The only thing that could work is if a game developer... It actually could be done even now via openCL...
Uses game effect data , such as FXAA(MSAA not possible) and some physics(HAVOK CPU does fine IMHO useless) or lighting reflection processing, fluids... that is actually blocked feature for nvidias department, then using the other card as processing cluster... but it ain't something new really... But again... accessing it to other card's driver kernel ain't a really fast solution, it still introduces lag(we can see it when enabling hybrid Physx), no matter how there will be two drivers that have their own capricious characters. The overall goodness in this is questionable really...