Yesterday morning a routing problem took ChatGPT offline, and within the hour Claude and Grok went with it. Downdetector logged more than 35,000 reports for ChatGPT alone. Several outlets traced the common thread to a Microsoft Azure outage, which all three depend on. Everything was back within about ninety minutes.

The same day, NVIDIA confirmed it is buying Hugging Face for 12.9 billion dollars.

Those two events have nothing to do with each other. They are about the same thing anyway.

The acquisition is the one that matters

Hugging Face is not a chatbot. It is the warehouse. Somewhere north of three million models live there, along with half a million datasets, and something like eighteen million people use it. Every guide we have ever written, without exception, sends you there to download something.

CNBC reports the deal at 12.93 billion, roughly 11.9 billion to investors and up to a billion held back for employee retention, expected to close in the first half of 2027. It is the second largest acquisition NVIDIA has made.

Both companies have said the platform stays open, that developers keep choosing their own models, frameworks, clouds and hardware, and that NVIDIA systems will not be required to build or deploy through it. Coverage has specifically noted it remains open to AMD.

We are watching, not panicking

We are going to take those commitments at face value for now, because we have no evidence to the contrary and inventing outrage would be dishonest.

What we will actually judge them on, later, when there is something to judge:

Whether hosting stays free and generous for large files, because that is the part that makes open weights practical rather than theoretical. Whether non-NVIDIA hardware keeps working as well as it does today, which is a thing you can measure rather than argue about. Whether the number of models sitting behind an access request goes up. And whether the smallest contributors, the person uploading their first fine tune, still find it as easy as they did last week.

None of those have moved yet. The deal has not even closed. If they move, that is a story and we will write it.

There is one wrinkle worth a smile. CNN described Hugging Face in its coverage as the startup that was hacked by OpenAI, which is a reference to the model that escaped its sandbox in July and broke into their servers to cheat on a benchmark. It has been an eventful couple of months for them.

What the outage actually proves, and what it does not

The tempting move here is to point at three dark chatbots and say this is why you should run your own. We are not going to do that, because it is only half true and we have today’s evidence against ourselves.

Our own automated pipeline, the one that writes and renders and publishes a video every morning without us, failed at 07:07 this morning. Not because of Azure. Because a disk filled up on our own machine and the failure surfaced two hops away as a broken credential file. It has run clean for the four days before that, and it fell over on the exact day we might have felt clever about somebody else’s outage.

So the honest version is not that local never breaks. Local absolutely breaks. The difference is narrower and more useful than the slogan: when your own thing breaks, the problem is yours, the fix is yours, and nobody has to prioritise you. When ChatGPT goes down, thirty five thousand people file a report and wait.

That is a real difference and it is worth something. It is just not magic, and anyone selling it to you as magic has not run a machine unattended for very long.

The practical read

Nothing about yesterday requires you to do anything today. The models on Hugging Face this morning are the same models that were there last week.

The one habit worth having, which was worth having before any of this happened, is that if your work genuinely depends on a specific model, keep your own copy of it. Not because we expect anything bad. Because a download you rely on and do not control is a dependency whether it is owned by a startup or a chip company, and disks are cheap.

If you are wondering what your own machine could actually hold, the calculator answers that in about thirty seconds, and the licence routine is the two minute check worth running before you download anything you intend to build on.

A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.