Latency has nothing to do with how full the RAM is. Again the point of RAm is that the latency is the same regardless of how the memory is ordered. The latency of a memory access will alwayse be roughly the same.

Unless you are talking about badwidth, which I guess given your "instantly call everything from the RAM" comment you are actually referring too. Of course reading 5GB of data from memory will take longer than reading 5MB. That is decided by memory bandwidth which for the PS4 for example is 176 GB/s which means each second you could read all 8GB of RAM 22 times every second. Of course having more data stored in RAM at once doesn't actually mean you have to use more bandwidth, it's actually the opposite. If you are using paging and swapping stuff into and out of memory constantly that consumes bandwidth. If however you fill all available RAM with data you don't have to swap out as much because more is pre-loaded. And for example for each room in a game if you only keep the lets say 1GB of textures and models etc you need in memory and then for the next frame you unload the lets say 100MB of textures you don't need and load in 200MB of new assets you need that takes extra time and bandwidth that you could be using for other things like say frame buffer effects, however if you pre load all 5GB (both consoles reserve around 3GB for the system so games only get around 5) of assets for the entire building you then don't have to waste bandwidth loading and unloading assets each time you move to a new room. And boom you have made your game more optimised by eliminating redundant asset loading and filled up all available RAM.

Makes you wonder how the PC market lasted this long. Thank mammon for hardware tiers!

But we are not discussing PC here. Devs are happy that they can work with fixed hardware. What shall they do if an expansion is released? Build two versions of the game like on PC? On PC you let "magic" do the work for you (drivers, apis) so you don't have to care for every cache-miss. On consoles it is especially wanted that there is only one known configuration so you can optimize with the given limits. If it's multiplat you have to optimize at least two times. But to optimize twice on each console? Especially by, let's say, twice the ram. You would have two different design goals. This would do more harm than be beneficial.

Just bundle the upgrade with the game that requires it.

Given how much Crytek sucks at optimization; I'll take someone else's word for it... Besides that though, of course console hardware will be a limiting factor for game development, that has always been the case.

Wouldn't that raise the costs of the game ? Packing in some components aren't cheap ... 

I know man, look at all them extortionate bundle prices.