Three AI providers went down at the same time on Thursday.
ChatGPT first, then Claude, then Grok, all within a few minutes of each other, around eight in the evening here. Cursor went with them. Gemini maybe, maybe not; the reports were mixed and Google never confirmed anything broad. Most of it was back inside about an hour and a half.
None of our own services were affected. But we have agents in CI and in code review now, so a good part of the afternoon’s work quietly stopped anyway, which is its own kind of outage even if nothing on our dashboards went red.
Nobody has explained it yet. Azure reported a network incident in East US during the same window, which is a fact and not an answer. Three separate companies failing together could mean one shared dependency underneath all of them, or it could be coincidence. Overlapping timing is not a cause. The postmortems should land in a week or two and they’re worth waiting for.
The part I keep coming back to is simpler than the cause.
We run more than one model provider on purpose. Split traffic, separate contracts, a self-hosted tier for the work that can tolerate it. The entire point of the arrangement is that no single vendor can take us out. On Thursday all three went down inside ten minutes.
That only works if the providers fail independently, and I have no evidence that mine do. If all three sit on overlapping infrastructure in overlapping regions, then what I bought was three of the same thing at three different prices.
Nvidia agreed on Tuesday to buy Hugging Face for about $12.9 billion. That is where most people go to download open models: eighteen million developers, three million models, and for a lot of small teams it is effectively the entire supply chain. Nvidia says it will keep the platform open and keep supporting other chip vendors, and there is no particular reason yet to doubt that. But the neutral place where everyone shares models is now owned by the company that sells the hardware to run them, which is a different arrangement than the one that existed last week. The deal is expected to close in the first half of next year if regulators agree.
A few other things worth noting from the past two weeks.
Tech layoffs passed 127,000 for the year across 281 companies, with Apple, TikTok, LinkedIn and Netflix all cutting in August. CareCloud disclosed a breach affecting 3.75 million patients: found in March, disclosed in August. Five months. And the UK government briefed energy companies that a cyberattack had knocked a small British generator offline for four days in July.
The thing most people I know actually spent the fortnight talking about was GTA 6. Gameplay clips from a playable build leaked, a lot of them, and Rockstar spent the following week firing off DMCA takedowns and subpoenaing Microsoft and Discord to work out who posted them. They had already ended remote work partly over security and it leaked anyway. The Netflix Extended Look went out on the 27th as planned and the November 19 date is still holding. The strange part is that the leaked footage is reportedly at least a year old, so a good chunk of the internet spent the week arguing about a version of the game that no longer exists.
TIL: there is no good way to check whether your providers fail independently. Correlation only shows up in outages you’ve already lived through. The best I’ve come up with is to log every provider failure with a timestamp and look once a quarter for clustering. Not clever, but it’s at least evidence.
See you in fifteen.