Microsoft's Exchange Online email services went down for millions of users worldwide, and the outage lasted nearly 24 hours. A source familiar with the outage told The Verge that it began in Switzerland, where Exchange Online servers that power Outlook started to fall offline. The problem then spread to Mexico and other countries as Microsoft's failover systems tried to shift email traffic elsewhere. Within a couple of hours, the source said, more than 80 percent of Exchange servers were unresponsive. Microsoft marked the outage "sev0" in its internal incident tracking system, the highest severity level.
Charles Lamanna, leader of Microsoft's Copilot, Agents, and Platform team, wrote in a post on X that "a latent defect caused servers to become overwhelmed and unresponsive under certain load conditions." Microsoft's preliminary post-incident report attributes the bug to "lock contention inside an internal Windows process" affecting select Windows servers. The report says the bug stemmed from a fix prepared specifically for Microsoft 365 networking gateway servers and does not affect regular versions of Windows.
According to Lamanna, the bug had been present on those gateway servers since July without causing impact. He wrote that the contention issue worsens on machines with more cores and newer hardware, while Microsoft's load testing focused on weaker hardware with fewer cores.
Lamanna wrote that the bug surfaced after two other incidents. Fifteen hours before the Outlook outage began, Microsoft Teams failed over from an alternative networking gateway to the main gateway supporting Exchange Online, increasing load and changing traffic patterns. At the same time, about 14 percent of that gateway's server capacity was out of service because of an unrelated issue from the previous day.
Lamanna wrote that recovery took longer than usual because servers returning to service were immediately overwhelmed by waiting traffic and became unresponsive again. He said Microsoft's automated recovery processes were not designed for restoring service from such a broad unresponsive state. The source told The Verge that overlapping interventions from teams working on the problem complicated coordination at times, and that an engineering bridge call reached its maximum of 1,000 participants.