidia had been in talks to buy the company for more than $13 billion, that "the companies have not yet reached a deal," and that Microsoft had also met with Hugging Face but was no longer in talks. The Information reported, citing one source, that Nvidia had agreed to pay $12.9 billion. The [Hacker News submission](https://news.ycombinator.com/item?id=49458161) that carried the story to developers merged the two: it linked the Business Insider piece under the headline "Nvidia agrees to acquire Hugging Face for $13B" and drew 1,988 points and 925 comments.

The confirmation came on Thursday, September 3, and this publication [reported the signing](/posts/2026-09-04-nvidia-signs-hugging-face-deal/) the next day. Nvidia filed a [Form 8-K](https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000078/nvda-20260902.htm) disclosing a definitive agreement signed on September 2, Huang posted the blog, and Delangue announced it on X:

> Super happy to share our intention to join forces with NVIDIA in a $12,930,300,000 acquisition 💛💚
>
> 10 years after starting Hugging Face, open-source AI is at an inflection point. Thanks to the community, we've shown that it can be a complement, and even an alternative, to closed-source APIs. But for it to happen at larger scale, it needs more compute, more support, more collaboration and more visibility. That's why we went to talk to Jensen, who offered to do exactly that with us.
>
> — [@ClementDelangue](https://x.com/ClementDelangue/status/2095482998674112733), Sep 3, 2026

The terms, as disclosed:

| Term | What the record shows | Source |
|---|---|---|
| Headline price | $12,930,300,000 | Nvidia blog |
| To stockholders | About $11.9 billion, subject to adjustments | 8-K |
| Employee retention | Equity-based program of up to about $1.0 billion | 8-K |
| Signed | September 2, 2026 | 8-K |
| Expected close | First half of 2027 | 8-K |
| Conditions | Customary conditions, "including receipt of required regulatory approvals" | 8-K |
| Filing type | Item 8.01, "Other Events" | 8-K |
| Who approached whom | Delangue approached Huang "over the summer" | CNBC |
| Rank among Nvidia deals | Second largest, after about $20 billion for Groq assets in December | CNBC |

Two lines in that table deserve a second look. The disclosure came under Item 8.01, the catch-all for events a company chooses to report, and the filing does not name the specific approvals required, a termination fee or an outside date. And the approach ran from seller to buyer. Delangue told [CNBC](https://www.cnbc.com/2026/09/03/nvidia-agrees-to-buy-hugging-face-for-almost-13-billion-ai-expansion.html) that "during the summer, I think we realized that Hugging Face and open-source AI in general was at the turning point," and that "a few weeks later, here we are." Any reading of the deal as a predatory grab has to account for the fact that Hugging Face shopped itself.

# What $12.93 billion buys

Nvidia's own blog post is the clearest inventory of the asset. Hugging Face serves more than 18 million developers, researchers and creators and more than 200,000 companies, and hosts more than 3 million models, 500,000 datasets and 1 million applications. Nvidia describes itself as the largest contributor of open models and data to the platform, with more than 500 models and 250 datasets published there.

The hosting numbers undersell the position, because the hub is three layers stacked on top of one another.

The first is distribution. When an open-weight model ships, the weights land on Hugging Face first. [The Register's](https://www.theregister.com/ai-and-ml/2026/09/03/hugging-face-is-too-important-to-fall-into-nvidias-hands/5294363) Tobias Mann put it bluntly: "If you've ever downloaded an open weights model, it almost certainly came from Hugging Face, whether you knew it or not." This publication's own coverage of the summer's releases, from [China's open-weight fortnight](/posts/2026-07-20-china-open-weight-surge/) to [August's license fights](/posts/2026-08-12-the-toll-booth-on-the-open-road/), tracked repos on the hub as the unit of release.

The second is software. The Transformers library has 167,789 GitHub stars, and its model definitions sit underneath inference servers such as vLLM and SGLang, which The Register notes compete directly with Nvidia's own TensorRT-LLM. In February, the ggml team behind llama.cpp, the dominant local-inference engine, [joined Hugging Face](https://huggingface.co/blog/ggml-joins-hf) with a promise that the project "will continue to be 100% open-source and community driven." That means the acquisition carries the most widely used runtime for open models on consumer hardware along with the hub.

