Microsoft’s engineers knew something was wrong within minutes. At 20:30 UTC on Sept. 30, customers across the company’s Azure cloud began reporting that ExpressRoute gateways, the private connections linking corporate networks to Azure, had gone down. VPN gateways and Azure VMware Solution failed alongside them. Before the hour was out, Azure Firewall, application gateways and the web application firewall had been pulled into the outage too, and the number of affected regions climbed from 18 to 19.
The services that failed are the doors and hallways of the cloud. ExpressRoute provides a private, dedicated connection between a company’s own data centers and Azure, bypassing the public internet. VPN gateways do the same job over encrypted tunnels. Azure Firewall and the application gateways in front of it sit at the perimeter, inspecting and routing traffic. When these components fail together, the practical effect is that applications stay up but become unreachable: workloads keep running while employees, partners and customers cannot get to them. For an airline or a bank, that is the difference between a slow day and a stopped one.
The cause, Microsoft said on its status page, was an infrastructure operating-system maintenance activity, in plain terms a routine upkeep job on the systems that run the cloud’s networking fabric. The company paused the maintenance once the failures surfaced, but the damage was already done. For hours, companies that had stitched their own networks to Azure found the stitches had come undone.
The geography of the outage showed how quickly a single maintenance error can propagate. Affected regions included US West, North Europe, West Europe, Southeast Asia and Japan West, among others. Some VPN gateways lost redundancy rather than connectivity outright, leaving customers running on degraded paths that could fail without warning. The pattern was characteristic of a change applied across a shared control plane, analysts said, a mistake made once and then replicated everywhere at once.
Microsoft declared the incident mitigated at 03:30 UTC on Oct. 1, roughly seven hours after it began. But five regions, France Central, North Europe, Southeast Asia, UK South and UK West, were still recovering as the company closed out the event, a sign that restoring redundancy takes longer than restoring connectivity. For cloud customers whose contracts promise a certain level of uptime, the clock keeps running until the last region is fully healed.
The timing made the incident awkward in another way. Oct. 1 was the deadline for customers to migrate off Azure’s retired VPN gateway SKUs, the older gateway tiers Microsoft has been winding down. The company did not link the maintenance work to the SKU retirement, but the coincidence meant thousands of customers were already touching their gateway configurations on the same day the gateways broke. Engineers working through the outage had to untangle which failures came from the maintenance and which from customers’ own hurried migrations.
The outage is the latest in a recurring pattern for the world’s second-largest cloud provider. Azure has suffered several networking incidents over the past year in which maintenance or configuration changes to its backbone took down regions at a time. Each episode follows the same shape: a change rolls out, a control-plane service fails, and the failure cascades to the managed services that depend on it before engineers can roll the change back. For Azure specifically, the incidents have been frequent enough that some large customers now run tabletop exercises for a gateway failure the way they once drilled for disk failures.
That pattern is not unique to Microsoft. Every major cloud has shipped a bad change, and the industry’s defenses, gradual rollout, automated rollback, region-by-region canarying, exist precisely because a single bad line of configuration can reach every customer at once. What varies is how quickly providers can stop the bleed. Microsoft’s seven-hour recovery, with several regions still limping afterward, will be weighed against rivals’ faster turnarounds.
For corporate IT teams, the practical lesson is a familiar one. The companies hit hardest were those that had built their network architecture around a single cloud’s gateways, without a redundant path through another provider or an on-premises fallback. Multi-cloud redundancy has been sold for years as insurance against exactly this failure. The invoice arrives on nights like Sept. 30, when the insurance turns out to have lapsed.
Microsoft has said it is investigating the root cause and will publish a post-incident review, the standard practice after an outage of this scale. Until that report lands, customers will be left with what the status page told them at 03:30 UTC: the worst had passed, five regions were still on their way back, and the company had not yet explained how a maintenance window became a seven-hour outage. For now, its reassurance is the same one it offered customers at every stage of the night: the maintenance had been paused, and the cloud’s lights would come back on, region by region.
