Interresting questions.
I think the core code would have to be run by the CPU. Cuda get's it's benifit from massive parrallism (sic) which in turn requires an algorithm designed to take advantage from it. I think it would be hard to keep our game universe consistent (in multiplayer games) if different PC's were uising different algorithms to calculate their 'game physics'
Of course the eye candy could take advantage from it giving a better experience for people with PhysX.
(Developers, please, cut in at any time and give use the low down

)