It’s probably the thing that everyone thinks it is. OpenAI, Anthropic and SpaceXAI are all routed through something that we’re not supposed to know exists and that thing had a whoopsie.
Wouldn't be surprised: Snowden's revelations 10 plus years ago already showed how the NSA was injected into the data centers of Social Media, it's only logical that they would now demand to be injected into the biggest, most information providing data stream of the planet of the present: LLM services.
If this hypothesis were true then a spy agency may want to rewrite responses. Every tool in an agent's harness becomes an remote procedure call you can make on that machine. Including a tool to execute a shell command, in many. A harness is completely isometric to a backdoor, it's the same code written with a different intention.
Maybe it was a power hub. And we're not allowed to know where the DC is. If we knew it was a power issue, then with other information (perhaps over time) we could determine the DC location. Or something along these lines.
Worth mentioning that this room takes a split from the main feed and is not in the path of traffic. Whatever is in this room could go down and it would not cause an outage.
Interestingly, fiber signals can be split passively[1] (without a transducer), which should be extremely reliable. No idea what technology this kind of application would use though.
Does the NSA even have the ability to monitor all AI traffic like this? Wouldn’t that require tons of data centers that there literally hasn’t been time to build yet? I really have no idea, maybe the asymmetry of the compute required to monitor is way lower than the compute required to serve inference?
I dont think it requires a lot. I think of them as data hoarders more than anything else. It doesnt seem out of their capabilities to store a ton of chats. Maybe they're having scaling problems with the increased data rates they're hoarding and it led to an outage.
"AI Traffic" is just traffic. If the infrastructure exists to monitor/buffer traffic (it does) then this can be monitored as well. Whether this hiccup was due to them hitting their limits briefly (or turning it on, or etc) who knows.
I bet it's even negligible traffic compared to e.g. Netflix or YouTube. Even a 'huge' context window is nothing compared to the random library of a basic Web page.
Were there significant API outages too? I didn’t notice any on my production workflows, and I’d assume what you’re implying would cover API routes too, otherwise it seems kinda pointless.
OpenAI and Anthropic have their systems in dozens of data centers, including using compute from the major cloud providers. Are you implying that all of these data centers (and many of their employees) are involved in helping the the US government secretly tap every single AI conversation by routing them through some unknown network/device?
Or did a couple of companies with poor uptime records happen to have overlapping downtime?
Occam's Razor heavily, heavily points us towards the latter.
What makes you think they'd need to touch every datacenter? All of these endpoints use existing providers with decades-long history at this point, and network monitoring is already a proven 'feature' of the agencies they'd need to co-exist with over their lifetimes.
If anything, Occam's Razor would point to a common denominator with all of them, given it wasn't network-wide, as far as i know.
Explain the system in which you could capture all of these chats with no knowledge of anyone in these data centers. How are they routed to this NSA system or through some NSA device when these companies' compute are spread over hundreds of data centers?
> All of these endpoints use existing providers with decades-long history at this point
That is just factually inaccurate. Their data centers aren't old and they lease a lot of compute from companies that didn't exist 5 years ago.
>Explain the system in which you could capture all of these chats with no knowledge of anyone in these data centers.
They all transit the same wires as all other traffic. Copy them at any regional bottleneck. https://en.wikipedia.org/wiki/Room_641A. Additionally, i'd admit that maybe someone(s) at these companies knows. But if we think there isn't any person who would agree to do this then I think we're being naive.
> Their data centers aren't old and they lease a lot of compute...
Again, they transit the same wires as everyone else. Here i'll also add that these companies have been actively courting government relationships (and Anthropic attempting to repair damaged ones), why would they stand on principles here and not any of the other many frontlines they've visibly acquiesced?
I just think it's easier to re-route their traffic than, as you say, touch every single datacenter and its employees in some way.
I'd say Occam's Razor leans easily to the former as well, given the history of projects that Snowden revealed and were never shut down, plus all of the cooperation with the federal government that's being touted in recent announcements from both companies.
