On September 3, the three biggest AI assistants — ChatGPT, Claude and Grok — went down or degraded within the same hour, with all three companies acknowledging problems on their status pages. All three services were reporting recovery by early afternoon US time (MacRumors, Lifehacker).
🔍 THE BOTTOM LINE
One chatbot going down is an annoyance; all three going down together is a structural diagnosis. Whether or not Azure turns out to be the common thread, the episode demonstrated that millions of people whose daily work runs through AI assistants have essentially no fallback when the top providers fail at once.
What actually happened
OpenAI’s status page showed issues across ChatGPT and Codex. Anthropic reported elevated errors across multiple Claude models — including Mythos 5.1, Fable 5.1 and Opus 5 — hitting claude.ai, the API and Claude Code. xAI’s status page showed all Grok functions down at one point (Lifehacker). Downdetector logged tens of thousands of user reports for ChatGPT alone, with thousands more for Claude and Grok.
The timing overlap is the part that got engineers talking. All three platforms rely heavily on Microsoft Azure for cloud compute — alongside other providers — and a report on the outage noted that speculation quickly settled on a shared Azure component such as a regional compute cluster, networking layer or identity service. To be clear: none of the three companies has confirmed Azure as the root cause, and it remains a suspect, not a finding.
Why a triple outage matters more than three outages
Outages happen. In June, Claude went down alone and the takeaway was “maybe keep an open-weights fallback ready.” The September 3 event sharpened that argument considerably: when the failure mode is correlated across providers — because they share infrastructure, or because traffic from a downed provider floods the others — the fallback you planned to switch to is the one that’s also struggling.
The maths is uncomfortable. If a large share of ChatGPT’s displaced users migrated to Claude during the outage window, that load spike lands on a platform already dealing with its own elevated errors. One user’s quip on X — “people were moving to ChatGPT because Claude is down and that caused an overload” — describes a genuine cascade dynamic, not just a joke.
It’s also worth noting what this means for anyone building on these APIs. Codex going down isn’t just chat windows greying out; it’s CI pipelines, coding agents and customer-support bots silently stalling. The more “agentic” the workflows get, the more a two-hour outage costs.
The unglamorous fix
The honest mitigation isn’t exotic: multi-provider routing, cached responses for critical paths, and a documented manual fallback for anything the business actually depends on. For individuals, it’s knowing which open-weight model you could run locally if everything hosted went away at once. None of this is glamorous, and most of it will never be needed — until the hour it is.
What it means for New Zealand
The outage window hit Thursday morning US time — early Friday morning NZ time — a period when many NZ developers were asleep and unaffected. But the lesson compounds: NZ businesses building on a single AI provider have no domestic fallback, and latency to alternative regions is a permanent reality of our geography. If your workflow died on September 3, you learned which of your tools share infrastructure. That’s cheap information now and expensive information later.
❓ FAQ
Was the outage caused by Azure? Unconfirmed. All three providers use Azure among other cloud services, and the overlapping timing made a shared Azure component the leading speculation in coverage — but no company has publicly confirmed a root cause as of writing (coverage).
How long did the outage last? Roughly a few hours in the US morning. By about 1:19 p.m. ET on September 3, all three status pages reported services running normally (Lifehacker).
Should I switch to an open-weight model as backup? Depends on your workload. For critical automated workflows, having a local or self-hosted fallback — even a smaller model — is cheap insurance. For casual chat use, probably not worth it; see our coverage of the open-weights debate.
Is a repeat likely? Yes, somewhere. AI usage keeps climbing and inference infrastructure keeps concentrating; correlated failures across providers sharing cloud infrastructure will recur. The question is whether providers disclose shared root causes when they do.
🔍 THE BOTTOM LINE
The September 3 outage will be forgotten within a week. The dependency it revealed won’t be: three nominally independent AI companies, three status pages, one shared infrastructure story. As more daily work moves onto these platforms, the interesting question stops being “will they go down” and becomes “who else goes down with them.”
📰 Sources
- MacRumors — Claude, ChatGPT, and Grok Are Down
- Lifehacker — A Bunch of AI Platforms Are Down
- Company status pages (OpenAI, Anthropic, xAI)
