The 2026 Cloud Outage Report: What 62 Major Outages Reveal

Published October 26, 2026 by Control Plane. Data: the Control Plane Cloud Outage Tracker, 62 incidents from January 2020 to October 2026, each linked to the provider’s own post-mortem or status page.

Most major cloud outages start with the providers’ own changes and software, not hardware. Of the 62 major AWS, Microsoft Azure, Google Cloud, and Cloudflare outages we documented since 2020, 34 (55%) started with a configuration or deployment change, a software bug, or automation acting on bad input. But the outages that last longest are physical: power and cooling failures ran a median of 21h 48m, about four times longer than configuration changes (5h 21m).

Key findings

• 55% of major outages start with a change or code. 34 of 62 began with a configuration or deployment change, a software bug, or automation acting on bad input.
• 2025 was the year of bad changes. 10 of 2025’s 16 outages were configuration or deployment changes, including Cloudflare’s November 18 and December 5 incidents and Azure Front Door on October 29.
• Physical failures stand out in 2026. 7 of 2026’s 18 outages so far were power, cooling, or physical damage, from lightning and transformer failures to a thermal event in AWS us-east-1.
• Power and cooling outages last about four times longer. Median: 21h 48m (8 timed incidents) versus 5h 21m for configuration changes. The median across all outages is 6h 12m.
• AWS us-east-1 appears most often. It is in 9 of the 62 incidents, 3 of them limited to one Availability Zone; no other region appears more than twice.
• Half of outages were not confined to one region. 31 of 62 hit more than one region or a provider’s global services. Among AWS, Azure, and Google Cloud alone, 18 of 46 (39%) did, so running in several regions of one provider may not be enough.
• Two AWS regions have been disrupted since March 2026. me-central-1 (UAE) and me-south-1 (Bahrain) are still listed as open on the AWS Health Dashboard.

How we built this report

We collected every major, customer-facing outage that AWS, Microsoft Azure, Google Cloud, and Cloudflare documented publicly from January 1, 2020 to October 26, 2026. 49 of the 62 come with an official post-mortem; the rest come from a provider status page or, where no provider record exists, a credible news report, marked as such in the tracker. Durations use each provider’s own start and end times and measure the full event window from first customer impact to recovery. Median durations use the 55 completed incidents with a provider-sourced duration; they leave out the two ongoing AWS Middle East events and incidents timed only by news reports. Root causes are paraphrased from the provider’s report and grouped into nine categories.

This is not a reliability ranking. Providers document incidents differently: Microsoft publishes reviews for more incidents than the others, so the tracker includes 20 selected Azure incidents rather than every Azure post-incident review. Coverage also varies by year, so year-over-year counts are a floor, not a precise trend.

Finding 1: Most outages start with a change

Configuration pushes, deployments, and automation acting on bad input caused more major outages (34) than power and cooling, physical damage, and network faults combined (18).

The pattern repeats across providers: a change that passed testing meets production data, configuration, or scale it was never tested against, and it propagates faster than humans can stop it. Cloudflare’s November 18, 2025 outage started with a database permissions change that doubled a feature file past a size limit. Google Cloud’s June 12, 2025 outage started with a policy change with blank fields that replicated globally.

Finding 2: Physical failures are rarer but last far longer

Power and cooling caused 11 of the 62 outages, and the 8 with a provider-stated duration had a median of 21h 48m. Configuration and deployment changes, the most common cause, had a median of 5h 21m. Changes can usually be rolled back; a failed transformer or an overheated data hall has to be repaired.

Physical failures stand out in 2026. 5 of the year’s 18 outages so far were power or cooling failures:

• Azure, Feb 7-8, 2026, West US: A transformer failure cut utility power and a control-system fault blocked the generator switch, draining UPS batteries.
• AWS, May 7-8, 2026, us-east-1 (N. Virginia), one AZ: A thermal event in one data center caused power loss, and servers shut down automatically when temperatures passed safe limits.
• Azure, May 29-30, 2026, West US 2: Lightning-driven voltage swings locked out cooling systems, and rising heat forced infrastructure shutdowns across two zones.
• Google Cloud, Jul 15-16, 2026, europe-west4 (Netherlands): A brief utility voltage drop tripped both power feeds, a backup unit failed, and a cooling failure followed.
• Google Cloud, Oct 19-20, 2026, europe-west2 (London): Google cites a power event in one London zone; a full analysis has not yet been published.

Add the two AWS Middle East regions disrupted by physical damage since March, and 7 of 2026’s 18 outages were physical.

Finding 3: AWS us-east-1 appears most often

Northern Virginia appears in 9 of the 62 incidents, 3 of them limited to one Availability Zone; no other region appears more than twice. Press coverage concentrates on us-east-1, which may inflate its count. It is AWS’s oldest region, and some AWS global features run through it, which is why a us-east-1 problem can affect customers who never deployed there.

The October 19-20, 2025 outage lasted 14h 32m across more than 100 AWS services. Parametrix estimated US companies lost $500 million to $650 million.

Finding 4: Half of outages crossed regions

31 of the 62 outages hit more than one region or a provider’s global services: 13 at Cloudflare, whose network is global by design, and 18 at the hyperscalers (13 at Azure). Running an application in two regions of the same provider protects against a regional failure, but not against a failure in that provider’s global control plane, identity, or edge services.

That is the case for spreading critical workloads across more than one provider, or at least removing single-provider global dependencies such as DNS, identity, and CDN from the critical path.

Finding 5: Outages by year

2026 has already passed 2025’s total, 18 to 16. Coverage varies by year and by which incidents are included, so the count is a floor, not a trend line. Causes shifted too, from configuration changes in 2025 to more physical failures in 2026, though this also depends on which incidents are included.