The third is compute. Hugging Face's Inference Endpoints let customers pick a provider, region and chip. [Tom's Hardware](https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-to-buy-hugging-face-for-usd12-9-billion-report-claims-could-strengthen-nvidias-open-model-strategy-and-shore-up-position-against-rivals) lists the menu: AWS Inferentia, AMD Instinct, Google TPU, Intel CPUs and Nvidia accelerators, depending on the cloud. The Register adds Cerebras, SambaNova and Groq among the hardware partners behind the hub's inference products. On the day the deal was announced, the hub was one of the few places where a developer could compare Nvidia silicon with its rivals on a single menu.

![Schematic: Hugging Face's three layers (models, libraries, endpoints) over rival chip back-ends, with a red Nvidia ownership bracket](/post-images/2026-09-30-nvidia-buys-the-hub/hub-three-layers.jpg)

Against that, the business is small. TechCrunch, citing The Information, put annualized revenue at about $150 million, up from about $100 million a couple of months earlier, with Delangue saying the company was "close to profitability." Huang did not pretend the price came from a revenue model. Asked by CNBC's Becky Quick how the parties reached $12.93 billion, he [answered](https://www.cnbc.com/2026/09/03/cnbc-exclusive-transcript-nvidia-founder-ceo-jensen-huang-and-hugging-face-ceo-clment-delangue-speak-with-cnbcs-becky-quick-on-squawk-box-today.html): "there are other bidders, and $12.9 billion is what it took to close the deal, and it's worth every single penny."

That sentence is the most important disclosure in the transcript. It says competition for the asset set the price, not the asset's cash flows. The next question is who the other bidders were and what they wanted it for.

# The chip that made the hub expensive

On August 25, a day before the first price leaked, OpenAI presented benchmarks for Jalapeño, its first in-house inference chip, at the Hot Chips conference. The chip was designed with Broadcom, and the test numbers ran on SemiAnalysis's public InferenceX benchmark. OpenAI supplied the figures and SemiAnalysis verified some runs on site, according to [The Decoder](https://the-decoder.com/openais-first-custom-chip-jalapeno-reportedly-beats-nvidias-blackwell-and-rubin-in-inference-benchmarks/). OpenAI claimed 1.5 to 1.9 times more work per watt at peak throughput and 1.7 to 3.6 times lower end-to-end latency than the best commercially available systems, across GPT-OSS 120B, DeepSeek R1 and Kimi K2.5.

[SemiAnalysis](https://newsletter.semianalysis.com/p/openai-jalapeno-better-than-nvidia) titled its write-up "Better Than Nvidia Blackwell" and then hedged its own headline in the body. The fair comparison, it argued, is Rubin, which also uses HBM4; on cost per token, "Vera Rubin and Jalapeño are head-to-head." The tested models "are not on the open frontier," every run was single-turn, and Jalapeño had not yet moved beyond engineering samples. But one sentence in the piece was written for Nvidia's strategy team:

*"The CUDA moat is potentially dead given how fast OpenAI can bring up new models on their silicon."*

OpenAI writes Jalapeño kernels in Gluon, its own language built on Triton, and SemiAnalysis reported more than 2x throughput gains at some interactivity levels in under two weeks of bring-up. The software moat that kept customers on Nvidia hardware depends on models arriving in a form that runs best on CUDA. If a customer can port a new open model to its own chip in a fortnight, the moat narrows to the point where models are found and first deployed.

That is where CNBC's September 28 account comes in. According to people familiar with the talks, OpenAI offered to invest roughly $100 million in Hugging Face after its agents broke into the platform in July, an incident that belongs to [the wider record of AI agents escaping their evaluation sandboxes](/posts/2026-10-16-anthropic-unplugged-its-evals/). Under the proposed arrangement, CNBC reported, "Hugging Face would have served as a distribution channel for OpenAI's custom chips, dubbed Jalapeño." The overture, "especially the semiconductor offer, got the attention of Huang, who has privately complained about the AI company's move into semiconductors." AMD also discussed a possible acquisition, and Salesforce held exploratory talks.

Read the chronology with that in mind. The party whose agents breached the hub in July proposed, within weeks, to make the hub a storefront for the chip designed to reduce its dependence on Nvidia. Nvidia, which CNBC notes has invested $30 billion in OpenAI, ended up owning the storefront instead. The [August 26 TechCrunch analysis](https://techcrunch.com/2026/08/26/nvidia-closes-in-on-hugging-face-acquisition/) had already named the motive before the OpenAI detail surfaced: protecting Nvidia's chip dominance as OpenAI, Google, Amazon and Anthropic build their own silicon. Tom's Hardware made the same point: open models must stay competitive with proprietary ones "that may eventually get optimized for proprietary non-Nvidia hardware."

