Nvidia Buys Hugging Face: What It Means for Open-Weight AI and Public-Sector Buyers
Updated: 6 days ago
Title: Nvidia Buys Hugging Face: What It Means for Open-Weight AI and Public-Sector Buyers
Date: 2 September 2026
Type: Paper
Author: SAASiQ (contact@saasiq.ai)
Word count: 2436 words
Reading time: 10 min
Published: 02-09-2026
Nvidia signed a definitive agreement on 2 September to buy Hugging Face, the main public platform for distributing open-weight AI models and datasets, in a deal Nvidia values at about $12.9bn. The talks were first reported on 26 August, and Nvidia expects to complete in the first half of 2027, subject to regulatory approvals. This paper sets out what was agreed, why the platform matters, what Nvidia has and has not promised about keeping it neutral, how competition authorities may respond, and what organisations that rely on open-weight models should do while the deal is under review.
What was agreed
The first report came from The Information, which said on 26 August that Nvidia was close to buying Hugging Face for about $12.9bn. TechCrunch's account of that report is dated 26 August, and other outlets date it 27 or 28 August. Fortune followed on 27 August and noted that Business Insider had reported Nvidia's takeover interest over the preceding weekend. On 28 August TechTimes set out the competition questions the deal would face. Nvidia signed a definitive agreement on 2 September, according to its filing with the US Securities and Exchange Commission. Hugging Face's chief executive is Clément Delangue.
The filing sets out the price in two parts. About $11.9bn is payable to Hugging Face's stockholders, subject to adjustments, and up to about $1.0bn goes into an equity retention programme for Hugging Face staff who join Nvidia. Nvidia's blog gives the total as $12,930,300,000. Completion is expected in the first half of 2027 and depends on customary conditions, including regulatory approvals.
In the same filing Nvidia commits to keeping the platform open, to allowing users to go on uploading and downloading models and datasets, and to supporting other chip vendors.
What Hugging Face does
An open-weight model is one whose trained parameters, the weights, are released for anyone to download. An organisation can then run the model on its own servers, or in a cloud account it controls, instead of sending its data to the model maker's service. Hugging Face is the main place those files are published and collected. It also hosts datasets and evaluation results, sells hosted inference through its Inference Endpoints service, and runs Spaces, which are small applications built on top of models and shared on the platform.
Nvidia's own figures give the size. The platform has more than 18 million users, about 3 million models, 500,000 datasets and 1 million applications, and more than 200,000 companies use it. Some earlier coverage put the user base at 13 million developers; the 18 million is Nvidia's figure. Nvidia is a large publisher there in its own right, with more than 500 models and more than 250 datasets.
On 1 September, the day before the agreement was signed, Perplexity released PII-Tracer, the classifier it uses to detect personal data in its new hybrid local and cloud product, as open source on Hugging Face, along with a benchmark of 13,148 synthetic conversations in 13 languages.
Delangue has described Hugging Face as the 'Switzerland of AI'. It was valued at $4.5bn in 2023, in a $235m round led by Salesforce Ventures, according to Fortune. In late 2025 it turned down a $500m investment from Nvidia at a $7bn valuation, reportedly to avoid having a single dominant investor. Less than a year later it has agreed to be bought outright by the same company.
Its revenue is about $150m a year, up from about $100m two months earlier, and Delangue told TechCrunch it was close to profitability. Progressive Robot calculates the price at about 86 times revenue; neither company has published a multiple.
How the deal is structured, and Nvidia's reasons
Nvidia has spent heavily on AI companies over the past nine months, mostly without buying them. According to TechTimes, it used licence-plus-talent arrangements, in which it licenses a company's technology and takes on many of its people while the company itself stays independent, for Groq (about $20bn, in December 2025), Enfabrica (about $900m) and Poolside (about $7bn). Those came to about $27bn in total. TechTimes calls them quasi-mergers, and none went through formal merger review.
Hugging Face is an outright equity acquisition. That means it has to be notified under the US Hart-Scott-Rodino Act, whose filing threshold is about $119m, and reviewed by the Federal Trade Commission or the Department of Justice. Of the deals TechTimes lists, it is the first to go through that process.
