On the morning of September 3, 2026, Claude had a service incident around 8:37 a.m. ET. Grok opened an investigation around 9:30. ChatGPT hit a routing error close to 10:43. Three companies, three separate infrastructures, degraded within about two hours of each other — and nobody has tied it to one shared cause.
The status pages told one story: elevated errors, a fix identified, service restored by early afternoon. Downdetector told a livelier one — more than 38,000 reports for ChatGPT alone before the day was out. Social media told the truest one. AI writer and Every CEO Dan Shipper put it plainly: "ChatGPT is down, and I am immediately 50 IQ points dumber." Chemist Simon Maechling asked the question a lot of people were quietly thinking: "Claude, Grok and ChatGPT. How can anyone get any work done?"
Funny, until you think about what "renting" actually means here. A lot of teams have made a single AI vendor load-bearing — for engineering, for writing, for judgment calls nobody ever wrote down as a dependency. When that vendor wobbles, so does the team, and the fallback is refreshing a different company's status page and hoping.
Three separate failures is the finding, not the reassurance
It would be easy to read September 3 as an argument for "just use two providers." That's not what actually happened. Three separate companies, on three separate infrastructures, degraded inside the same two-hour window, for reasons nobody has connected. That's not a coincidence a second vendor protects you from — it's a reminder that the whole category shares failure modes even when the root causes don't overlap: the same clouds underneath, the same identity providers in front, the same instinct in every client to retry harder the moment the primary starts timing out.
the window in which all three labs' AI products degraded on the morning of September 3, 2026
Downdetector reports for ChatGPT alone during the outage
confirmed shared root cause connecting the three incidents, as of publication
The fix is the layer underneath, not the vendor on top
Picking a favorite provider and hoping its outage math is better than the others' is not a resilience strategy — it's a preference. The actual fix is architectural: AI deployed in a way that degrades gracefully instead of stalling out completely, with a fallback path that doesn't depend on the same vendor being the one to fix itself. "The vendor is down" should be an inconvenience your workflow absorbs, not a Thursday-morning crisis that stops the team.
What "resilient" actually requires
- A dependency map that includes AI, not just infrastructure. Most business continuity plans list cloud and payment processors as critical vendors. Few list the AI product a team now can't function without.
- A degraded mode, not just a backup vendor. What does the workflow do when the primary AI product times out — wait, retry, or fall back to a defined manual or secondary path? If the answer is "wait," that's not a plan.
- Detection independent of the vendor's own status page. By the time a status page says "investigating," your team has usually already noticed. Closing that gap is what a resilience plan is actually for.
This week didn't prove you picked the wrong provider. It proved that whichever one you picked, the team's ability to work still depends entirely on a system it doesn't control.