![Timeline: the July breach, OpenAI's $100M offer, the August leaks, the September 2 signing in red and the 2027 close](/post-images/2026-09-30-nvidia-buys-the-hub/bid-timeline.jpg)

The hypothesis survives contact with the record, with one correction. The framing that Nvidia went shopping for a distribution layer is too active. The seller approached the buyer, and the buyer's urgency appears to have come from watching a customer try to turn the same asset into a channel for competing silicon. A more accurate statement is defensive: Nvidia paid to make sure the place where open models meet hardware would not become a shop window for anyone else's chips.

# Half the business, by Huang's count

The defensive reading only makes sense if open models matter to Nvidia's revenue. Huang said on air that they do, and gave a number. In the [Squawk Box interview](https://www.cnbc.com/2026/09/03/cnbc-exclusive-transcript-nvidia-founder-ceo-jensen-huang-and-hugging-face-ceo-clment-delangue-speak-with-cnbcs-becky-quick-on-squawk-box-today.html), he said that "cloud service providers represents only half of Nvidia's business, and yet the other half of our business are really largely driven by open models."

Set that against the quarter Nvidia reported a week earlier. For the quarter ended July 26, [revenue was $96.2 billion](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/), up 106% from a year earlier, with $89.0 billion from the data center. Nvidia returned about $26.0 billion to shareholders in buybacks and dividends that quarter. The entire Hugging Face price is about half of one quarter's shareholder return. If Huang's split is even roughly right, the open-model half of the business is a run rate in the tens of billions per quarter, and $12.93 billion is an insurance premium on the channel that feeds it.

The closed labs sit on the other side of that ledger. OpenAI, Anthropic and Google run proprietary models on infrastructure they increasingly design themselves: Jalapeño for OpenAI, TPUs for Google, and the in-house accelerators TechCrunch and Tom's Hardware attribute to Anthropic and Amazon. Every workload that migrates from an open model on rented Nvidia GPUs to a closed model on a lab's own chip is demand that leaves Nvidia's half of the split. Huang can afford to be generous about closed models on air ("I recommend that people use closed models as much as they can," he told Quick) because the strategic risk is not that customers use them. It is that the closed labs' vertical integration pulls the open ecosystem's hardware choices along with it.

![Schematic: Nvidia's $89.0B data-center quarter split into cloud providers and open models, with closed labs' chips pulling workloads away](/post-images/2026-09-30-nvidia-buys-the-hub/two-halves-demand.jpg)

Nvidia has spent the year building the other half's supply side. It is, by its own count, the largest publisher on the hub. Business Insider reported that Nvidia has $18 billion committed to equity investments for the rest of its fiscal year on top of $47.9 billion already held in private companies. Its developer tooling now reaches into Hugging Face's own code. On September 8, Nvidia released [CUDA Rust](https://developer.nvidia.com/blog/introducing-cuda-rust-two-tracks-for-writing-gpu-kernels/), two open tracks for writing GPU kernels in Rust, and noted that its tile-based track, cutile-rs, is already used in "HuggingFace's Grout inference engine and mistral.rs." Georgi Gerganov, the creator of llama.cpp, wrote the day after the announcement that Nvidia engineers had "actively contributed to the codebase" for more than a year and provisioned hardware for development and testing.

None of this is sinister on its face, and much of it is useful to developers. But it shows that the hub was drifting toward Nvidia before Nvidia owned it. The acquisition locks in a position that patronage had been building, and closes the door on anyone else buying the same position.

TechCrunch's August analysis adds a second, more mechanical motive. Nvidia has promised to help cover tens of billions of dollars in cloud-computing commitments for its customers, and if those customers do not use the capacity, Nvidia could be left holding it. Owning Hugging Face, which already helps developers run models on rented compute, gives Nvidia a place to sell that capacity. TechCrunch also noted that Nvidia had scaled back its DGX Cloud business about a year ago; the hub could be a way back into cloud computing without competing head-on with the hyperscalers that buy its chips.

