Slow startup was definitely one reason to have long lived servers, but I’m surprised not to see the major other reasons:

- Keep-alive/pooled connections to remote services can significantly reduce average latency for making those calls.

- In-memory caches that allow amortizing repeated lookups across requests.

Just those two alone mean that a serious high performance server probably couldn’t get away with just CGI even ignoring startup time.