What happened
- On September 3, 2026, between 6:30 and 10:00 a.m. Pacific time, ChatGPT, Claude and Grok reported overlapping outages on their official status pages. Gemini users also reported failures on Downdetector.
- OpenAI described a routing error. Anthropic reported elevated errors across several models and services. xAI documented a Grok outage and mentioned its compute providers.
- Downdetector logged about 37,000 ChatGPT reports at the peak, from more than 70 countries.
- There is no common cause confirmed by the four companies. Various outlets attributed the episode to a regional Microsoft Azure failure and to an outage at a Memphis data center, with no cross-confirmed official statement to date.
Why it matters
- Any Chilean product that calls one of these APIs in real time — a support assistant, a ticket classifier, a product-listing generator — went three hours without a response during European and U.S. business hours.
- The operational lesson isn’t to switch providers. It’s to treat the provider the way you treat your bank, hosting or payment gateway: with monitoring, graceful degradation and a response the end user doesn’t notice.
- If the service agreement your team signed doesn’t define compensation for downtime, the outage is absorbed by your operation, not the provider’s.
The number
37,000 ChatGPT reports at the peak, according to Downdetector, from more than 70 countries.
Context
The four platforms share layers that aren’t always visible from outside: cloud regions, networks, name resolution, authentication and hosting capacity. Several independent services failing in the same window doesn’t prove a single point of failure, but it does show how much common infrastructure lies beneath products sold as alternatives to one another. Concentration is reinforced by moves such as Nvidia’s purchase of Hugging Face.
What’s next
- None of the companies announced a joint post-incident analysis.
- No timelines announced for individual root-cause reports.
Bottom line
In August, a U.S. court filing had already recorded the warning that a power outage at a compute plant would take several assistants offline at once. The hypothesis stopped being a hypothesis on September 3.
Sources
- OpenAI Status — official incident page
- Anthropic Status — official incident page
- Caída simultánea de ChatGPT, Claude y Grok — Neoteo, comparative timeline (in Spanish)
Edited by Rodrigo Cornejo. How we select and verify the facts: who writes these notes.
