Insight / AI and infrastructure
Published
NVIDIA's leadership change, Hugging Face and your model exit plan
What NVIDIA's leadership transition and proposed Hugging Face acquisition mean for model users, and five practical alternatives for an open-model supply chain.
On 12 November 2026, NVIDIA announced that Jensen Huang had resigned and that Colette Kress would serve as interim chief executive. The announcement raises a fair question about the future of NVIDIA’s planned acquisition of Hugging Face and its public support for open models. Hugging Face chief executive Clément Delangue responded that the company would keep working for open-source AI. A leadership change and that reassurance are both important signals; neither settles future platform policy.
In September 2026, NVIDIA announced an agreement to acquire Hugging Face. An agreement is a material change in direction; it is not evidence that the transaction has closed or that every product decision, integration or commercial term has changed. NVIDIA says the platform will remain open to models, frameworks, clouds and hardware beyond its own. Delangue described open-source AI as needing more resources and scale in a joint interview about the proposed deal.
Those statements matter, but a model supply chain should work even if a platform changes ownership, pricing, access rules or priorities. The question for a team using Hugging Face is not whether to leave tomorrow. It is whether the team could reproduce and relocate a working model if it needed to.
Leadership is a reason to watch the commitments
Huang personally championed open-weight AI in NVIDIA’s acquisition announcement. Kress’s transition statement says the company intends to continue its customer and developer commitments. Those statements should be assessed against what the platform actually offers after a deal, if it closes: neutral hardware support, exportable artifacts, predictable terms and a functioning community. It would be speculation to say the leadership change has already withdrawn any of those commitments.
Separate the platform from what you depend on
Hugging Face is several things at once: a discovery catalogue, a place to publish model and dataset files, a set of libraries, a community, and a route to hosted inference. Losing access to a catalogue page is different from losing a private repository that holds the only copy of production weights. A decision to change inference providers is different again from moving a dataset with permissions and history.
List the dependencies in those terms. For each production model, record its publisher, exact revision, license, files and checksums, tokenizer, prompt template, runtime version and permitted uses. Record who can retrieve the artifact, where a verified copy lives, and what will happen if the original URL disappears. A model card is useful documentation; it is not a backup of the bytes or of the rights to use them.
If the workflow calls a hosted endpoint, also record the request and response contract, rate limits, authentication, region, data-handling terms and failure behaviour. The runbook for a model change that breaks a business pipeline starts with that inventory. For sensitive material, our data-sovereignty guide explains why the location of weights alone does not settle where your data goes.
What to watch after an ownership change
Look for changes in artifact access, commercial terms, hardware neutrality, moderation and publishing rules, and community governance. These are monitoring questions, not predictions. A developer can compare the written policy with actual behaviour over time: can a non-NVIDIA backend still download and run the same model? Can an organisation still export its own private artifacts? Are licenses and revision histories preserved? Are changes announced with enough notice to test a replacement?
Keep a small, licensed mirror of your critical artifacts outside any single hub. Run a scheduled restore test, not just a download test. The restore should recreate the tokenizer, runtime and a few representative outputs on a machine your team controls. If the model’s license restricts redistribution, keep the backup within the rights it actually grants.
Five alternatives, each for a different job
There is no honest ranking that makes five services interchangeable with all of Hugging Face. These are five useful routes to evaluate:
- Kaggle Models supports model discovery, versions, documentation and downloads, with strong links to notebooks and competitions. Check the original publisher and license; some catalogue entries are integrated from Hugging Face, so a listing there is not automatically an independent copy.
- ModelScope offers model and dataset repositories with its own upload and download tooling. Test the regions, authentication, artifact formats and support path your organisation needs before treating it as a replacement.
- Ollama’s model library is a convenient route to models intended to run through Ollama. It is a narrower runtime-oriented catalogue, useful for local experimentation and deployment but not a full substitute for a general research hub or dataset repository.
- Replicate community models make published model implementations runnable through hosted inference. This is a service alternative when an API is the dependency. It does not, by itself, give you the same artifact custody or the same runtime as a local deployment.
- A controlled repository and artifact store can combine GitHub with Git LFS and your own object storage or registry. This gives a team more control over its release process, while making it responsible for bandwidth, access control, integrity checks, discoverability and long-term maintenance.
For each route, try a real migration of one model and its documentation before announcing a platform policy. Compare the same revision, not a similar model name. Check whether private artifacts, dataset permissions, large files and automated downloads work under the target’s terms.
An exit drill is more useful than a prediction
Choose a low-risk model, export its files and metadata, restore it from the alternative location, run an evaluation set, and time the recovery. Then repeat with the production model whose loss would hurt most. The exercise exposes hidden dependencies: a tokenizer fetched from elsewhere, a gated license, a conversion script on one person’s laptop, or an agent tool that expects a specific API response.
The same method applies to local engines. Strata’s approach to local inference shows how a model artifact, inference runtime and agent harness can be separated. Our local-system performance checklist turns that separation into measurements. An ownership announcement should prompt those preparations, not an unsupported claim that open models have already become unavailable.