One region, three availability zones or a global edge do not help when they share the mechanism that failed.
Google Cloud London: the spare zones need capacity too
On 20 October 2026, a power failure in one London zone triggered failover pressure and shared-dependency failures elsewhere in the region.
The boundary
The physical outage affected one zone. Recovery and impact varied by workload; owning hardware would still require tested power and failover arrangements.
Azure ExpressRoute: redundant gateways, shared recovery controls
On 30 September–1 October 2026, an Azure control-plane overload and premature health checks disrupted some redundant gateways.
The boundary
The overall incident window was not every customer's downtime. Local configuration fallback helped restore affected connections.
AWS: one empty DNS record, many broken services
On 19–20 October 2025, a latent race condition in DynamoDB DNS automation produced an empty US-EAST-1 endpoint record. AWS recorded impact across EC2 launches, Lambda, ECS, EKS, Fargate, Connect, Redshift, authentication and support access over different periods of an event lasting roughly 15 hours.
The boundary
This was not a global disappearance of AWS. Existing EC2 instances remained healthy and some cross-region configurations continued to work.
Azure: the global front door closed
On 29 October 2025, incompatible configuration metadata crashed Azure Front Door edge sites. For more than eight hours, customers saw timeouts and DNS failures affecting Azure and Microsoft services across regions.
The boundary
Some dependent services failed over successfully. Microsoft also said some services had no established fallback strategy.
Azure OpenAI: retries became the outage
On 29 May 2026, an upstream change caused Microsoft 365 workloads to amplify capacity failures into an internal retry storm. A shared Azure OpenAI inference-routing layer overloaded across multiple regions; Microsoft said one failed request could trigger up to 48 additional attempts.
The boundary
Impact varied by customer, resource and region. Microsoft isolated the first-party workload and scheduled stronger routing isolation and overload controls.
AWS Middle East: multi-AZ did not survive the region
From March 2026, physical damage that AWS attributed to the conflict in the Middle East took down multiple zones in Bahrain and the UAE. On 15 September 2026, AWS said it could not restore data held only in its Bahrain Region or in one UAE zone.
The boundary
AWS advised customers to migrate from March and said most did so before the region became unavailable. The loss applies to data that existed only there.
Railway: one account, three providers unreachable
On 19 May 2026, Google Cloud automatically suspended Railway's production account. When cached routes from a Google-hosted control plane expired, workloads still running on AWS and Railway's own hardware became unreachable for hours.
The boundary
This is Railway's account; Google has not published its own. The non-Google hardware kept running and failed only through a shared dependency.
IDCF Cloud: the provider told 495 customers to restore from their own backups
On 7 October 2026, a ransomware attack stopped virtual servers in four zones of IDC Frontier’s East Japan Region 1. The provider said data in those zones was expected to be difficult to restore and that it could be restored only from backups customers held themselves.
The boundary
Other zones, regions and the provider’s private cloud were reported unaffected. By 14 October it found no sign of data theft and said snapshot data for 194 of the 495 customers could be extracted.
Cloudflare: four failures, four mechanisms
Between June 2025 and February 2026, Cloudflare documented a third-party storage dependency failure, two global-change incidents and an accidental withdrawal of customer BYOIP routes.
- 12 June 2025: Workers KV dependency cascade
- 18 November 2025: oversized Bot Management feature file
- 5 December 2025: 28% of served HTTP traffic affected
- 20 February 2026: BYOIP prefixes withdrawn
Open the Cloudflare incident dossier →