TechCrunch and Fortune report three reasons for the purchase. It protects demand for Nvidia's GPUs at a time when OpenAI, Google, Amazon and Anthropic are all building their own chips. People who download open models need compute to run them, and Hugging Face is where they download them. And it gives Nvidia a route back into cloud services, after it scaled back its own DGX Cloud offering about a year ago, along with a way to sell capacity that would otherwise go unused.
Delangue's account is that the approach came from his side. He told CNBC he went to Nvidia first because, in his view, open-source AI 'was at the turning point' and 'needed more resources, more scale, more visibility'. He described the goal as 'empowering 100 million AI builders to own their intelligence rather than rent it'.
In SAASiQ's view the two accounts are compatible, and they point to the same pressure on Nvidia after completion. What Nvidia is paying for is the traffic of developers and models through the platform. That traffic depends on publishers and users continuing to treat Hugging Face as neutral ground, and a platform seen as tied to one chip vendor would give them a reason to publish and download somewhere else.
What Nvidia has promised on neutrality
Nvidia's commitments are in two documents. The SEC filing says it will keep the platform open, allow continued uploads and downloads of models and datasets, and support other silicon vendors. The blog adds that Nvidia compute will not be required to build or deploy through Hugging Face, that the platform will support multiple clouds and multiple accelerators, and that the Hugging Face brand will be kept.
Jensen Huang wrote that 'Hugging Face will remain an open platform for the entire AI ecosystem', and that 'open weights broaden access to AI and help ensure that AI leadership is distributed across companies, institutions and communities.' Delangue told CNBC that Nvidia had 'committed to strongly supporting Hugging Face and our mission while keeping the platform open, independent, and compute agnostic.'
Other points are not covered. The blog does not say where Hugging Face will sit inside Nvidia, and neither source sets out how the platform will be governed after completion, who will decide changes to its policies, or what access Nvidia will have to the usage data Hugging Face holds.
PCMag's analysis names three risks that sit in those gaps. The first is ranking and discovery. Hugging Face decides how models are ordered in search and surfaced to users, and that affects which ones get downloaded. The second is competitor telemetry: the platform holds data on what is downloaded and deployed, including by companies that compete with Nvidia. The third is favouritism in the back end, where the hosted services on the platform could be tuned to run better on Nvidia hardware.
As we read the published commitments, none of those three would breach them. A platform can stay open to uploads and downloads, and still allow other hardware, while its search ranking, its defaults or its hosted services lean towards one vendor. Any buyer or regulator looking for protection on those points would need it written down separately, covering how models are ranked, how usage data about other vendors is kept apart from Nvidia's own business, and whether hosted services are offered on the same terms whatever the hardware.
Competition review and timing
Only the US filing is certain. Coverage names the FTC or the Department of Justice, the European Commission, the UK Competition and Markets Authority and China as possible reviewers, the last of these according to PCMag. Progressive Robot goes further and says the deal faces mandatory reviews in the US, the EU and the UK, but no source found for this paper confirms a filing with the Commission or the CMA. The UK merger regime is generally voluntary, meaning the parties are not obliged to notify the CMA before completing. UK and EU review should be treated as likely or possible, and neither is confirmed.
Nvidia's earlier deals had already drawn comment in Washington. In March 2026 Senators Warren and Blumenthal asked whether the Groq deal had been 'structured to evade scrutiny by antitrust regulators'. In January the FTC's chair, Andrew Ferguson, said the agency was examining structures 'being constructed to try to escape Hart-Scott-Rodino review'. Both comments were aimed at the licence-plus-talent arrangements. Hugging Face, as a conventional purchase, goes through the review process those arrangements did not.
The nearest precedent is Nvidia's own. Its $40bn purchase of Arm collapsed in 2022 after opposition from the FTC, the CMA, the European Union and China, over concerns that Arm would not remain neutral under Nvidia's ownership. That was chip design and this is model distribution, but the question regulators asked then is the one this deal raises now: whether a supplier that many competitors rely on can stay even-handed once the dominant AI chip supplier owns it.
