I think one response that's missing here is that application servers might use a lot of RAM while they're running, so you can only afford to have a few instances running. In such cases, it's better to have nginx buffer the response and deal with the slow clients, than holding up an application server for the duration of the transfer.