It's just not worth it. The technical back-end challenge isn't the hard part here, it's the game design. WoW with 10,000 players idling in Orgrimmar isn't fun. WoW with 100 players in a single dungeon isn't fun either, for that matter.
http://www.wired.co.uk/article/improbable-spatialos-simulate...
I don't see the need for this tech in the high fidelity city simulation cited in the linked article: that would much easier to implement as an offline simulation.
Training seems like the most likely usage: eg. disaster response or large scale military scenarios.
I have to agree with the earlier poster that tech is not the limiting factor for massively multiplayer games.
It would most likely be used to test infrastructure or policy decisions. For example, what happens to your traffic in your city if you add this bus route, or how would a contagious bioweapon spread under certain conditions.
If you want to simulate a new bus route then run an offline simulation using whatever tools you like with as much complexity as you like.
In fact you'd likely want to run your simulation much faster than real time to answer specific questions quickly, like finding traffic or passenger hotspots.
I doubt you'd want to sit there watching virtual buses driving around a 3d city, but rather quickly see heat maps change as you play with parameters.
I also wonder how it handles very large entities that might span across clusters (say a large spaceship with thousands of entities spread across multiple decks and in different area). Since all of those things are entities in a sense, and there might be many ships, how does it handle this?
No idea, but the fact that this can be asked at all is pretty interesting.