ChatGPT, Claude, and Grok All Went Down at Once; OpenAI Tweeted Something Cryptic About Stars

On the morning of September 3, ChatGPT, Claude, and Grok all stopped responding within minutes of one another, in the kind of overlapping, three-service outage that would normally read as a boring infrastructure story about shared cloud dependencies. It did not stay boring, because at some point during the disruption, OpenAI’s official account on X posted four words with zero context: “The stars are almost aligned.”

Nobody at OpenAI explained what that meant. Nobody needed to — the internet had already done the work. OpenAI’s next model, code-named Astra, has been the subject of weeks of speculation about an imminent launch, and a chunk of the internet spent the outage rearranging the letters of “ASTRA” into “ATRSA” and then squinting at that until it produced “RESET,” which is either an elaborate marketing tease, a coincidence, or proof that given an unexplained outage and a cryptic tweet, people will find a hidden message in literally anything, including an anagram that requires swapping two letters and ignoring that it isn’t actually a word.

The far less cinematic explanation, and the one every outlet that actually checked converged on, is Microsoft Azure. The trigger, per multiple reports tracking the incident, was a regional failure inside Azure’s East US infrastructure — the same cloud backbone that happens to underpin three of the four largest AI chatbot services on the market. OpenAI attributed its own portion of the disruption to a routing error, saying the ChatGPT and Codex outage began at 7:43 a.m. Pacific and was resolved by 8:17 a.m., a 34-minute window, though other trackers logged a considerably longer and messier disruption spanning several hours across the affected services before everything was reported fully operational the next morning. Google’s Gemini, notably, kept working, because Gemini doesn’t run on Azure — which is either an argument for infrastructure diversification or the least interesting sentence anyone said about this whole event, and it’s the second one.

None of the companies involved has connected OpenAI’s tweet to the outage, and taking it at face value requires believing OpenAI chose the exact fifteen-minute window of an unplanned, embarrassing, multi-vendor cloud failure to also drop a deliberately mysterious teaser about an unrelated model launch — a level of either confidence or indifference to optics that would be its own small miracle of corporate communications. The more mundane read is that the post was scheduled or planned independently, landed in the worst possible moment by accident, and OpenAI’s comms team spent the rest of the morning watching a hype campaign and a service outage merge into one story against their will.

What’s genuinely notable here has nothing to do with stars, resets, or acronyms, and everything to do with how load-bearing a single cloud region has quietly become for an entire category of consumer software. Three of the highest-traffic AI products on Earth, run by three separate companies that spend considerable energy competing with and occasionally suing each other, all went dark at roughly the same time because they’re all, in the end, tenants in the same building. That’s a genuinely useful piece of information about systemic risk in AI infrastructure. It got about a tenth of the attention that “the stars are almost aligned” got, because a vague tweet about destiny is a much better story than a regional Azure failover, and the internet, given the choice between infrastructure diversification and a conspiracy theory, will reliably pick the conspiracy theory — especially when the company involved appears to have accidentally handed it to them for free.

Sources: