From a hardware perspective, I have no doubt that dedicated physics hardware (like physX cards) are dead in the water.
the direction of CPU development is now largely parallel, and particularly in games this is something which is very difficult to take full advantage of. Physics calculations can be easily parallelised, and so can take advantage of 'idle' CPU cores.
Similarly, GPUs are (and always have been) massively parallel, so can be (relatively) easily modified to run physics-type calculations. So, having another bit of hardware specialised to do physics (which is basically just another floating point co-processor - just like a graphics card) makes little sense when you already have a bunch of un-tapped parallel potential sitting idle in your CPU and GPU.
I am writing a "physics" code to take advantage of the GPU through CUDA (nvidia's API for using the GF8 for general floating point calculations). It's an engineering code (computational fluid dynamics) rather than a games-related engine, but the principles are the same. Anyway, the GPU has orders of magnitude more power for such calculations than a CPU, but in the end I think CPUs will win out for game physics, because:
a) There will be no conflicting standards (GPU acceleration would almost certainly require games developers to write two different physics implementations for ATI and nvidia cards)
b) future multi-core processors (8 and 16 cores) will have more than enough spare capacity for high en physics
c) Programming for the CPU allows more flexibility than programming for specialised FPU accelerators (like a GPU or physics card). The CPU is more more "forgiving" of poor coding practices.
d) Programming for CPUs will not require extensive re-education for programmers. They can use the compilers and other tools that they are already familiar with.
Anyway, that's my 2p. I think the GPU has massive potential for general FPU calculations, but I think that the (relative) simplicity of programming for multi-core CPUs will win out. Physics-specific hardware is an unhappy compromise, since it has all the drawbacks of GPU implementation (specialised parallel architecture that is difficult to program, along with not having the raw power of a GPU), and also requires the user to buy an extra bit of hardware.