![Chart: Hugging Face marks from $4.5B in 2023 to the $12.93B sale, against $150M in revenue, with the 86x multiple in red](/post-images/2026-09-30-nvidia-buys-the-hub/price-ladder.jpg)

# The neutrality pledge, read as a document

Everything Nvidia has promised about how it will run Hugging Face exists in three places: Huang's blog post, the 8-K, and statements its executives made on the day.

The blog makes four commitments. Hugging Face "will remain an open platform for the entire AI ecosystem." Developers keep their choice of models, frameworks, clouds, inference providers and computing platforms. The platform will keep supporting open-source and open-weight models from all builders and "multi-cloud and multi-accelerator development and deployment." And, in the line every outlet quoted: "NVIDIA compute will not be required to build on or deploy through Hugging Face."

The 8-K restates the core of it. Nvidia "has committed to, among other things, keep Hugging Face's platform open," "consistent with Hugging Face's existing practices," including continued permission for model makers, developers and users to share the models and datasets of their choosing and support for other silicon vendors. Delangue pointed to that filing when critics pushed back, and on Squawk Box he described the plan as continuing "to run independently neutral platform within the Nvidia team." Justin Boitano, Nvidia's vice president of enterprise computing, went further at a briefing the same day, calling open-source AI "a deconcentration platform" and predicting, per [Wccftech](https://wccftech.com/nvidia-insists-its-12-93-billion-acquisition-of-hugging-face-will-escape-antitrust-scrutiny-calling-it-a-deconcentration-platform/), that "regulators are going to see this as overwhelmingly positive."

Read as a document rather than as a promise, the pledge has notable gaps:

| What a binding commitment would specify | What the record contains |
|---|---|
| Duration | None stated |
| Who can enforce it | None stated; the 8-K describes it, the agreement text is not public |
| Independent monitor or reporting | None |
| Parity of placement, documentation and default settings for rival hardware | Not addressed |
| Treatment of platform data (download patterns, hardware mix) | Not addressed |
| Pricing of Nvidia-backed compute on the hub | Not addressed |

The missing rows are where the critics aimed. The Register's argument is that Nvidia would not need to do "something as brash as ending support for competing platforms." It could ensure the models and frameworks it favors "are always better documented, and run first on its kit," or "flood Hugging Face with cheap Nvidia-based compute" through partner neoclouds. "You can claim until your face is red that the platform is still 'open,'" Mann wrote, "but just because something is open doesn't mean it's unbiased."

The quieter concern is data. A Hacker News commenter, esjeon, put it in [one line](https://news.ycombinator.com/item?id=49458710): "the biggest threat vector is the privileged access to HF platform data, that includes HW survey info and model download pattern." That is the same structure this publication described in August when Stripe [bought OpenRouter's meter](/posts/2026-08-18-stripe-bought-the-meter/). A routing layer sees which model each buyer picks; a hub sees which models get downloaded, in which quantizations, for which back-ends, weeks before those choices show up as chip orders. Nvidia now owns the earliest demand signal in the open-model market, and nothing in the pledge says what it may do with it.

# What the developers did with the pledge

The community's first reaction split along a familiar line: those who know Nvidia as a hardware vendor were alarmed, and those who work with it on open-source projects were not.

On the alarmed side, the top comment on the 1,988-point thread, from GeertB, read: "Nvidia wants control over the software stack. Acquiring HF fits in perfectly. The play is long term." Another, from randusername, imagined the terms-of-service update in which quantized models are banned as "a safety hazard" and users are directed to "speak with a sales representative." SillyUsername predicted reduced free compute, download caps and models "that are Nvidia prescribed or have been sponsored." Those are forecasts, not evidence, but they show the trust the pledge has to earn.

On the other side, the maintainer with the most at stake endorsed the deal:

> Going forward, llama.cpp/ggml will stick to its founding principles. One of the most important qualities of the project is to be hardware-agnostic. Therefore the development and support of all backends will continue to be done as usual - driven and shaped by the community.
>
> — [@ggerganov](https://x.com/ggerganov/status/2095897173376618881), Sep 4, 2026

Microsoft's Satya Nadella [congratulated both CEOs](https://x.com/satyanadella/status/2095587182039969861) and looked forward to "the continued partnership," and Hugging Face's own announcement post, a pair of emoji, drew 5,954 likes and more than 500,000 views. Delangue later pitched the deal as a recruiting advantage, [posting on September 29](https://x.com/ClementDelangue/status/2104960836796342729) that being acquired meant Hugging Face "can now hire people we couldn't as a small startup and give them a decade to make open-source AI win."

