Regarding using a graphics card as a physics card in addition to being a graphics card...is there enough spare processing power and memory? I'm not convinced that there is, particularly since any spare capacity on a graphics card can be used for more realistic graphics.
The PhysX PPU contains about 125 million transistors and is designed specifically for that one task. The cards also have 128MB of their own memory. I am not convinced that the card can be adequately replaced by whatever spare processing power and memory is available on a graphics card. It is not at all common that a complex specialist piece of kit can be replaced by spare capacity on something else.
I think we'll have a 3-tier system in a while. Low physics with just a CPU, low-medium with a powerful enough graphics card (which would probably also require reducing graphics settings) and high physics with a physics card (which will probably be a choice of PhysX cards, since I don't see any other company in the market).
The next question is - will games developers consider it worth the extra time needed to develop a game to run at 3 different levels of physics?
I think they probably will in terms of realism, e.g. more complex particle effects, more accurate deformation and destruction of game objects (like the tower collapsing in the demo video from the website) and stuff like that. That is relatively easy to implement in three levels, as it doesn't affect the game. PCs with more physics power will get a more realistic-looking game, PCs with less physics power will get a less realistic-looking game, but it will be the same game. It's analogous to graphics settings today.
I think they probably won't in terms of what is probably the greatest potential of a physics card - player-gameworld interactivity. I don't think we'll be seeing much of things such as the ability to pick up a gameworld rock, use it as a tool to break off a gameworld branch from a gameworld tree and use that as a club, except in the form in which it currently exists, i.e. in a limited context where written in by the programmer. That sort of thing affects the game itself, so the game developer has several bad options. Firstly, create a game which only runs on a PC with a physics card. Secondly, create a game in which players on PCs with a physics card have a massive advantage over players on a PC without a graphics card. You'd essentially have to create two versions of the game and keep them separate. Not just for multiplayer (which could be done by using different servers) but also for single-player, because the two versions of the game would have to be balanced very differently. Thirdly, don't support physics cards at all.