Nvidia owns Hugging Face: keeping your open-weight pipeline independent

On September 3 Nvidia confirmed what The Information had reported a week earlier: it is acquiring Hugging Face for $12.93 billion, the largest acquisition in the company's history and nearly double the $6.9 billion it paid for Mellanox. The multiple is extreme; Hugging Face's annualized revenue was reported at about $150 million. What Nvidia is buying is not revenue. It is the registry: more than 3 million models, about 500,000 datasets, more than 1 million apps and 18 million users. Jensen Huang has pledged that the platform stays open and that developers will not be pushed toward Nvidia hardware. I believe the pledge is sincere today. I also think you should plan as if it were not.

What was actually bought

Hugging Face is the package registry of open-weight AI. It is where weights are published, where the model card lives, where the GGUF and MLX conversions are hosted, where datasets are versioned, and where the inference endpoints and Spaces run. Every local runtime I have written about, from Ollama in practice to the MLX daily driver, resolves model names against it. The owner of that layer now also owns the dominant GPU, the CUDA ecosystem, and, per an August report in The Information, a roughly 1-trillion-parameter open model called Nemotron 4 that exists explicitly as a demand engine for Nvidia's own inference stack. The regulatory question writes itself: one company owning both the compute and the main marketplace for the software that runs on it. Scrutiny is likely, and the terms could move.

Why Apple Silicon and AMD users should care most

Hardware neutrality is easy to promise and hard to audit. The risk is not that Nvidia deletes the MLX weights. It is a slow, reasonable-sounding tilt: CUDA-first optimization in the hosted inference endpoints, Nvidia-tuned quantizations on the front page, integration work prioritized for TensorRT before Metal or ROCm, and pricing for storage and bandwidth that happens to favor the formats Nvidia's stack reads natively. None of that requires bad faith; it is just what a hardware company does with a registry. If your inference runs on a Mac, on AMD, or on the edge, you are the constituency with the least leverage inside the new owner, and the one I described in Nvidia vs AMD for local inference as already fighting a CUDA-first ecosystem.

Hugging Face was never your infrastructure. It was a very good CDN with a social layer, and the acquisition is the reminder to stop treating it as a source of truth.

Keeping your open-weight pipeline independent

• Mirror what you depend on. Every model in production gets pulled once, hashed, and stored in your own object storage. Your build pulls from your mirror, not from the hub.
• Pin revisions, not names. A model name on the hub can be updated in place. Record the commit hash of the revision you validated and fail the build if the hash changes.
• Keep the conversions. If you ship GGUF or MLX-quantized weights, keep the conversion scripts and the converted artifacts yourself. Community conversions on the hub are the first thing that could quietly go stale.
• Run your own registry for internal models. Fine-tunes, adapters and embeddings should never have the hub as their only home. A private registry with the same hash discipline is a weekend of work.
• Watch the terms. Storage quotas, bandwidth pricing, endpoint availability for non-CUDA backends, and licensing of the hosted tooling are the four things to re-read every quarter under the new owner.

The honest limitation

Nothing has changed yet. The weights are where they were, the API is the same, and hardware-neutral is the stated intent of the buyer. Everything above is a hedge against a future that may not arrive, and the deal itself still has to clear regulators who may attach conditions that make the hedge unnecessary. The cost of the hedge is a mirror, a hash and a habit, which is cheap enough that I would pay it even if Nvidia turns out to be the best steward Hugging Face has ever had. The cost of not hedging is finding out on a Tuesday that the model your product depends on has a new default revision and a CUDA-only fast path.