When a cloud provider has an outage, services with no apparent connection fail together. The pattern reflects how infrastructure is organised beneath applications that appear independent.
Regions are the unit of concentration
Providers group data centres into regions, and customers deploy into one because keeping data close together is cheaper and faster than spreading it.
Several regions are far more heavily used than the rest, largely because they were built first and defaults still point toward them.
A fault in one of those regions therefore reaches an unusually wide set of customers, which is why some outages feel like a failure of the internet itself.
Control planes fail differently from workloads
The systems that create, configure and route resources are separate from the ones running customer code, and they are far more interconnected.
Existing workloads often keep running during a control plane failure, but nothing new can be started and nothing broken can be replaced.
That distinction explains outages where a site stays partially up yet cannot recover, because every automatic remedy depends on the component that is down.
Dependencies are deeper than they appear
An application may run in several regions while relying on authentication, storage or configuration services concentrated in one.
Those dependencies frequently go unmapped because they were adopted separately and never appear in an architecture diagram as shared risk.
Outages reveal them all at once, which is why post-incident reviews so often describe surprise at what turned out to be a single point of failure.
Recovery creates its own load
When service returns, every affected client reconnects simultaneously, and that surge can exceed the traffic the system was sized to handle.
Retry logic amplifies it further if clients retry aggressively, producing a second failure moments after the first appeared to be resolved.
Providers therefore restore capacity in stages, which is why full recovery takes considerably longer than fixing the original fault.
Redundancy is bounded by cost
Running fully across multiple regions requires duplicated data, tested failover and continuous expense for capacity that is idle most of the time.
Most organisations conclude that the cost exceeds the expected loss from rare outages, and that calculation is defensible rather than negligent.
The result is an infrastructure where concentrated failures remain possible, and where the practical question is how gracefully a service degrades rather than whether it can avoid failing.