Thursday morning my whole feed was one story. Nvidia has agreed to buy Hugging Face for $12.9 billion. The Information broke it on the 26th, CNBC and everyone else had it within hours, and as I write this neither company has confirmed anything. TechCrunch notes there's no signed agreement yet and the whole thing could still fall over.
Assume it doesn't fall over.
My first reaction was a shrug. Open weights are open weights. A Llama checkpoint does not become less permissively licensed because the company hosting the download button changed hands. My second reaction, maybe ten minutes later, was a lot less relaxed, and I've spent the days since trying to work out which of the two is actually right.
Still not sure. That's most of what this post is.
The detail I keep coming back to
Per the FT reporting summarised here, Nvidia put another $500 million on the table at the end of 2025, at a $7 billion valuation. Hugging Face said no. Not because the money was bad, but because they didn't want a single dominant investor with that much sway over their decisions.
Nine months later they sold the entire company to that investor.
I don't mean that as a gotcha. Twelve point nine billion for the whole thing is a completely different conversation than half a billion for a slice, and people are allowed to change their minds when the number moves that much. But it does clarify that the independence thing was a preference rather than a principle, and knowing which one you were dealing with is useful.
What probably doesn't change
Most of the doom I read on Thursday was overcooked.
The licenses don't change. Apache-2.0 stays Apache-2.0. Weights already on disk stay on disk, and anything mirrored stays mirrored. The Hub is, underneath, a very large pile of git repos with an LFS backend and a good front end, and that shape is hard to quietly close without everybody noticing the same afternoon.
Nvidia also has close to zero reason to wreck it. They sell shovels. A busy open model ecosystem is the best demand generator they have, because every fine-tune somebody kicks off at 3am is compute that gets paid for eventually. The read that this is about defending chip dominance seems obviously right to me. You don't spend thirteen billion on something and then break it.
So on "will open source AI survive this", I lean yes, fairly comfortably.
What might change, quietly
The registry is not the models. That's the whole thing.
The registry is defaults. It's what comes up first when you search for an embedding model. It's which quantisation formats get a first-class badge and which ones you find in a forum thread. It's which inference runtimes the Spaces support properly versus technically. It's what the download counts say, and download counts are how a lot of people pick models whether they'd admit it or not.
None of that requires a single sinister decision. You just ship TensorRT support a bit sooner than ROCm support, every time, for reasons that each look like sensible prioritisation on the day. Five years of that and the default path through the world's model registry runs through one vendor's stack.
The scale makes it matter. The Hub is somewhere around 2.9 million public model repos, a million datasets and 13 million users, and usage is brutally concentrated, with roughly 1.5% of repos accounting for 99.2% of downloads. "Just fork it" is true and close to useless. You can mirror the weights. You can't mirror the gravity.
Where it gets actually interesting
The regulators.
Nvidia has spent the last couple of years growing through arrangements that mostly don't trigger a merger filing. Investments, licensing deals, compute commitments, hiring most of a team without buying the company. Those structures dodge review by design. An outright $12.9 billion purchase does not. This is the first move in a while that has to walk through the front door.
And it walks through that door with history attached. The DOJ has been looking at Nvidia's position in high-end AI silicon since September 2024. The European Commission has been poking at the GPU market since 2023. China's SAMR opened its own probe in December 2024, partly over commitments made back during the Mellanox deal, then dropped it once trade politics shifted.
What makes this one unusual is how clean the vertical story is. Chips at the bottom. The software stack nearly everybody trains on in the middle. And now the discovery layer where the results get published and found. If you wanted a teaching example of vertical integration you would have a hard time building a tidier one.
The question a competition authority would ask is almost exactly the one I was circling above. Does neutral hosting stay neutral when the host sells the hardware, and what does that mean for AMD, for Intel, for everyone doing custom silicon.
Which brings up the part I find genuinely hard. Say a regulator agrees there is a problem. What do you actually order? Divestment is disproportionate. Structural separation of a model registry is a strange thing to even describe. So you land on behavioural commitments, and behavioural commitments on a ranking surface are close to unenforceable. "Maintain neutral treatment of competing hardware in search results and documentation" is a sentence you can write into an agreement and then never meaningfully supervise for the next decade.
My guess is it clears, somewhere in the nine to fifteen month range, with undertakings nobody can measure. I would be pleased to be wrong about that, and I don't expect to be.
The thing I'll actually watch
Not the ruling. Rulings are slow and the interesting part will be over by the time one lands.
The measurable signal is much smaller. Watch how quickly non-Nvidia runtimes keep showing up as first-class on the Hub. ROCm builds, Apple silicon, the custom accelerator crowd, whether their formats stay one click away or drift onto the second page. That's observable, it's observable by anyone, and it starts answering the question long before any competition authority does.
If that gap widens over the next year, the shrug was wrong.