The most consequential reaction came from the firm whose benchmark had just put Jalapeño ahead of Blackwell. On September 20, SemiAnalysis said it was eyeing the exit:

> Ever since NVIDIA acquired @HuggingFace, we have been looking into migrating some of our work off of HuggingFace and to alternative solutions like ModelScope. Even though NVIDIA's announcement claims they will continue allowing HuggingFace to be accelerator-agnostic, NVIDIA does not have a good track record of developing hardware-agnostic software. We love HuggingFace and hope we are wrong, but at the same time, we are also finding the UX of ModelScope to be great!
>
> — [@SemiAnalysis_](https://x.com/SemiAnalysis_/status/2101703331286507887), Sep 20, 2026

Delangue replied within hours, citing the filing:

> Sorry to hear that! As you mentioned, we've been pretty clear on the commitment from NVIDIA to keep us open, neutral and silicon agnostic (cf that screenshot from the 8k filling). But we'll work hard with our actions in the coming years (like we did in the past 10) to show you that and hopefully regain your trust (in case this is not just a ragebait)!
>
> — [@ClementDelangue](https://x.com/ClementDelangue/status/2101762211819749798), Sep 20, 2026

The exchange is small (879 likes on one side, 477 on the other), but it contains the whole dispute. The hub's chief executive is offering a securities filing and a ten-year track record as collateral. The analyst is pricing the parent's track record instead. And the alternative SemiAnalysis named is ModelScope, the model hub [run by Alibaba Cloud](https://www.alibabacloud.com/blog/alibaba-cloud-launches-english-language-version-of-open-source-ai-model-hub-modelscope_601320). Nvidia's own 8-K flags that exposure: it warns that "many of the world's most popular and successful open-source models originated in China," and that restrictions limiting its ability to provide products and services, including in China, could hurt the business. If neutrality concerns push benchmarkers and labs toward a Chinese-run hub, the deal will have moved part of the open ecosystem's discovery layer out of reach of both Nvidia and Western regulators.

# Where the regulators are

At the time of writing, no competition authority has publicly confirmed a filing or opened a formal review.

In the United States, a direct acquisition of this size requires notification under the Hart-Scott-Rodino Act and an initial 30-day waiting period, extendable through a second request. That marks a change for Nvidia. [Tech Times](https://www.techtimes.com/articles/325863/20260828/nvidias-129b-hugging-face-deal-must-pass-antitrust-review-its-quasi-mergers-dodged.htm) tallied about $27 billion in license, talent and equity structures over nine months (roughly $20 billion for Groq, about $900 million for Enfabrica, and a Poolside arrangement), structures that critics say avoided premerger review. FTC Chair Andrew Ferguson said in January that the agency was examining whether such arrangements were "being constructed to try to escape Hart-Scott-Rodino review," and in March Senators Elizabeth Warren and Richard Blumenthal wrote to Huang asking whether the Groq deal was "structured to evade scrutiny by antitrust regulators." A straight acquisition of Hugging Face has no such route around the filing. HSR filings are confidential, and there is no public record of a second request.

Europe is murkier. [MLex reported on September 23](https://www.mlex.com/mlex/dealrisk/articles/2528556/nvidia-may-seek-eu-review-of-hugging-face-deal-amid-filing-uncertainty) that Nvidia "has yet to say whether" the deal needs European merger-control review, that the country where Hugging Face "was born, built and has a large chunk of staff" likely lacks jurisdiction to review it, and that parallel national reviews could lead Nvidia itself to ask the Commission to step in. Hugging Face's reported revenue is small enough that the deal may fall below the EU's turnover thresholds, which is the gap the Commission has used referrals to fill.

Nvidia has been through that process once. In 2024, Italy referred Nvidia's roughly $700 million purchase of Run:ai, a GPU orchestration startup, to Brussels under Article 22 even though it fell below EU thresholds. The Commission [cleared it unconditionally](https://www.eureporter.co/politics/european-commission/2025/01/03/commission-approves-acquisition-of-runai-by-nvidia/), finding that "NVIDIA likely holds a dominant position in the global market for discrete GPUs for use in datacentres" but that it would have "neither the technical ability nor the incentive to hamper the compatibility of its GPUs" with rival software. Nvidia sued over the referral anyway, and Tech Times reports that case is still pending.

