As an example, consider the PCIe bus controller. Say it had some non-volatile memory so that when it powers on, it just reads the bus configuration from the NVM instead of enumerating and initializing the devices on the bus. Lots of time saved.

Well that wouldn't really work without the PCIe devices also storing the configuration in NVM. If I plug a x16 device into a x4-only slot (as is typical for the third PCIe x16 slot on motherboards that have it), the device needs to know it has to only use 4 lanes, not the full 16.

Ok, so say the controller has some way to tell the devices to save their state, and when powering on they all read from NVM and can just go on talking with each other right away, no further initialization required. Yay!

Except then you power off your machine, and swap that GPU for the new one you bought from your friend. Now the new GPU has the configuration from your friends motherboard, and your bus controller has no idea about this new card... so this is doomed to fail.

Of course this could be worked around by querying the bus for the devices we expect to be there... except that's the initialization step we were trying to avoid in the first place...