Nvidia acquires Hugging Face for $12.9B, the hub 18M devs build on
Jensen Huang vows the Hub stays open and Nvidia-optional. But your prod pipeline just gained one owner — here's what changes and how to stay portable.
Copy markdown
The deal, by the numbers
Nvidia is paying $12.93B for Hugging Face — the Hub that hosts 3M models, 500K datasets and 1M apps for 18M+ developers. HF had turned down an earlier $500M Nvidia offer and runs around $150M in annual revenue, so this is a control premium, not a rescue.
'Open and Nvidia-optional' — the pledge
Jensen Huang says the Hub 'will remain an open platform for the entire AI ecosystem' and that 'NVIDIA compute will not be required to build on or deploy through Hugging Face.' Model, framework and cloud choice stay intact — on paper. Clem Delangue framed it as more compute and reach for open source.
The neutral middle of the stack is thinning
A chipmaker now owns where open weights are distributed — weeks after Stripe bought model-router OpenRouter for roughly $8B. Even if nothing breaks today, the independent layer builders lean on is consolidating into platform owners, which is the real reason to pay attention.
Do this today: mirror and pin
Treat HF like any upstream you don't control. Keep internal copies of production weights, tokenizers and configs where licenses allow, pin exact revisions by checksum so nothing shifts silently, and archive the licenses and model cards for the versions you actually shipped.
Design for model-agnosticism
Separate artifact storage from runtime inference, keep direct-provider fallbacks for anything mission-critical, and actually test-deploy on a second provider or accelerator before you need to. The teams that win are the ones that can swap models without rebuilding the system around them.
Elsewhere: npm ships multiple trusted-publishing configs
GitHub moved multiple trusted-publishing (OIDC) configurations per npm package to GA, so one package can authorize several CI workflows to publish without long-lived tokens — useful if you release the same package from more than one pipeline.