Isn't that the same as loading them from file on each request?
No, because you don't have additional IO overhead.
I think that overhead would be negligible, especially if the other examples are all based on micro-frameworks (i.e. loads of files).
Premature optimisation is the root of all evil, right?
Disk IO is a massive problem at scale. This obviously isn't "at scale", so it works fine here, but avoiding hitting the disk (edit: any slow IO, really) is Performance 101.
You would also want to install an opcode cache like APC, XCache, etc. That way the code is only compiled on the first page load.