I work at OpenAI and I was the Incident Commander for yesterday's outage.
We had a routing error within our infra that caused issues for some of our products. It was not related to the Astra launch. We don't comment on other providers' outages.
Neither of these companies have stellar uptime records. Their downtime episodes overlapped in this instance. In this case, it was a partial downtime for both.
Also, OpenAI is saying what caused it:
> "A routing error starting around 7:43 am PT on Thursday, September 3, made ChatGPT and Codex unavailable for some users across platforms"
Anthropic stated their issue started earlier:
> "The company began alerting about a “partial outage” at 6:23 am PT on Thursday that involved “elevated errors on requests to Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5.”
I don't get why everyone reaches for an extraordinary explanation when the ordinary will do: both of these companies have quite a bit of downtime.
I thought the consensus on here yesterday was that it was likely caused by cascading failures. OpenAI had an issue during their GPT-6 rollout, taking down their service. This caused a lot of OpenAI users to push their requests (or a larger share of their requests) to Claude and/or Grok, which pushed their load high enough to cause outages.
We used to experience similar effects when I worked at a CDN. If one CDN would go down, we would see immediate spikes in traffic. Luckily, we had procedures for that to prevent overload, but the AI folks might not have the capacity/capabilities to handle that sort of cascade yet.
It could be as simple as a new model (astra) was released which takes more resources combined with a surge in usage due to novelty took down OpenAI. Meanwhile everyone at big companies have the ability to switch models and moved to Anthropic pushing it too over the edge.
People keep saying this "everyone can switch" thing, but it's not my experience at $VeryBigCorp. We don't have an Anthropic contract at all. Is this different in other places? The bigger and more bureaucratic an org is, the less I would expect it to have contracts with all the providers. Curious about others' experiences.
At my company (not massive but not tiny either - I think its about 8000 global employees) we get a choice between pretty much all available Google, OpenAI, Anthropic, XAI models.
Using an agent-agnostic harness like Pi switching is trivial - I run into occasional disconnects and slowdown and switch quite easily.
Yeah, I think there are lots of places that have a semi-official preferred AI vendor but also some backup subscriptions floating around. For example, at my workplace, we generally use Claude, but I also have some kind of Codex subscription too, which I'd use if Claude went down.
The OpenAI outage lasted only 15 minutes and when it happened everyone started to use the other models which created super heavy load for them. This then cascaded into them all being down.
it wasn't timed like a cascade, and it relies on the premise that every single frontier provider is working so efficiently that they spend exactly what they need to provide for their exact market with perfect margins.
I do not believe personally that 1) they can forecast their load that perfectly 2) they chose to remain that inflexible in a world where they are at each others' throats and a single meme can cause bursts of activity.
Or it could simply mean that they're operating at the limit of the capacity they were able to purchase and do not have headroom to handle load spikes. From what I understand, that's the situation Anthropic is in. And since Anthropic is now leasing a large portion of xAI's datacenter capacity, it's plausible that an Anthropic load spike could cause issues for xAI as well.
Your comment makes it sound like they can just push a button and spin up more capacity -- but at this scale and in this GPU-constrained environment, that's not really how it works.
9 AM PST / 12 Noon EST on weekday. All my co workers immediately went "oh codex is down lemme try Claude". Multiply that by millions. Easy to see how they all went down. The OpenAI downtime also coincided exactly with their tweets announcing GPT-6 and about an hour before they started to role it out.
Not that uncommon. It's so easy to just put up the new thing and make the old thing the failure route. But the old thing nearly never had the bandwidth for today's traffic. A famous EBay outage some years ago was just such a scenario.
OpenAI was rushing and making last minute changes for a big product launch. Easy to mess something up in that situation. The outage lasted like 20 minutes.
Anthropic is down a lot regardless, and in this case only specific models were affected.
xAI probably couldn't handle the extra traffic it was getting from the other two.
Watched it happen. Doesn't look like traffic moving off OpenAI took out the others. Looked like the opposite. The Anthropic thread was initially chock full of marketing accounts claiming "Last straw! I finally moved to Codex and I'm so happy! No problems there!" - and then they all got deleted when Codex went down too.
