On 29 May 2026, customers using Azure OpenAI saw increased latency, timeouts and 5XX errors, with the impact most pronounced in Europe and Australia East. The trigger was an upstream change that altered how capacity failures surfaced, flooding a shared inference load balancer with retry traffic until it buckled. Around the same window, a thunderstorm took out cooling in a US region. Different causes, same lesson.
The point isn’t this specific incident — it’s the new dependency it exposes. As AI features get embedded into the tools your staff use all day — Copilot in Office, an AI step in a workflow, a customer-facing assistant — the model service becomes part of your uptime story. “Is email working?” now quietly includes “is the AI region healthy?”
You don’t need to over-engineer for it, but you do need to design for it: know which of your workflows now depend on an AI service, make sure the human fallback still exists when the model is slow or down, and don’t let an AI step become a single point of failure in a process that has to keep running.