I think most people will skip the step of buying the physics card and just wait for it to be integrated into the GPU. It will be an interesting marriage, though, since the Nvidia and Ageia architectures have some significant differences.
They are somewhat different, and I suspect the Ageia PhysX line as it is now will be discontinued. It's not exactly been selling like crazy, and I suspect nVidia are buying them out for the software, not the hardware side.
If you look at benchmarks, SLI isn't exactly what they make it out to be: it scales very badly in many games, presumably because of render-to-texture and vertex feedback rendering, which either have to be duplicated or copied between the cards.
Those types of effects are getting more common not less, so I suspect it'll become commonplace for games to do physics calculations on one GPU and the actual rendering on the other. Right now, driver support for this kind of thing isn't great, and nVidia don't have their own physics engine.
After a graphics processor, you'll buy a physics engine, and then vector math, AI, and Vision engines.
Graphics card makers know this, and will just put all those chips on a single card.