Video for a national audience does not travel from one server to every household. It is copied to edge locations close to viewers, and what sits at each edge is determined by demand.
The origin is a fallback, not the source
Encoded segments are stored once at an origin. Delivery networks place caching servers in metropolitan areas, and those caches serve almost every request that viewers actually make.
When an edge lacks a requested segment, it fetches it from a parent cache or the origin, stores a copy, and serves it. Subsequent viewers in that region are served locally.
The origin therefore handles a small fraction of total traffic. Its job is to be authoritative, not to be fast.
Popularity decides what stays resident
Cache space is finite, and every stored segment competes with others. Caches evict what has not been requested recently, so widely watched titles remain resident while obscure ones do not.
This produces a measurable difference in startup behavior. A popular series begins almost instantly, while a rarely requested catalog title may need a fetch from further up the chain.
Services counter this for known events by pushing content to edges in advance, a practice generally called prepositioning, so the first viewers do not pay the fetch cost.
Live traffic breaks the usual assumptions
A live event has no back catalog behavior to learn from. Every viewer requests the same newly created segment within a few seconds of its existence.
That concentration is actually favorable for caching, since one fetch to the origin can satisfy an entire region. The difficulty is the sheer request rate arriving simultaneously.
Edges handle this by collapsing identical simultaneous requests into a single upstream fetch, a behavior that prevents a live audience from overwhelming the origin.
Not everything is cacheable
Video segments are static files and cache cleanly. Manifests, which list the available segments and rungs, change constantly during a live stream and must be treated as short-lived.
License requests for protected content are personalized and never cached, which is why playback of a protected title depends on a service reachable from the viewer rather than on the nearest edge.
Advertising inserted into a stream is also personalized, so those segments follow a different path than the program around them, with its own timing consequences.
Why regional outages look like service outages
Because delivery is regional, a failure at one set of edges affects one part of the country while everyone else sees normal service.
Traffic is usually rerouted to a neighboring region automatically, which restores playback at the cost of longer round trips and more upstream fetches until caches refill.
The visible symptom for viewers is slower startup and lower sustained quality rather than a hard failure, which makes such events harder to describe than a simple outage.