Cloud Reliability Comparison
AWS vs Azure vs GCP: Outage Frequency & History Compared
Updated September 2026 · 28 documented major incidents across three clouds
Outage Frequency, In Brief
By documented major incidents, the three hyperscalers cluster around one to two significant outages per year each. Our database records 10 major AWS incidents since 2012, 9 for Azure since 2018, and 9 for GCP since 2016. Raw counts are not a clean reliability ranking - they reflect how long each provider has been tracked and how visible each event was. AWS outages concentrate in us-east-1 and affect the most customers per event; Azure incidents often hit Microsoft 365 and Teams; GCP fails less often but its global control plane lets a single failure scale worldwide within minutes (June 2025). Full side-by-side table, combined timeline, and SLA credit comparison below.
Outage Frequency & History, Side by Side
| Metric | AWS | Azure | GCP |
|---|---|---|---|
| Documented major incidents | 10 | 9 | 9 |
| Tracking window | since 2012 | since 2018 | since 2016 |
| Approx. frequency | ~1 / yr | ~1-2 / yr | ~1 / yr |
| Costliest incident | S3 Feb 2017 ~$150M (Cyence); Dec 2021 est. similar | Oct 2025 Front Door global outage (~8.5h) | Jun 2025 Service Control / IAM (dozens of products) |
| Most recent major | Jul 24 2026 us-west-2 network outage (~80 min; Apple Pay, Reddit, DoorDash) | Feb 2026 multi-region control-plane failure (~12h) | Sep 1 2026 us-central1 fiber-disconnection (~4h 11m, 19 products) |
| SLA credit range | 10% -> 30% (EC2, <99%) | 25% -> 100% (<99%) | 10% -> 50% (Compute, <95%) |
| Failure pattern | Concentrated in us-east-1; affects the most customers per event | Often hits Microsoft 365 and Teams - high enterprise visibility | Rarer, but a global control plane means failures scale worldwide fast |
Incident counts are the major outages documented on each provider page below; they are not exhaustive logs of every status-page event. No cloud provider discloses customer-impact figures for its own outages, and for eleven of the thirteen events in the timeline below no named analyst or insurance-modelling firm has published one either, so those read "None published" rather than carrying a guess. The two that do are October 2025 (CyberCube's preliminary insured-loss range) and July 2024 CrowdStrike (Parametrix, US Fortune 500 only). To size an event against your own revenue, the AWS and GCP chronologies show revenue at risk per $100M of annual revenue with the arithmetic shown. Updated October 2026.
Combined Outage Timeline (Most Costly, Recent)
| Date | Provider | What Happened | Published Cost |
|---|---|---|---|
| Sep 2026 | GCP | us-central1-b (Iowa) fiber-disconnection during maintenance; all fibers pulled at once, 19 products degraded ~4h 11m | None published |
| Jul 24 2026 | AWS | us-west-2 (Oregon) network path to Seattle metro failed ~80 min; cross-region traffic timed out; Apple Pay, Reddit, DoorDash, Hulu, PlayStation Network offline | None published |
| Jul 16 2026 | AWS | CloudFront VPC Origins control-plane failure; Frankfurt AZ (euc1-az2) capacity limit served global 5xx ~3.5h; Canvas, Blackboard, Hugging Face, UK National Lottery hit | None published |
| May 2026 | AWS | us-east-1 (use1-az4) data-center cooling/thermal event; EC2/EBS, ELB, EKS, Redshift, MSK offline ~28h; Coinbase down ~7h | None published |
| Feb 2026 | Azure | Remediation workflow disabled anonymous read on storage hosting VM extension packages; VM/VMSS/AKS provisioning failed multi-region ~12h, then Managed Identities token failures in East US & West US | None published |
| Nov 2025 | Cloudflare | Config-propagation failure degraded a large share of the web (~2.4B users routed through Cloudflare) | None published |
| Oct 2025 | AWS | DynamoDB DNS race condition cascaded to EC2/Lambda/STS; ~70K orgs | $38-581M insured (CyberCube) |
| Oct 2025 | Azure | Front Door data-plane config change failed globally (~8.5h); cascaded to M365, Outlook, Copilot, Azure Portal, Xbox; Alaska Airlines, Costco, Starbucks hit | None published |
| Jun 2025 | GCP | Invalid quota policy crashed Service Control; IAM auth failed across dozens of products; cascaded to Cloudflare, Spotify, Discord, OpenAI | None published |
| Jul 2024 | CrowdStrike / Azure | Falcon sensor update crashed Windows hosts globally, incl. Azure VMs; airlines, banks, hospitals hit | $5.4B (US Fortune 500, Parametrix) |
| Jan 2023 | Azure | Global WAN/BGP router config change broke connectivity across Azure and Microsoft 365 (Teams, Outlook, SharePoint) worldwide (~5h) | None published |
| Dec 2021 | AWS | us-east-1 EC2/ECS/Lambda/SNS outage (~7h), major global disruption | None published |
| Feb 2017 | AWS | S3 us-east-1 outage (~4h) disrupted a large portion of the internet | $150M, S&P 500 (Cyence) |
Per-provider chronologies: AWS, Azure, GCP. CrowdStrike was a software-update incident, not a cloud-provider failure, but it crashed Windows hosts including Azure VMs - see the CrowdStrike case study.
SLA Credits: What Each Cloud Actually Pays
| Provider (service) | Lower-tier credit | Maximum credit | Claim window |
|---|---|---|---|
| AWS (EC2) | 10% (<99.99%) | 30% (<99%) | 30 days |
| Azure (most services) | 25% (99%-99.99%) | 100% (<99%) | Auto for some since 2024 |
| GCP (Compute Engine) | 10% (99%-99.95%) | 50% (<95%) | 30 days |
All credits are a percentage of the monthly fee for the affected service, applied to future invoices - never your revenue loss. Full SLA credit vs actual loss analysis, or calculate your own exposure.