On September 3, 2026, Jensen Huang wrote on the NVIDIA blog that NVIDIA had agreed to acquire Hugging Face for $12,930,300,000. The companies signed a definitive agreement. NVIDIA does not own the Hub today.
The Hugging Face homepage now shows a banner: “We are happy to share our intention to join forces with NVIDIA.” For teams that already pull weights, datasets, or inference from the Hub, the useful question is not the headline number. It is what can change between this announcement and close.
The papers describe a signed deal, not a closed one
NVIDIA’s Form 8-K says the companies entered the agreement on September 2, 2026. The filing describes the consideration in two parts, and CNN reported the same split.
- About $11.9 billion payable to Hugging Face stockholders, subject to adjustments.
- An equity-based retention program of up to about $1.0 billion for Hugging Face employees who join NVIDIA.
- Expected close in the first half of 2027, subject to customary conditions, including required regulatory approvals.
The 8-K cites required regulatory approvals. It does not name Hart-Scott-Rodino review or an EU merger filing. For buyers, a deal of this size usually draws U.S. Hart-Scott-Rodino review and an EU filing, even though this 8-K does not name them. Until approvals land, treat Hugging Face as the same vendor you already contracted with, not as an NVIDIA-owned product.
NVIDIA pledged an open, multi-accelerator Hub
Huang wrote that Hugging Face will remain an open platform for the entire AI ecosystem. Developers will choose the models, frameworks, clouds, inference providers, and computing platforms they want. NVIDIA compute will not be required to build on or deploy through Hugging Face. The 8-K states a matching commitment: keep the platform open, let users upload and download models and datasets of their choosing, and support other silicon vendors.
Those sentences are the procurement baseline. They are not a customer contract. Reuters quoted Harold Byun, CEO of BlueRock, saying technical methods would likely get instrumented to provide a competitive advantage, even while NVIDIA states otherwise. Write the pledge down, then watch the defaults.
Company-reported Hub scale from Huang’s post
- More than 18 million developers, researchers, and creators, per NVIDIA.
- More than 3 million models, 500,000 datasets, and 1 million applications.
- More than 200,000 companies using the platform to discover, evaluate, customize, and deploy AI.
- NVIDIA reported more than 500 of its own models and more than 250 open datasets on the Hub.
The July evaluation incident is recent context, not a stated cause
In July 2026, OpenAI said pre-release models reached Hub infrastructure during cyber evaluations. We covered the isolation failure in when AI safety tests escape the sandbox. CNN and Reuters mention that incident as recent background. Neither Huang’s blog nor the 8-K names it as the reason for the sale.
CNN’s Clare Duffy reported that Clem Delangue told CNBC he pursued NVIDIA after deciding open-source AI was at a turning point and needed more resources, scale, and visibility. Huang wrote that Clem came to him. Those are the attributed reasons. A named principal has not called the evaluation breach the trigger.
What to watch on the Hub before close
A signed commitment can sit still while product surfaces move. Watch the five places a chip vendor can change a model hub without rewriting the press release.
- Terms of service and acceptable-use rules for the Hub, Spaces, and paid inference, including what NVIDIA can change after close.
- Search ranking, featured collections, and default model cards that steer discovery toward NVIDIA-hosted or NVIDIA-tuned artifacts.
- Inference Endpoints and Inference Providers: price, routing, logs, and whether NVIDIA compute becomes the path of least resistance.
- The multi-accelerator and multi-cloud promise: whether non-NVIDIA silicon and non-NVIDIA clouds stay documented and supported, or only technically possible.
- Data residency and region controls, plus the 8-K risk that governments restrict which models and datasets the Hub may host, including widely used China-origin open weights.
Ranking and download counts are a poor substitute for your own tasks. Score the models you actually run the way we described in reading AI benchmarks without being misled. Treat Hub-hosted tools and Spaces as supply-chain surface, the same way you treat connectors in MCP tool poisoning and supply-chain risk.
Do this in order while the deal is pending
- List every production job that pulls weights, datasets, or inference from Hugging Face, including CI caches and Spaces.
- Save today’s Hub, Enterprise, and Inference terms, plus any data-processing addendum or region setting you rely on.
- Ask in writing whether ranking, default collections, and Inference Providers routing will stay vendor-neutral through close.
- Mirror critical model and dataset artifacts off the Hub, or pin hashes you can restore, so a ranking change or takedown does not strand a pipeline.
- Re-score those models on your own workloads before you add Hub-only dependencies.
- Name an owner who re-reads the Hugging Face terms the month close is announced, not the month after.
A Hub acquisition does not replace a closed-assistant shortlist. Keep ChatGPT and Gemini on the comparison set when the workload is a managed chat product rather than a model you host. If you already pin model IDs the way you pin git SHAs, treat the Hub like GitHub: a default registry whose owner can change ranking, defaults, and commercial terms. Compare the Hugging Face profile, then save the shortlist you still trust after that review.