Actually, the Wii, 360, and PS3 recently got PhysX support. The only reason ATI hasn't implemented support is becaues they don't want to support it.
As for the whole "proprietary standards" argument, which is more popular: DirectX or OpenGL?
As I said before, Non-linear physics, which is a necessity to get accurate physics effects, requires a large data bus to enable loading of mass amounts of data using a single LOAD operation in order to execute at a reasonable speed. Regardless of how fast a CPU's IPC is, if it can't load the necessary data in a timly manner, then how fast it is capable of doing the actual computations is meaningless.
A decent Physics implementation should be around as fast as rendering is, and we know what happens to FPS when you have the CPU start to render as well...Hovok, at best, implements standard 'textbook' liner-physics formula, just like every other standalone engine out there, which leads to limitations to how far you can take physics.
For example: Destorying an individual object once created is impossible under current physics implementations, unless the object is created with breakaway zones (Company of Heros, Battlefield: Bad Company), as opposed to dynamically destroying an object as it takes damage. I want an implementation that is capable of fully dynamic destruction of a solid object, and based only at the mathematical formula required, I have come to the conclusion that a CPU based implementation simply will not be able to constantly LOAD and EXECUTE the necessary data in a timly manner.
Also remember, any data sent to the CPU needs to be stored in its CPU registers prior to any mathematic formula being applied. As those registers are limited, you may only have access to one or two registers while playing a game (due to other resources eating the other registers). Hence why the GPU has an advantage: As GPU's use a higher bus length (128/256), you can theoretically send/hold more data (by segmenting a single 256bit register, you can hold 8 32-bit integers) at a single time, cutting down on how many LOAD operations (Instruction Cycles) needed to get, load, and execute the data. Hence why I feel a large data bus is a necessity for any decent implementation of a physics engine.