Annual report
Analysis of 47 major public incidents across AWS, Microsoft Azure, and Google Cloud, 2017–2025.
Cloud reliability is high and improving on paper, yet the incidents that do occur are longer, wider, and more correlated than most architectures assume. Four findings define 2026: multi-hour recovery norms, change error as the dominant cause, flagship-region concentration, and a persistent gap between SLA credits owed and claimed.
3.4h
median duration of a major cloud incident — recovery is measured in hours, not minutes
CloudDowntime, State of Cloud Reliability 2026 — median major-incident duration (204 minutes).
45%
of major incidents trace to configuration or change error — humans and automation, not hardware
CloudDowntime, State of Cloud Reliability 2026 — root-cause distribution.
63%
of major AWS incidents involve us-east-1 — flagship-region concentration is a measurable risk
CloudDowntime, State of Cloud Reliability 2026 — regional concentration analysis.
68%
of major incidents breached at least one monthly SLA target — most of the credits went unclaimed
CloudDowntime, State of Cloud Reliability 2026 — SLA breach and credit-gap analysis.
The corpus counts 47 major public incidents 2017–2025. Frequency is remarkably stable year over year — five to seven majors annually across the big three — suggesting a floor set by complexity growth offsetting reliability investment.
Median major-incident duration is 204 minutes (~3.4 hours); the 90th percentile reaches 11 hours. The longest event in the corpus — AWS us-east-1 DynamoDB DNS event, Oct 2025 — ran 15 hours, most of it recovery cascade rather than the original fault.
Configuration and change errors cause 45% of major incidents — more than software defects (21%) and hardware/network failure (11%) combined. The industry’s biggest reliability lever is change safety, not redundancy.
Failures concentrate where control planes live: 63% of major AWS incidents involve us-east-1, and roughly half of Azure and GCP majors involve global layers (identity, networking, API management) that cross every region at once.
68% of corpus incidents breached at least one monthly SLA target, yet practitioner estimates put claim rates below 10%. With claim windows of 30–60 days, most owed credits expire quietly.
68%
of major incidents breached at least one monthly SLA target
State of Cloud Reliability 2026
<10%
estimated share of owed credits actually claimed (practitioner estimate — see methodology)
State of Cloud Reliability 2026
Claim windows run 30 days (Google Cloud) to ~60 days (AWS, Azure). Independent tooling exists for the math — e.g. the SLA credit calculator and per-provider recovery playbooks — but the structural fix is treating the SLA check as part of incident closure.
Expect fewer but wider incidents: consolidation of internal platforms means single faults reach more services (the 2025 DynamoDB DNS and Service Control events are the template). Architectural independence from provider control planes becomes the differentiating resilience investment.
The full chapter — including the consolidation analysis behind the fewer-but-wider thesis and provider-by-provider outlooks — ships in the complete report. Get the PDF below; the research blog carries updates between editions.
A branded PDF of the key findings, provider tables, and chaptered summaries with a copy-ready citation. Free — no email required.
Download the PDFPrefer it emailed — plus monthly reliability research?