Anything under like 100ms load is instant to the user, especially a page load.

True, but it accumulates. Every ms I save on templating I can “waste” on I/O, DB, upstream service calls, etc.

If you run it in old-school CGI mode, no, because each request would spawn a new process. But that’s nowhere near state-of-the-art. So typically you would still have a long-running process somewhere that could manage a connection pool. No idea if it does, though. Can’t imagine that it wouldn’t, however, since PHP would be slaughtered in benchmarks if there was no way to keep connections (or pools) open across requests.