This isn't entirely accurate though. It's only replicated if people request it. If everyone only requests the content through `ipfs.io` then the content will only be living on the original seeder, and the IPFS gateway nodes which run GC on a regular basis.
Depending on how the content is moved, it _might_ end up on a few other nodes (a way around nats, by using a third party peer to stream content between two peers), but I believe it's only ever moved through the network, and will not be stored on the nodes. If it is, it's only there till the next GC (disabled by default I believe).