Jigawatt
Arcane
Because it is really fucking hard and new like most of SC things when it comes to tech.
Basically once they implement it player instancing will be limited only by game designers not by technology. So bigger ship will have its own server and if it will reach some huge number of players it will have its second or third server all dynamically done without any problem for player who won't notice people being de-spawned and so on. So space battles will be limited mostly by hardware of players who will have to render shitload of ships.
Right now servers are static.
It really isn't that hard though - at least it shouldn't be for people who claim to be experts in networking. I wouldn't call what you're describing a solved problem but it's had a lot of thought put into it for years - as an example see this documentation from Unreal
Relevancy
A level can be huge, and at any time a player can only see a small fraction of the Actors in that level. Most of the other Actors in the level are not visible, are not audible, and have no significant effect on the player. The set of Actors that a server deems are visible to or capable of affecting a client are deemed the relevant set of Actors for that client. A significant bandwidth optimization in Unreal's network code is that the server only tells clients about Actors in that client's relevant set.
Unreal applies the following rules (in order) in determining the relevant set of Actors for a player. These tests are implemented in the virtual function AActor::IsNetRelevantFor().
... rules follow
Seems to map pretty nicely to what you're describing, and it's been that way since at least UDK3 which was 15 years ago.
In SC it is the other way. Each character has a server of its own and if you go somewhere you still have your own server following you. When you meet someone then this is where your server or his server takes your instance or his and combine (meshing) on the fly. This frees up one server leaving both player using only one server.
This too doesn't seem that complicated, and there are off the shelf solutions to it. Off the top of my head I could see something like Firebase used to orchestrate this 'meshing' (bullshit term btw) and then traditional in-engine networking once the right connections are made. For a small team it's a < 1 month job to roll out something usable, I'd be embarrassed to estimate higher than 3 months.