Nvidia agrees to buy Hugging Face for $12.9 billion and pledges the hub stays open
The largest open-model hub is moving inside a chipmaker, which makes knowing how model hubs, licences and hardware backends work a practical skill.
Nvidia said on 3 September that it has agreed to acquire Hugging Face, the platform that hosts most of the world's openly shared machine learning models, for $12.93 billion. The announcement came in a blog post signed by Nvidia chief executive Jensen Huang, who framed the plan as scaling the platform, strengthening its infrastructure and widening access to AI for developers and institutions.
Hugging Face was built over the past decade and its hub now holds more than 3 million models, 500,000 datasets and 1 million applications. The company says more than 18 million developers, researchers and creators use it, and over 200,000 companies build on it. Nvidia is already the single largest contributor of open assets to the hub, having released more than 500 models and 250 datasets there.
The chipmaker made several commitments about how the platform will operate after the deal. It says Hugging Face will remain open to the whole AI ecosystem, that Nvidia hardware will not be required to build or deploy through the platform, that multi-cloud and multi-accelerator development will continue, and that open source and open weight models from any lab will still be supported. The founders, referred to in the post as Clem, Julien and Thomas, will carry on their work within Nvidia under the existing Hugging Face brand.
According to The Next Platform, the transaction is expected to close in the first half of 2027, subject to regulatory approval. Its report quotes Huang describing the current period for open models as a delicate one, and Hugging Face chief executive Clément Delangue explaining that the company saw two possible paths ahead and chose to commit further to open source AI. The Next Platform also reported Nvidia's view that open models account for roughly half of its business, with cloud service providers making up the other half.
Why it matters
The deal moves the main distribution point for open models inside the company that sells most of the hardware those models run on, which makes the openness commitments the part to watch rather than the headline number. It confirms that open weight models are now a core commercial channel rather than a research sideline. It also suggests that the infrastructure for hosting, versioning and serving models is becoming as strategic as the models themselves. Regulators will have until 2027 to test whether a hub can stay neutral when its owner sells accelerators.
Applications and demos
Spaces, chat apps and agents that call a model; Hugging Face hosts about 1 million of these
Models and datasets (the hub)
Git-backed repositories holding weights, model cards, licences and data; over 3 million models and 500,000 datasets
Libraries and runtimes
Transformers, PyTorch, JAX and ONNX Runtime translate a checkpoint into operations for a device
Accelerators
Nvidia, AMD, Apple, Google and other silicon; Nvidia says its compute will not be required to use the hub
Clouds and local machines
Where the hardware lives; the pledge covers multi-cloud and multi-accelerator deployment
What you can learn from this
A model hub is a version-controlled file store with a social layer. Hugging Face repositories are Git repositories underneath, with large binary weight files handled by a large-file extension so that multi-gigabyte checkpoints can be pushed, pulled and tracked by commit. Model cards, licences and tags sit alongside the weights as ordinary files, which is why a hub can be mirrored, cached or self-hosted. Seeing the hub as Git plus storage plus metadata explains both why it scaled to millions of repositories and why an acquirer must invest heavily in bandwidth and storage.
Open weights and open source are different promises. An open source model publishes training code, data recipes and weights under a permissive licence; an open weight model publishes only the trained parameters, often under a custom licence with usage limits. Both can be downloaded and run locally, but only the first can be fully reproduced. Nvidia's pledge to support both categories matters because most popular models today are open weight, and the licence file in each repository governs what you may legally build.
Hardware neutrality is enforced by software abstraction layers. A checkpoint is a set of tensors; what makes it run on a particular chip is a runtime such as PyTorch, JAX or ONNX Runtime with a backend for that accelerator. Libraries such as Transformers sit above those runtimes, so the same model code can target Nvidia, AMD, Apple or Google silicon. The commitment that Nvidia compute will not be required is therefore a promise about keeping those backends first-class, which a learner can check by looking at which devices a library release actually tests against.
Regulatory review of vertical deals focuses on foreclosure. When a supplier of one layer of a stack buys a widely used platform at another layer, competition authorities ask whether the owner could favour its own products, for example by degrading rival accelerator support or steering deployments to preferred cloud partners. A close in the first half of 2027 reflects the time those reviews take across several jurisdictions. Behavioural commitments published at announcement often become the conditions regulators later formalise.
Platform network effects belong to the ecosystem, not the owner. Hugging Face's value lies in the millions of models, datasets and Spaces contributed by others, and those contributors can leave if trust erodes. Mirrors, alternative hubs and internal registries already exist, so switching costs are real but not prohibitive. That dynamic is why an acquirer publicly commits to openness on day one: the asset being purchased is a community, and communities are held by consent.
We teach this
How to use this in practice
Audit the licences of every model your projects download. Run
huggingface-cli scan-cacheto list the checkpoints already on your machine, then open each repository's model card and record the licence and any usage restrictions in a small spreadsheet. Done looks like a table with one row per model, a licence column, and a note on whether commercial use is permitted. Flag any model whose licence you cannot identify as unusable until you can, and note the date you checked.Pin models by commit hash rather than by name. In any script that calls
from_pretrained, add therevisionargument with a specific commit SHA taken from the repository's history page, and store that SHA in your requirements or config file. Done means a fresh clone of your project downloads exactly the same weights next month, regardless of what the maintainer pushes to the main branch. This also gives you a reproducible artefact to cite in a bug report or an evaluation.Prove one model runs on two backends. Pick a small open weight model, load it once with the default CUDA or CPU device and once with a different backend available to you, such as Apple's MPS, an AMD ROCm build or ONNX Runtime, and compare outputs on ten identical prompts. Done is a short markdown note recording which backend worked, the tokens per second you saw, and any operator that failed to convert. Expect small numerical differences between backends; note them rather than treating them as failures.
Set up a local mirror for the models you depend on. Point the
HF_HUB_CACHEenvironment variable at a shared drive, download your critical models withsnapshot_download, and then run your test suite withHF_HUB_OFFLINE=1to prove nothing reaches the network. Done means the tests pass with the network disabled, which is your insurance if hub access, pricing or terms change after the acquisition closes. Record the total size of the mirror so you know the storage cost of that independence.
Sources
- NVIDIA to Acquire Hugging Face — NVIDIA Blog
- Nvidia Expands Its Open Source AI Presence With $12.9 Billion Hugging Face Buy — The Next Platform
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report