Microsoft's preliminary report, issued on November 15, says a software defect in a July 2026 update for Microsoft 365 network gateway servers caused last week's outage. The change had shown no impact after deployment. It then left more than 80 percent of those servers unresponsive and cut off email for users and organizations across more than 70 countries.
Tests that missed newer machines
Microsoft says the July update cleared load testing and went live without immediate trouble. The company says that testing tended to use weaker hardware with fewer cores. Under heavy load, the new code created lock contention inside an internal Windows process and used up CPU and memory until the machine froze. Microsoft says the contention is worse on machines with more cores and newer hardware. The change was built for the gateway servers that take incoming connections in front of Outlook and many other Microsoft 365 services. Microsoft has not said what the July work was meant to fix. The report says the software ran worldwide for more than three months and showed no trouble.
Microsoft's report says two ordinary conditions arrived together. Teams failed over to the primary gateway that also supports Exchange Online, a surge the company described as not unusual on its own, while almost 14 percent of networking-gateway capacity was out of service and being restored. Microsoft says spare capacity still looked adequate. GeekWire, describing the same report, places the automatic Teams move about 15 hours before the outage, calls that switch normal operations, and says the failure began when traffic peaked in the U.S. workday, with nearly 14% of capacity already down for a separate problem the report does not describe.
A patch, because rollback was impossible
When a server lost connectivity, traffic moved to the next healthy machine, which then hit the same defect. The fault had spread through Microsoft's fleet within two hours. At the worst point, more than 80 percent of the servers were unresponsive. Putting machines back into service did not help: queued traffic overwhelmed them and they failed again, a pattern called a thundering herd. Microsoft says its usual automated recovery was not built for an unresponsive state this broad. A direct rollback was impossible because the defect was months old. Engineers instead spent roughly 15 hours building and deploying an update that restored the earlier code path in Windows, applied it to groups of servers, and brought service back gradually.
Microsoft says the period of active disruption ran nearly 24 hours. Delayed messages were still going out after service came back Friday morning. The company posted an all-clear on X late Friday night, nearly 38 hours from the start of the outage.