ChatGPT, Claude, and Grok Went Down Together: What Actually Happened
On the morning of Thursday, September 3, 2026 (US time), three of the world's biggest AI chatbots -- ChatGPT, Claude, and Grok -- all reported elevated error rates and failed requests within roughly the same window. Google's Gemini stayed largely unaffected. The timing sparked an obvious question: were three competing AI companies somehow sharing a single point of failure? The honest answer, based on what's actually been confirmed, is "probably some of them, but not as cleanly as the headlines suggested."
⚡ Quick facts
- When: Thursday, September 3, 2026, over roughly a 90-minute-to-2-hour window during US morning hours; most services were restored by early afternoon US Eastern time
- Hit hardest: ChatGPT and Codex, with tens of thousands of user reports; Claude and Grok each logged over a thousand reports with several models degraded
- Mostly unaffected: Google Gemini, which saw only a small, largely unrelated bump in reports
- Suspected cause: Trouble in Microsoft Azure's East US region, per multiple outlets -- but not officially confirmed by any of the three companies, and xAI's own statement pointed instead to its own Memphis compute center
The timeline
Reports differ slightly on exact minute-by-minute timing -- not unusual when three separate companies each publish their own status-page updates on their own clocks -- but the broad shape is consistent across outlets. Grok and Claude users began reporting problems first, around 9:00 a.m. US Eastern time, followed by a sharper spike in ChatGPT and Codex error reports around 10:30 to 11:00 a.m. Eastern. OpenAI's own status page opened an incident titled "Elevated errors across ChatGPT and Codex," classified as degraded performance, listing components across both products as affected. Anthropic's status page confirmed multiple Claude models -- including Mythos 5.1, Fable 5.1, Opus 5, Opus 4.8, and Opus 4.6 -- were throwing elevated error rates, and said it had identified the cause and was working on a fix. xAI acknowledged Grok was affected too.
By the numbers reported via Downdetector-style tracking, ChatGPT and Codex drew the largest response by far -- more than 37,000 reports at peak, over 66,000 combined across the outage -- while Claude logged roughly 1,300 and Grok roughly 1,000 to 1,400. Gemini, by contrast, saw only around 500 reports, which analysts attributed to routine, unconnected client-side issues rather than a real service disruption. Most services were back to normal within about two hours; Anthropic reported Claude fully restored by roughly 12:27 p.m. Eastern, and OpenAI's incident closed out shortly after.
Why everyone jumped to "it's Azure"
Multiple outlets, including Computing.co.uk and several AI-infrastructure newsletters, pointed to a breakdown in Microsoft Azure's East US region as the likely common thread, since both OpenAI and Anthropic run meaningful portions of their production infrastructure on Azure-hosted compute and networking. Cloudflare and AWS also reportedly showed unrelated issues during the same general window, which muddied the picture further. None of the three AI companies has officially named Azure as the root cause in their own incident write-ups -- OpenAI's status-page language stopped at "investigating," without attributing blame to a specific upstream vendor.
There's also a wrinkle that complicates the tidy "all three ran into the same Azure outage" theory: xAI's own explanation for the Grok disruption pointed not to Azure but to its own Memphis compute center -- home to xAI's Colossus supercluster. SpaceX (which now operates xAI as SpaceXAI following their 2026 merger) said in a public apology, "We are sorry for the issues you may have experienced with Grok following an outage at our Memphis compute center this morning." That's a self-hosted infrastructure problem, not a shared third-party cloud failure -- so even if Azure trouble explains OpenAI and Anthropic's overlapping issues, it doesn't cleanly explain why Grok broke at roughly the same time. Some coverage has since softened to noting the timing overlap without asserting a single, confirmed shared cause.
Why Gemini didn't blink
The most consistent explanation for Gemini's relative calm during the incident is vertical integration: Gemini runs end-to-end on Google's own cloud infrastructure, the same stack Google built and operates for its search, Android, and Workspace products, rather than leasing compute from an external provider the way OpenAI and Anthropic lean on Azure. That doesn't make Gemini immune to outages in general -- Google Cloud has had its own incidents before -- but it does mean Gemini has no direct exposure to an Azure-specific regional failure the way a company built on top of Azure would.
The bigger point: concentration risk
Setting aside the exact root cause, the incident put a spotlight on something enterprise buyers and everyday users don't always think about: three competing AI products can still share the same underlying failure point if they're all renting infrastructure from the same handful of cloud giants. A company that switches between ChatGPT and Claude for redundancy, assuming they're independent, may not actually be diversifying its risk if both ultimately depend on the same regional data center infrastructure going down at the same time. That's prompted renewed calls -- echoed across outlets like Axios and TradingKey -- for AI vendors to publish clearer uptime SLAs and disclose their cloud-dependency footprint, and for enterprise customers to build real multi-vendor, multi-cloud failover into anything mission-critical rather than assuming "using two different AI companies" is enough.
Should you be worried?
For casual, everyday use, no -- this was a roughly two-hour disruption that resolved on its own, and outages of this kind happen occasionally to essentially every major internet service, AI or not. But if you or your business depend on ChatGPT, Claude, or Grok for something time-sensitive -- a customer-facing chatbot, an automated workflow, a coding assistant mid-deadline -- this is a useful reminder to have a fallback tool ready, and to not assume that because a service is AI-powered and from a well-funded company, it's immune to the same infrastructure failures that occasionally take down any other cloud-hosted product.
Frequently asked questions
What happened to ChatGPT, Claude, and Grok on September 3, 2026?
All three services reported elevated error rates and failed requests within roughly the same window on the morning of Thursday, September 3, 2026 (US time). ChatGPT and Codex were hit hardest, with tens of thousands of user reports; Claude and Grok each logged over a thousand reports and had several models degraded. Services were largely restored within about two hours.
Was Microsoft Azure to blame for all three outages?
It's the leading theory, not a confirmed fact. Several outlets pointed to trouble in Microsoft Azure's East US region as a common thread, since OpenAI and Anthropic both lean on Azure infrastructure. But none of the three companies officially named Azure as the cause, and xAI's own statement blamed an outage at its own Memphis compute center for Grok's problems -- which doesn't fit a single shared-cloud explanation.
Why did Google Gemini stay up while the others went down?
Gemini saw only a small bump in reports, generally attributed to unrelated client-side issues rather than a real outage. Analysts pointed to Google's vertical integration -- Gemini runs on Google's own cloud infrastructure end-to-end, rather than leasing capacity from a third-party provider like Azure, which several competing AI companies both rely on.
Does this mean I shouldn't rely on ChatGPT, Claude, or Grok?
Not necessarily -- outages happen to every major tech service occasionally, and all three were restored within about two hours. But the incident is a reminder that if you depend on an AI tool for something time-sensitive or work-critical, it's worth having a backup plan or a second tool ready, since a cloud-infrastructure problem can take out a service even when the AI model itself is working fine.