I think Factorio is a better example; items still exist when you don't see them on the map, and assuming you have a somewhat optimized base the performance is very acceptable until you are around 383k items/s being created and destroyed (50kSPM in the in-game parlance).
Do they have to? Couldn't you come up with some chunking based system where the game only monitors the 'inputs' for given area and as long as those are stable, skip simulating entire chunk and only generate the output?
* a mechanism to constantly evaluate whenever a active chunk is worth promoting to a "hibernated" chunk, with its overhead and;
* a very steady state factory chunk. But endgame steady state factories already overperform your average factory by a lot in the first place.
It is the chaotic state non-endgame factories that currently need optimizations the most. And it is speficially them that don't reach steady state; one small moment of instability in the inputs and you need to start evaluating the state of your entire factory again.
I think that clustorio or the next expansion approach, where you have several "surface" maps that each have well defined limited I/O interfaces "chest or spaceships", in a way that you could split off the massive state of the game into a cluster of smaller servers is the way to go.