Yet another blatant proof that those companies are not smarter than anyone else. They might be focusing on intelligence and yet in practice we can all see they are not doing better than most.
Not doing better than most at routine IT/platform stuff, they seem to be doing fantastically at co-opting the US Gov into their "vision" and ramming through DC's getting built.
[1] https://en.wikipedia.org/wiki/Fiber-optic_splitter
Having a magical cert doesn't mean you can just intercept everything.
No mainstream browser (or any browser?) is doing cert pinning.
What "other methods" are there that are deployed and actually in use?
https://theintercept.com/2016/11/16/the-nsas-spy-hub-in-new-...
They still do, but they used to, too.
Or did a couple of companies with poor uptime records happen to have overlapping downtime?
Occam's Razor heavily, heavily points us towards the latter.
If anything, Occam's Razor would point to a common denominator with all of them, given it wasn't network-wide, as far as i know.
> All of these endpoints use existing providers with decades-long history at this point
That is just factually inaccurate. Their data centers aren't old and they lease a lot of compute from companies that didn't exist 5 years ago.
They all transit the same wires as all other traffic. Copy them at any regional bottleneck. https://en.wikipedia.org/wiki/Room_641A. Additionally, i'd admit that maybe someone(s) at these companies knows. But if we think there isn't any person who would agree to do this then I think we're being naive.
> Their data centers aren't old and they lease a lot of compute...
Again, they transit the same wires as everyone else. Here i'll also add that these companies have been actively courting government relationships (and Anthropic attempting to repair damaged ones), why would they stand on principles here and not any of the other many frontlines they've visibly acquiesced?
I just think it's easier to re-route their traffic than, as you say, touch every single datacenter and its employees in some way.
We had a routing error within our infra that caused issues for some of our products. It was not related to the Astra launch. We don't comment on other providers' outages.
Also, OpenAI is saying what caused it:
> "A routing error starting around 7:43 am PT on Thursday, September 3, made ChatGPT and Codex unavailable for some users across platforms"
Anthropic stated their issue started earlier:
> "The company began alerting about a “partial outage” at 6:23 am PT on Thursday that involved “elevated errors on requests to Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5.”
I don't get why everyone reaches for an extraordinary explanation when the ordinary will do: both of these companies have quite a bit of downtime.
Biggest competitor goes down and all of a sudden you have a lot more traffic...
And not only that, when one goes down a bunch of API traffic switches over to the other, spiking demand and knocking it down.
We used to experience similar effects when I worked at a CDN. If one CDN would go down, we would see immediate spikes in traffic. Luckily, we had procedures for that to prevent overload, but the AI folks might not have the capacity/capabilities to handle that sort of cascade yet.
Using an agent-agnostic harness like Pi switching is trivial - I run into occasional disconnects and slowdown and switch quite easily.
Does this really need an explanation?
I do not believe personally that 1) they can forecast their load that perfectly 2) they chose to remain that inflexible in a world where they are at each others' throats and a single meme can cause bursts of activity.
Your comment makes it sound like they can just push a button and spin up more capacity -- but at this scale and in this GPU-constrained environment, that's not really how it works.
It was 1000% a cascade. I would bet on it.
It was extremely weird... and if it was a load thing they probably would have explained it by now?
why?
Anthropic is down a lot regardless, and in this case only specific models were affected.
xAI probably couldn't handle the extra traffic it was getting from the other two.
Not everything is a conspiracy.
Anthropic: 6:23 am PT.
OpenAI: 7:43 am PT.
Watched it happen. Doesn't look like traffic moving off OpenAI took out the others. Looked like the opposite. The Anthropic thread was initially chock full of marketing accounts claiming "Last straw! I finally moved to Codex and I'm so happy! No problems there!" - and then they all got deleted when Codex went down too.
Ask HN: Why were OpenAI, Claude, and Grok simultaneously down?
https://news.ycombinator.com/item?id=49551096