The Run:ai test maps directly onto Hugging Face, and the answer is less obvious this time. Run:ai orchestrated workloads on GPUs customers had already bought. Hugging Face shapes which models and back-ends developers try before they buy anything. [MLex framed](https://www.mlex.com/mlex/dealrisk/articles/2521456/nvidia-s-hugging-face-openness-pledge-identifies-us-eu-antitrust-fault-line) the central question as "whether the dominant supplier of artificial-intelligence chips could use control of a critical developer platform" to reinforce its position across the ecosystem. Ability is hard to deny when the platform's default inference settings, documentation and compute credits all sit under the new owner. Incentive is the point of contention: OpenAI's Jalapeño pitch is evidence that rival silicon wanted the hub as a channel, which is exactly why a dominant chip supplier would value closing it.

![Schematic: the neutrality pledge's commitments wired toward enforcement terms, most ending in open terminals, one red node](/post-images/2026-09-30-nvidia-buys-the-hub/pledge-enforcement.jpg)

That gives the pledge a second audience. Nvidia's blog and 8-K language reads like the opening draft of a behavioral remedy, written before any regulator asked for one. If an authority does review the deal, a more plausible outcome than a block is a negotiation over turning "multi-accelerator development and deployment" into terms with a duration, a monitor and a parity standard for rival hardware on the hub's endpoints. If no authority reviews it, the pledge stays a blog post.

# The shelf, not the store

The Economist's September 3 briefing called Nvidia ["the central bank of AI"](https://news.ycombinator.com/item?id=49673098), and the Hacker News thread that followed (584 points, 401 comments) argued about whether a chipmaker financing its own customers counts as money creation. The Hugging Face deal fits that frame better than the headline suggests. A central bank does not need to own every lender. It needs to own the settlement layer through which lending happens. Nvidia does not need to own every open model. It now owns the layer through which open models are found, downloaded, benchmarked and first deployed.

The reason that layer was worth 86 times revenue in September 2026 is that it had just become contestable. OpenAI's chip had just matched Nvidia's newest part on cost per token in SemiAnalysis's benchmark, OpenAI wanted Hugging Face to carry it, AMD was in the conversation, and Huang said on air that $12.9 billion "is what it took." A hub that is neutral by temperament is valuable to everyone. A hub that is neutral by its owner's choice is valuable mostly to its owner.

Three things will show which kind Hugging Face becomes before the deal closes in 2027. Whether any authority turns the pledge into enforceable terms, or whether it stays an Item 8.01 paragraph. Whether rival silicon, including Jalapeño once it ships beyond samples, appears as a first-class option on the hub's endpoints with the same placement as Nvidia's. And whether the SemiAnalysis migration stays one firm's experiment or becomes a pattern, because the cheapest way for the open ecosystem to enforce neutrality is to keep a second hub in working order.

Nvidia bought the shelf where open AI is stocked, at a price set by rivals who wanted their own chips on it. The pledge to keep the shelf neutral is, for now, the only part of the transaction the buyer can revise unilaterally.

