Cloud Outages, 2020 to 2026: A Timeline of Who Pays First
Archi4 min read
Contents
- The timeline at a glance
- 2020: the year of the shared control plane
- 2021: DNS and the long tail
- 2022: capacity, not code
- 2023: the configuration era
- 2024: the noisy neighbor returns
- 2025: supply chain meets cloud
- 2026 so far: the bill arrives
- The pattern behind the dates
- 1. Tenants without an exit
- 2. Teams without on-call depth
- 3. Regulated and time-sensitive businesses
- What the timeline suggests for the next twelve months
The timeline at a glance#
Cloud failures get reported as engineering stories. The more useful reading is a distribution story: when a region goes dark, the damage does not land evenly. Below is a rough chronology of the last six years, drawn from public postmortems and incident reports, with an eye on who carried the cost.
2020: the year of the shared control plane#
When a major provider's identity and authentication service wobbled in late 2020, the blast radius reached far past its own customers. Companies running on other clouds found their single sign-on and support tooling dependent on the same vendor. The most affected group was not the giant tenants. It was small teams with no second identity provider and no staff on call.
A pattern showed up early: enterprises with dedicated platform engineers rode out the day. Startups doing their own infrastructure work did not sleep.
2021: DNS and the long tail#
A DNS provider outage took a large slice of the consumer web offline for roughly an hour. Retailers lost checkout windows. Media sites lost ad revenue. But the sharpest pain was among businesses whose entire uptime story rested on one DNS vendor with a default TTL and no secondary.
Affected most: small ecommerce operators and anyone selling time-sensitive inventory. Affected less than headlines suggested: large retailers with multi-provider DNS, who saw elevated errors but kept taking orders.
2022: capacity, not code#
Several regions saw capacity crunches as demand for GPU and general compute climbed. This was less dramatic than an outage and more expensive for a specific group: teams on committed-use discounts who could not burst, and research groups renting spot capacity that vanished mid-job.
Affected most: cost-sensitive workloads with no checkpointing. A training run that dies at hour nine is a budget event, not an engineering one.
2023: the configuration era#
A run of incidents traced back to permission changes and network policy edits rather than hardware. One widely discussed case involved a storage service where a configuration push removed access to a large volume of customer data for hours.
Affected most: regulated industries. Banks, health providers, and insurers had the least room to absorb downtime and the most paperwork afterward. A four-hour outage became a six-week evidence exercise for compliance teams.
2024: the noisy neighbor returns#
Multi-tenant noise came back into focus as providers packed more workloads onto shared hardware. Incidents ranged from degraded networking to storage latency spikes. Public postmortems rarely named the tenants involved, but the pattern was consistent: workloads without resource isolation felt it first.
Affected most: teams running latency-sensitive services, and anyone who assumed a managed database's performance tier was a guarantee rather than a target.
2025: supply chain meets cloud#
A cluster of incidents involved dependencies outside the provider's own walls: a certificate authority mistake, a package registry compromise, a third-party API that failed closed instead of open. Each one showed how thin the layer of trust between a business and its customers has become.
Affected most: companies with deep dependency trees and shallow inventories. If you cannot list your critical third parties in an afternoon, you are exposed.
2026 so far: the bill arrives#
The current stretch has been quieter on headline outages and louder on cost. Several providers repriced egress and committed-use terms. Finance teams that had treated cloud spend as a fixed line item started asking platform engineers for forecasts.
Affected most: mid-sized companies. Large enterprises have reserved capacity and negotiating room. Small startups can move. The middle gets squeezed: too big to replatform on a weekend, too small to command a custom contract.
The pattern behind the dates#
Three groups keep appearing on the losing side.
1. Tenants without an exit#
Every incident above punished the same behavior: depending on one provider, one DNS vendor, one identity service, with no tested fallback. Exit plans are cheap to write and expensive to test. Most teams write them and never test them.
2. Teams without on-call depth#
A single engineer who knows the failover procedure is a single point of failure dressed as a person. The 2020 and 2023 incidents both separated organizations that could hand off at 3 a.m. from those that could not.
3. Regulated and time-sensitive businesses#
A bank and a hobby site can both go down for two hours. Only one of them has to explain it to a regulator. Time-sensitive businesses, from ticketing to payments to logistics, lose revenue in the same window that other sites lose page views.
What the timeline suggests for the next twelve months#
Expect fewer dramatic regional outages and more incidents that start in a dependency, a certificate, or a billing change. The groups most exposed are the ones that have not done the boring work: inventory your third parties, test your failover, write down who gets paged, and know your egress numbers before finance asks.
The cloud did not get less reliable over six years. The failure modes moved from hardware to configuration to contracts, and the people who feel them first have not changed much at all.
More articles