Nine months ago Hugging Face reportedly turned down $500 million from Nvidia because it did not want a single investor able to steer the company. On August 27 The Information reported it has agreed to sell the whole thing to Nvidia for $12.9 billion.
Neither company has confirmed it. Neither has denied it, and Nvidia is usually fast to deny things that are wrong. Treat the deal as reported, not done. Treat the shape of it as the story.
What Is Being Bought
Hugging Face makes about $150 million a year. The price is roughly 86 times that. Nobody pays 86 times revenue for revenue. They pay it for position, and the position is that Hugging Face is where open-weight models live: two million repositories, the place Kimi K3, GLM-5.3 and DeepSeek all landed on release day, the place every local inference tool downloads from by default.
In July, Nvidia signed a letter arguing Washington should not restrict open weights, and Jensen Huang told Axios why: more open models means more use, and more use means “you’ll have to sell a lot more NVIDIA computers.” Hugging Face co-signed. Six weeks later the letter’s author is buying its co-signatory.
Then look at the week. On August 25 OpenAI published results for Jalapeno, its own inference chip, and Apple announced a Mac Studio with 512GB of unified memory, the best non-Nvidia box for running a large open model on a desk. Nvidia’s biggest customers are building their own silicon. Owning the shelf is the hedge. If the workloads leave Nvidia’s chips, the downloads still go through Nvidia’s site.
— dpoloncsak, Hacker News, August 27 2026locking the next generation of programmers into CUDA
That was the top fear in a 908-comment thread. The most-cited escape hatch was llama.cpp, “the best way to run models on hardware that isn’t nvidias.” Nobody mentioned MLX. That is the part I want to fix.
Is MLX in Trouble?
MLX is Apple’s array framework for its own silicon, and mlx-lm is how most people run a model on a Mac. Its supply chain, from Hugging Face’s own documentation: mlx-lm downloads weights through huggingface_hub and uploads converted models through huggingface_hub. The mlx-community organisation, thousands of pre-converted models, is run by volunteers, not Apple. The Hub has a dedicated MLX filter and docs page, which is real infrastructure Hugging Face built and can change.
There is no MLX mirror. huggingface_hub honours an environment variable that points it at a different endpoint, so technically the escape hatch is one line of config. The problem is what to point it at. ModelScope, the largest alternative, has around 80,000 repositories against two million and is built for the Chinese market. Nobody has sized a migration, because nobody has needed one.
Compare the other two local-inference formats.
- GGUF and llama.cpp. Hugging Face bought ggml.ai, Georgi Gerganov’s company, in February. GGUF is not hosted on the Hub. It is the Hub’s. A buyer inherits it and has no reason to break something it owns.
- AMD and ROCm. A partnership layer plus AMD’s own cloud running Hub notebooks. Dependent, not owned. But AMD signed Nvidia’s letter and sits in every enterprise buyer’s second-source column. Nvidia has to be seen to keep AMD working.
- MLX. A client library, a volunteer organisation, and a filter. Nothing owned, nothing signed, and the hardware it serves is the direct competitor to DGX Spark, Nvidia’s own 128GB desk-side box with its own local agent stack.
So: not today. The deal is not signed, review will take a year, and the day after it closes nothing will have changed. But of the three, MLX has the least structural protection and the most direct commercial conflict with the new owner.
The threat to MLX is not deletion. It is defaults. A “run this with NIM” button above the fold, Nvidia-optimised checkpoints ranked first, Inference Providers routing to Nvidia silicon unless told otherwise, and MLX still there, one filter click away, maintained by volunteers. Microsoft bought GitHub in 2018 and it stayed a neutral host for years. Then Copilot made it the front door to Microsoft’s AI. Nobody was locked out. The path of least resistance just moved.
The Part Hugging Face Already Told Us
On August 2, before the deal broke, Clement Delangue told CBS how Hugging Face handled the July breach, when commercial guardrails locked its defenders out of their own forensics. They used “the American version of a Chinese model… from Nvidia.” A company broken into by one frontier lab’s models defended itself with a chipmaker’s, co-signed that chipmaker’s lobbying letter, then reportedly agreed to sell to it. Read as a sequence, the deal is the next step in a relationship that was already load-bearing.
And the GLM-5.3 licence I wrote about two days ago has a clause for affiliates of companies over $10 billion in revenue. Hugging Face serves models commercially through Inference Providers. As an Nvidia subsidiary it is over the line. A reading, not a ruling, but the kind of reading a Chinese lab may have had in mind.
What Would Change My Mind
- Apple funds a mirror, or runs mlx-community itself. The fix is cheap and Apple has not done it in two years.
- Nvidia commits to neutrality in writing, with specifics. Ranking, defaults, routing, and the MLX and ROCm integrations named. Nothing like that exists, because there is no signed deal to attach it to.
The Takeaway
The open-weights fight all summer was about whether the file could be published. The file is published, thousands of times over, and the question is now who owns the shelf. Nvidia’s answer is that it should, and it has the money to make that true. MLX is not in trouble this week. It is the format that has to be protected by someone other than its host, and nobody has volunteered. The hardware just got very good. The distribution just got a landlord.
Related Posts