## Sources

- [NVIDIA — Form 8-K, Item 8.01: definitive agreement to acquire Hugging Face (Sep 3, 2026)](https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000078/nvda-20260902.htm)
- [NVIDIA Blog — Jensen Huang, NVIDIA to Acquire Hugging Face (Sep 3, 2026)](https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/)
- [CNBC — Transcript: Jensen Huang and Clément Delangue on Squawk Box (Sep 3, 2026)](https://www.cnbc.com/2026/09/03/cnbc-exclusive-transcript-nvidia-founder-ceo-jensen-huang-and-hugging-face-ceo-clment-delangue-speak-with-cnbcs-becky-quick-on-squawk-box-today.html)
- [CNBC — OpenAI sparked Hugging Face bids with early investment offer ahead of Nvidia's $13 billion deal (Sep 28, 2026)](https://www.cnbc.com/2026/09/28/openai-spark-hugging-face-bid-war-early-investment-bid-ahead-of-nvidia.html)
- [CNBC — Nvidia agrees to buy Hugging Face for almost $13 billion (Sep 3, 2026)](https://www.cnbc.com/2026/09/03/nvidia-agrees-to-buy-hugging-face-for-almost-13-billion-ai-expansion.html)
- [Clément Delangue on X — deal announcement (Sep 3, 2026)](https://x.com/ClementDelangue/status/2095482998674112733)
- [Business Insider — Nvidia has been in talks to acquire Hugging Face for more than $13 billion (Aug 26, 2026)](https://www.businessinsider.com/nvidia-in-talks-to-buy-hugging-face-13-billion-dollars-2026-8)
- [TechCrunch — Nvidia closes in on Hugging Face acquisition (Aug 26, 2026)](https://techcrunch.com/2026/08/26/nvidia-closes-in-on-hugging-face-acquisition/)
- [TechCrunch — Nvidia confirms it will buy Hugging Face for $12.9 billion (Sep 3, 2026)](https://techcrunch.com/2026/09/03/nvidia-confirms-it-will-buy-hugging-face-for-12-9-billion/)
- [Tom's Hardware — Nvidia to buy Hugging Face for $12.9 billion, report claims (Aug 27, 2026)](https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-to-buy-hugging-face-for-usd12-9-billion-report-claims-could-strengthen-nvidias-open-model-strategy-and-shore-up-position-against-rivals)
- [SemiAnalysis — OpenAI Jalapeño: Better Than Nvidia Blackwell (Aug 25, 2026)](https://newsletter.semianalysis.com/p/openai-jalapeno-better-than-nvidia)
- [The Decoder — OpenAI's first custom chip Jalapeño reportedly beats Blackwell and Rubin in inference benchmarks (Aug 25, 2026)](https://the-decoder.com/openais-first-custom-chip-jalapeno-reportedly-beats-nvidias-blackwell-and-rubin-in-inference-benchmarks/)
- [NVIDIA — Financial results for second quarter fiscal 2027 (Aug 26, 2026)](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/)
- [NVIDIA Technical Blog — Introducing CUDA Rust: Two Tracks for Writing GPU Kernels (Sep 8, 2026)](https://developer.nvidia.com/blog/introducing-cuda-rust-two-tracks-for-writing-gpu-kernels/)
- [Hugging Face — GGML and llama.cpp join HF to ensure the long-term progress of Local AI (Feb 20, 2026)](https://huggingface.co/blog/ggml-joins-hf)
- [The Register — Hugging Face is too important to fall into Nvidia's hands (Sep 3, 2026)](https://www.theregister.com/ai-and-ml/2026/09/03/hugging-face-is-too-important-to-fall-into-nvidias-hands/5294363)
- [MLex — Nvidia's Hugging Face openness pledge identifies US, EU antitrust fault line (Sep 3, 2026)](https://www.mlex.com/mlex/dealrisk/articles/2521456/nvidia-s-hugging-face-openness-pledge-identifies-us-eu-antitrust-fault-line)
- [MLex — Nvidia may seek EU review of Hugging Face deal amid filing uncertainty (Sep 23, 2026)](https://www.mlex.com/mlex/dealrisk/articles/2528556/nvidia-may-seek-eu-review-of-hugging-face-deal-amid-filing-uncertainty)
- [Tech Times — Nvidia's $12.9B Hugging Face deal must pass antitrust review its quasi-mergers dodged (Aug 28, 2026)](https://www.techtimes.com/articles/325863/20260828/nvidias-129b-hugging-face-deal-must-pass-antitrust-review-its-quasi-mergers-dodged.htm)
- [Wccftech — Nvidia calls Hugging Face a 'deconcentration platform' (Sep 3, 2026)](https://wccftech.com/nvidia-insists-its-12-93-billion-acquisition-of-hugging-face-will-escape-antitrust-scrutiny-calling-it-a-deconcentration-platform/)
- [EU Reporter — Commission approves acquisition of Run:ai by NVIDIA (Jan 3, 2025)](https://www.eureporter.co/politics/european-commission/2025/01/03/commission-approves-acquisition-of-runai-by-nvidia/)
- [Hacker News — Nvidia agrees to acquire Hugging Face for $13B (1,988 points, Aug 26, 2026)](https://news.ycombinator.com/item?id=49458161)
- [SemiAnalysis on X — migrating work off Hugging Face (Sep 20, 2026)](https://x.com/SemiAnalysis_/status/2101703331286507887)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-09-30-nvidia-buys-the-hub/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt