The Information reported on Wednesday that Nvidia has agreed to buy Hugging Face for $12.9 billion. TechCrunch had it the day before as closing in, and by Thursday morning Bloomberg, CNBC and Fortune had all matched some version of it, which usually means it's real. Worth keeping the asterisk visible anyway: neither company has confirmed a thing. Nvidia didn't comment, Hugging Face didn't comment, and the closest CNBC got was a source allowing that an acquisition has "been part of ongoing and recent talks." So: almost certainly real, not yet officially real.

Everyone led with the number, and the number is the least interesting part of this. Hugging Face pulls in somewhere around $150 million a year depending on whose estimate you take — the company itself said it crossed $100M ARR back in June. Call it eighty-something times revenue, for a business whose core product is free hosting of other people's model weights.

Which tells you straight away that Nvidia isn't buying revenue.

What's actually on the table

Think about the last time you pulled a model. You wrote from_pretrained("...") and never once thought about where the weights lived. That's the whole thing. That's the asset.

The Hub sits on roughly three million public models and about a million datasets, but the raw counts undersell it. What Hugging Face really owns is a decade of muscle memory in which "get the model" and "hit huggingface.co" are the same physical action. It's baked into every README, every Colab notebook, every CI pipeline, every tutorial written since about 2019. It is the default. And defaults are worth more than revenue, because nobody ever renegotiates a default.

Nvidia won the compute layer so thoroughly that the interesting question stopped being who makes the chips. Buying the Hub is buying the layer above it — the place where developers decide what to run, before they get anywhere near deciding what to run it on. That's not a hosting company. That's a toll booth nobody currently perceives as a toll booth.

The awkward part

Here's the detail I can't stop chewing on. In late 2025 Nvidia reportedly offered Hugging Face a $500 million minority investment, and Hugging Face turned it down — the reported reasoning being that leadership didn't want one dominant investor with outsized influence over a platform whose entire value proposition is not belonging to anybody.

Ten months later: same investor, all of it.

And it's not like Nvidia was a stranger. It was already on the cap table from the 2023 Series D — $235M at a $4.5B valuation, a round that also included Google, Amazon, Intel, Qualcomm, IBM and Salesforce Ventures. Read that list again. A good half of it is Nvidia's direct competition in silicon. They all bought in precisely because the Hub was neutral ground — the Switzerland where AMD, Intel and AWS could each ship an integration without any of them having to host a rival's ecosystem.

Switzerland has now been bought by one of the armies.

What breaks, and what doesn't

Nothing breaks on Monday. Your from_pretrained still resolves, your tokenizer still downloads, nothing in your requirements.txt moves. Anyone telling you the Hub enshittifies by Q4 is selling engagement.

The thing to actually watch is the optimum family. Optimum Intel, Optimum AMD, Optimum Neuron for AWS silicon, sitting alongside the TensorRT-LLM path — these are the libraries that let you take a Transformers model and run it fast on hardware that isn't an Nvidia GPU. They're maintained in the open, largely alongside the vendors themselves, and they're exactly the sort of thing that never gets killed so much as quietly under-resourced.

Nobody has to do anything villainous for this to drift. You just need eighteen months of the Deploy button defaulting to one backend, the featured quantisations being the ones that happen to shine on one architecture, the benchmark tables in model cards quietly assuming one runtime. Defaults drift. That is what defaults do once somebody owns them.

The counter-pressure is real, mind you. If AMD and Intel decide the Hub is no longer neutral ground they have both the money and the motive to fund an alternative, and the models themselves are just files — weights are portable in a way a social graph never was. That's the genuinely healthy thing about this corner of the industry: the lock-in here is convention, not physics.

But conventions are sticky, and "just self-host a registry" is the kind of advice that sounds great right up until you try to make an entire ecosystem edit one line in ten thousand READMEs.

So

I don't think this is a disaster, and I'm faintly annoyed at myself for not being angrier about it. Nvidia has been a genuinely decent open-weights citizen — Jensen Huang signed that letter alongside two dozen other companies pushing Washington to support open models rather than restrict them, and Hugging Face's own CEO signed it too. The alignment is real, not just a story about alignment.

The thing is, "we're aligned right now" is a thing you say about people, and "who owns the default" is a question that outlives whoever happens to be running the place. The Hub was valuable precisely because no single vendor could answer it. Once regulators have had their look, one can.

Anyway. Buy your GPUs from whoever you like. The download's still free — for now, and probably for a good while yet, which is either reassuring or exactly the point.