Blog
/ Network | Cloud | Colocation

Public Cloud Outages: What the Data Shows and How Ark Approaches Resilience

10/30/2026

Public cloud adoption has made enterprise IT more flexible and scalable than ever. It has also concentrated a large share of the world’s digital infrastructure in a small number of providers and regions. When one of those platforms has a problem, the effects reach well beyond a single company. Over the past year, AWS, Microsoft Azure, Google Cloud and Cloudflare have each had major incidents that disrupted thousands of downstream businesses. These events are worth a close look, because the patterns behind them apply to any organization that depends on cloud services.

How Often Major Outages Happen

Control Plane maintains a public tracker of major incidents at the four largest platforms since 2020. It is built from the providers’ own post-incident reports and documents 62 major outages in that period:

Across all 62, the median outage lasted just over six hours (Control Plane).

These counts reflect how much each provider discloses publicly, not how reliable each one is. Microsoft, for example, publishes reviews for more incidents than its peers. Still, multi-hour outages at the largest platforms happen several times a year across the industry.

The past twelve months illustrate the range:

  • AWS, October 2025. A software defect in DynamoDB’s automated DNS management disrupted the us-east-1 region for more than 14 hours. Some global features were affected too, including IAM sign-in to the AWS console, because they depend on that region (AWS).
  • Azure, October 2025. An invalid configuration took down Front Door, Microsoft’s global content delivery and routing layer, for more than eight hours.
  • Cloudflare, November 2025. A routine database permissions change produced an oversized internal file that crashed traffic-handling software across its network (Cloudflare).
  • AWS, 2026. In May, a thermal event in a Northern Virginia data center took about 21 hours to largely recover from. Separately, physical damage to facilities in the Middle East has left two AWS regions impaired since March.
  • Google Cloud, October 2026. A power event in a London availability zone disrupted services for roughly eight hours.

What Causes Them

In the Control Plane dataset, 32% of major outages came from configuration or deployment changes. Another 23% came from software bugs or automation errors, and 18% from power or cooling failures. Network issues, capacity problems, DNS failures and physical damage make up most of the rest.

A few patterns show up again and again.

Scale amplifies small mistakes. Hyperscale platforms are built to push changes across global fleets quickly. That speed also lets a single faulty change reach every location before anyone can stop it.

Dependencies are often invisible. A workload deployed in one region may still rely on identity, DNS or management services hosted elsewhere. Nine of AWS’s 14 major outages in the tracker occurred in us-east-1, which hosts some of those shared functions.

Physical infrastructure still matters. Power and cooling failures tend to produce the longest outages, because recovery means restoring physical systems rather than rolling back software.

Concentration turns local failures into broad ones. When many organizations run in the same regions and rely on the same edge and identity services, a single incident can affect entire sectors at once.

The Business Impact

Uptime Institute’s 2026 outage analysis found that outage rates per site have declined for five straight years, so facilities are getting more reliable. The cost of each outage, however, keeps climbing. In Uptime’s 2025 survey, 57% of operators said their most recent major outage cost more than $100,000, and one in five reported costs above $1 million. Uptime also found that third-party providers account for about two-thirds of publicly reported outages. That group includes cloud platforms, telecom carriers and colocation companies (Uptime Institute).

For the October 2025 AWS event alone, Parametrix estimated losses to U.S. companies of $500 million to $650 million (Parametrix).

Regulators and boards are paying closer attention, too. More organizations are being asked to show that they have planned and tested for the failure of a critical technology provider.

What Resilient Architecture Looks Like

Most organizations are converging on a few practical steps.

  1. Map every dependency. Know which services your critical workloads rely on and where those services physically run.
  2. Spread risk deliberately. Distribute workloads across locations and providers so that no single failure takes everything down.
  3. Keep core systems on dedicated infrastructure. Run critical systems on infrastructure the organization controls, with direct connections to public cloud.
  4. Keep workloads portable. Applications should be able to move between environments when needed.
  5. Test failure scenarios regularly. Rehearse realistic outages end to end rather than assuming the plan will work as written.

Few companies are moving away from public cloud. They are treating it as one part of a planned infrastructure strategy rather than the whole of it.

How Ark Approaches these Challenges

At Ark, these principles shape how we design and operate our infrastructure.

We build in distributed markets. Our enterprise data centers are in secondary and edge markets across the United States, outside the few metro areas where hyperscale capacity is most concentrated. That puts infrastructure closer to where our customers’ data is generated. It also gives organizations a way to diversify geographically, away from the regions most exposed to large-scale incidents.

We deliver hybrid infrastructure from one provider. Customers can combine colocation, private and multi-tenant cloud, network connectivity and managed security with us. This year we launched Ark Cloud on the Nutanix platform, which includes built-in disaster recovery along with managed backup and security monitoring. Customers can keep critical systems on dedicated infrastructure and still connect to the public cloud services they already use.

We invest in network diversity. Alongside our data centers we operate fiber networks and internet exchanges, including exchanges we launched in Boise and Tucson. They give customers more direct and varied paths between their sites, cloud platforms and carriers. With connectivity-related outages on the rise, this matters as much as power and cooling.

We design for geographic redundancy. With facilities in multiple markets, customers can run primary workloads in one Ark data center and keep redundancy in another, with the same operating model at both.

We grow thoughtfully. As our customers’ needs change, we expand mainly within our existing campuses rather than developing new sites. We also prioritize energy efficiency; our Akron facility, for example, has earned ENERGY STAR certification. Our aim is to add capacity in step with real customer demand while being a good long-term neighbor in the communities where we operate. 

Public cloud outages will keep happening; they come with the scale and complexity of these platforms. What organizations can control is how much an outage affects them. Our job is to give enterprises the infrastructure, connectivity and expertise to stay up and running when a single provider or region goes down.

Related Resources