Think of it this way: if the low-millisecond latency internet is good enough for multiplayer FPSs, i.e. exchange of player/missile coordinates, then why not use it for other aspects of game data too?
The article may have been a bit fuzzy in its discussion in general, but I believe it communicated clearly enough that gfx are on the GPU and longer latency stuff such as AI (which is _not_ computed in nanoseconds, nor does it need to be completely deterministic) can be done via external network interface.
Cloud AI has all kinds of other advantages, too, such as dynamically scaling difficulty, which we already see examples of in Starcraft 2...
OnLive, despite failing, was not dissimilar to "cloud rendered graphics" and it worked. Lots of things people thought impossible 10 years ago are now commonplace.
The basic question is, is it more efficient and overall cheaper to have #(max. concurrent clients) times the hardware running in a hosted (remote) environment than having #(all clients) times the hardware running locally. Clearly #(max. concurrent clients) is much lower than #(all clients). MS, as the owner of the platform is in a pretty good position of balancing between thin and thick client and forcing developers to work within this model, ie. making the client thin (read: cheap) enough that critical calculations need to be done on the server but thick enough to run some latency critical stuff and probably some of the rendering on the client.
If my SimCity simulation needs more than a quad i7 to simulate, how many high-CPU instances is that? How many to handle every connected user? Even EA's current SimCity instances on EC2 can't handle basic save operations.
[1] http://www.wpcentral.com/xbox-one-no-bc-not-always-online
So, will it need always online? In that case, here is a console I will never buy.