Finding 6: By provider

Cloudflare outages are short (median 1h 15m) but often global; hyperscaler outages are usually regional but last hours. Counts reflect what each provider documents, not how reliable it is.

The longest outages in the tracker

• Azure, East US 2 (Jan 8-11, 2025): 50h 13m. While enabling a feature, networking PubSub replicas were recycled in parallel, breaking quorum and losing index data.
• Cloudflare, Hillsboro, Oregon (control plane) (Nov 2-4, 2023): 40h 42m. Utility problems, a ground fault and generator failures cut power to a data center hosting control-plane services not in the HA cluster.
• Google Cloud, europe-west2 (London) (Jul 19-20, 2022): 38h 47m. Redundant cooling systems failed together during record UK heat, so Google powered down part of a zone.
• Azure, Southeast Asia, East Asia (Feb 7-9, 2023): 32h 11m. A voltage dip tripped chillers in one zone, forcing shutdowns, and regional failover and ARM dependencies also failed.
• Google Cloud, europe-west9 (Paris) (Apr 25-26, 2023): 27h 14m. A cooling water leak flooded a UPS room and started a fire, and the building shutdown took the region’s control plane down.

Cloudflare’s 2023 Hillsboro event left its edge network running, and most control-plane services were restored at a backup site within about 6 hours. Still open: AWS me-central-1 (UAE) and me-south-1 (Bahrain), disrupted since March 1 and 2, 2026 after physical damage to facilities in the region.

What outages cost

Downtime cost depends on the business, but the published research points the same way:

• 57% of operators said their most recent major outage cost more than $100,000; for the second consecutive year, 1 in 5 reported costs above $1 million. (Uptime Institute)
• One hour of downtime costs more than $300,000 for over 90% of mid-size and large enterprises; 41% put it at $1 million to over $5 million. (ITIC, 2024)
• High-impact business outages cost respondents an average of $1.85 million per hour, over $30,000 a minute; the annualized cost of high-impact outages was $73.5M in 2026, down from $76M in 2025. (New Relic, 2026; respondents skew toward large enterprises.)
• Downtime costs Global 2000 companies $400 billion annually, or 9% of profits. (Splunk and Oxford Economics; this covers all downtime, including security incidents.)

Here is a worked example from the tracker: a company with $50 million in annual revenue through its main application has a 3-hour regional outage that disrupts 80% of that revenue, half of which is lost rather than delayed. With 12 engineers responding, 150 employees unable to work, and $30,000 in SLA credits, the outage costs about $50,620. It assumes $120 an hour for engineers, $70 an hour for other staff, and 30% lost productivity. The tracker has the full formula and a calculator.

What limits the damage

The post-mortems point to a few practices that limit the impact of an outage:

• Run active-active. Serving real traffic from more than one location means failover is already exercised. A warm standby that never serves traffic is untested until the outage.
• Remove hidden single-region dependencies. Identity, DNS, secrets, CI/CD, and observability often live in one region even when the application does not.
• Plan the data layer per service, choosing replication based on how much inconsistency each service can tolerate.
• Route by health, not by hand. Health-checked global routing moves traffic in seconds; manual DNS changes take minutes to hours.
• Limit what automation can remove. Some outages grew because automation took healthy capacity out of service in response to bad signals.

During the October 2025 AWS us-east-1 outage, 20% of Control Plane customers had us-east-1 in their Global Virtual Clouds. Their workloads were active and running in healthy locations within 10 seconds, with no customer downtime, according to Control Plane internal incident telemetry. The 99.999% availability SLA applies to workloads running at least two replicas in at least two locations. Those locations can be two regions of one cloud or locations on different providers. Our guide to building active-active high availability across regions and clouds covers the architecture, and the best multi-cloud failover solutions compares the options.

Frequently asked questions

What causes most cloud outages?

Changes, not hardware. 34 of the 62 major outages in this report (55%) started with a configuration or deployment change, a software bug, or automation acting on bad input. Power and cooling failures caused 11.

How long do major cloud outages last?

The median major outage since 2020 lasted 6h 12m, measured from first customer impact to recovery using each provider’s own timeline. Power and cooling failures ran a median of 21h 48m. The longest completed outage in the tracker was Azure East US 2 on Jan 8-11, 2025, at 50h 13m.

Which cloud region has the most outages?

AWS us-east-1 in Northern Virginia appears most often: 9 of the 62 incidents, 3 of them limited to one Availability Zone. No other region appears more than twice. Press coverage concentrates on us-east-1, which may inflate its count.

Are cloud outages getting more frequent?

This report counts 18 major outages in 2026 through October 26, already more than the 16 in 2025. Coverage varies by year, so treat the count as a floor. Causes also shifted: configuration changes dominated 2025, and physical failures stand out in 2026.

Does running in multiple regions protect against cloud outages?

It protects against regional failures, which made up half of the outages here. The other 31 of 62 hit more than one region or a provider’s global services (18 of the 46 at AWS, Azure, and Google Cloud), so critical applications also need to remove single-provider global dependencies or run across more than one provider.

How much does a cloud outage cost?

57% of operators said their most recent major outage cost more than $100,000; for the second consecutive year, 1 in 5 reported costs above $1 million. The October 2025 AWS us-east-1 outage alone cost US companies an estimated $500 million to $650 million, according to Parametrix.

Cite this report

Control Plane, “The 2026 Cloud Outage Report,” October 26, 2026. Data: the Control Plane Cloud Outage Tracker, updated after each major outage. Free to cite with a link.

Next steps

See every incident, filter by provider and cause, and estimate your own downtime cost in the Cloud Outage Tracker. To keep applications online through the next one, talk to the Control Plane team or sign up free.