The first half of 2027 is Nvidia's expected completion date and depends on those approvals. Until then Hugging Face is an independent company, and the commitments in the filing and the blog describe what Nvidia intends to do once it owns the platform.
Who is affected
Model publishers are the most directly affected. PCMag reports that Chinese labs account for roughly 41 per cent of the models on the platform, a figure from a single source. After completion those labs would be distributing their models through a platform owned by a US chip company, and China is among the possible reviewers named in coverage.
Rival chipmakers, AMD among them, and the companies building custom silicon need open models to run well on their hardware and to be easy to find and deploy. The commitment in the SEC filing to support other silicon vendors is the one written for them, and its value to them depends on the ranking and hosting points above.
Cloud providers are affected through Nvidia's reported interest in selling cloud services. The hyperscale providers accounted for $48.71bn of Nvidia's $89.0bn data-centre revenue in the quarter to July. If Hugging Face becomes Nvidia's route back into selling compute, as TechCrunch and Fortune suggest, Nvidia will be selling capacity on the platform alongside its largest customers.
Enterprises and public bodies that host models themselves are the group this paper is mainly written for. Open-weight models are the preferred route for many public-sector and regulated buyers, because a model the organisation runs on its own infrastructure keeps personal and sensitive data inside its own controls. Where the model came from Hugging Face, any deployment pipeline that pulls it by name at build time depends on the platform each time it runs.
SaaS vendors that embed open models in their products carry the same dependency one step removed. A finance or HR application that uses an open-weight model to read invoices or classify documents sourced it from somewhere, and its customers may not know where. That is a question to add to supplier due diligence: which open-weight models the product uses, whether the vendor pulls them from Hugging Face or keeps its own copies, and what it would do if the platform's terms or availability changed after the sale.
What buyers should do now
Start by finding out where Hugging Face sits in the organisation's AI supply chain. The places to look are model downloads in development and deployment pipelines, any use of Inference Endpoints, datasets pulled for training or testing, and Spaces that staff use as tools. Suppliers should be asked the same question about the models inside their products. The output is a list of dependencies with an owner against each one.
Each entry should record the model's name and exact version, where it was downloaded from, the licence it was released under, where it runs, what data it reads and who is responsible for it. Hosted services such as Inference Endpoints need a separate line, because they are a running service under contract with Hugging Face, whereas a downloaded model is a copy the organisation already holds. After completion that contract would be with a company owned by Nvidia.
For models that matter, keep a copy. Mirroring, or pinning, a model means storing the exact files at a specific version in the organisation's own repository, so that a deployment does not depend on the platform being available or unchanged. The licence text should be stored with it, as it stood on the day of download, and the record should show which version is in use and where it came from.
Add change-of-control and neutrality clauses to AI supplier contracts. Last week's decision by OpenAI on Cursor is the current example of access changing after an acquisition. SpaceX completed its purchase of Cursor's maker, Anysphere, on 14 August, and OpenAI then said it would stop supplying models, with a proposed end date of 12 November. A supplier whose product depends on a model or platform should say in the contract what happens if that access is withdrawn or its terms change after a sale, and how much notice the customer gets.
Keep a route to other hardware. The deal's published commitments include support for multiple accelerators, and a buyer should check that its own deployments would still run if it moved from Nvidia GPUs to another vendor's chips or another cloud. For most organisations that means testing each critical model on a second platform once and recording the result.
Watch the CMA and the European Commission. If either opens a review, its published decisions will set out the concerns and any remedies Nvidia offers, which may include binding versions of the commitments in the blog.
Put the dependency on the risk register. Oracle Fusion and other finance and HR programmes that use open-weight models, directly or through a partner's extension, should record Hugging Face as a supplier dependency with Nvidia as its prospective owner, and review the entry when a regulator reports or the deal completes.
Dates to watch
The agreement was signed on 2 September. OpenAI's proposed end date for supplying models to Cursor is 12 November 2026. Nvidia expects the Hugging Face acquisition to complete in the first half of 2027, subject to regulatory approvals.
SAASiQ - Intelligent Solutions for SaaS ©
