Erlang will work spectacularly well in those scenarios and is explicitly aimed at them, definitely not just for 'distributed computing within a single machine' (which is nonsense anyway). The reason why it can't deal well with network splits is better explained by internalizing this:

http://en.wikipedia.org/wiki/CAP_theorem

If I remember correctly, Erlang was designed for fault-tolerant telephone switches and there were always two of them (the switches) in one box, because the system had to cope with even hardware failures.

So Erlang is pretty much designed to deal with systems distributed on different machines, it's nice behaviour on modern multicore machines is just a byproduct of that.