ChatGPT, Claude and Grok all went down at the same time on September 3, and for hours nobody at OpenAI, Anthropic or xAI could explain why.
If you couldn't get ChatGPT to load on Wednesday morning, you weren't alone: Claude and Grok went down too, at the same time, and for a few hours nobody at OpenAI, Anthropic or xAI could say why. Three companies, one window, on September 3. That kind of alignment is rare - outage trackers treated it as one story rather than three. According to 9to5Google, Downdetector logged more than 20,000 reports tied to OpenAI's services alone, with thousands more each for Anthropic's Claude and xAI's Grok. Tom's Guide put the OpenAI number even higher: over 40,000 reports for ChatGPT. Microsoft's Copilot and the coding tool Cursor were struggling too. Web, mobile and desktop clients were all hit.
Google's Gemini stayed up throughout. That single exception is what turned a routine outage into a genuine puzzle. Three competing AI labs, run by three separate companies with three separate engineering teams, do not normally fail at the same moment for the same reason. When they do, the obvious suspect isn't the AI models themselves. It's whatever plumbing sits underneath them.
What actually broke #
Multiple outlets, including 9to5Mac and MacRumors, reported that the trouble traced back to Cloudflare, the content delivery and security network that a huge share of the internet quietly depends on. Microsoft Azure was also seeing an elevated outage count at the same time, and OpenAI, Anthropic and xAI all lean on Azure or Cloudflare somewhere in their stack, whether for content delivery and DDoS protection, or just raw compute. None of the three companies had confirmed a root cause as of publication. Their status pages simply acknowledged elevated errors and promised updates.
Anthropic Open-Sources Claude Commerce Agents Blueprint Ahead of Black Friday Anthropic has open-sourced Claude Commerce Agents, a free GitHub blueprint for building AI shopping and merchant agents, claiming pilot retailers saw carts grow up to 35% and purchases complete 60% more often. The release lands three months before Black Friday, just as OpenAI has pulled back from running checkout directly inside ChatGPT after weak... - how to build AI shopping agents with Claude - commerce agent blueprint for increasing customer cart conversions
Here's the thing worth sitting with: none of this required a bug in GPT-5, Claude or Grok. It just required one shared vendor to hiccup. Cloudflare alone has knocked large chunks of the internet offline before, including a November 2025 outage that hit X and ChatGPT together. Wednesday's incident looked like a repeat of that same failure mode. Same script, more AI products caught in the blast radius this time.
For founders and teams who've built workflows around these tools, that's the real story - not the outage itself, but what it reveals about how thin the layer is between three independent AI companies and one shared point of failure. If your customer support bot runs on Claude and your code review pipeline runs on Cursor, you might assume you've diversified. You haven't, necessarily. Wednesday showed that diversification can be an illusion if all three sit on the same Cloudflare edge network or the same Azure region underneath.
The Astra coincidence #
The timing added an extra layer of speculation. OpenAI had spent the two days before the outage talking up Astra, the next model it plans to release. On September 1, the company said Astra was the first of its models to cross what it calls a Critical cybersecurity capability threshold, according to CNBC and TechCrunch: capable of finding and exploiting unknown software vulnerabilities without step by step human guidance. That's the top tier. OpenAI has said only that a public release is coming soon, not that Wednesday was the day, and the company hasn't confirmed whether Astra and the long-rumored GPT-6 are even the same model. Still, the coincidence of a major outage landing the same week as that announcement was enough to send speculation running on social media that something bigger was happening behind the scenes.
There's a simpler, less dramatic explanation sitting right there, and it's the one the evidence actually supports: shared infrastructure, not shared code, took three chatbots down at once.
Don't treat this as a one-off. If your business runs on a single AI provider with no fallback, you found out on September 3 what that costs you. The fix is simple: a backup model or provider you can flip to when the primary one goes dark, the same way smart teams already keep a backup payment processor or a second cloud region. Most founders haven't bothered, because outages like this have been rare enough to ignore. They just got a lot harder to ignore.
Also read: Saudi Arabia Built Its National AI Model on Top of China's MiniMax • Multiverse Computing's Quasar 438B Becomes Europe's Top AI Model • Songcove Turns Your Actual Love Story Into the Gift for Wedding Anniversary You Cannot Buy in a Store