I've done some work with both PhysX and the things that AMD is pushing for. I try to keep with the Physics Abstraction Layer, which lets me plug in whatever physics engine as the backend, which gives a pretty damn good apples-to-apples performance metric. Personally, my ultimate choice of physics engine is the one which exhibits the best performance. My experience may differ from others, but I generally get the best performance from PhysX on with an nVidia GPU and BulletPhysics with an AMD GPU. Sometimes, the software version of PhysX outstrips the competition, but I have never seen anything beat PhysX in performance with GPU acceleration turned on. And with PAL, it is easy to check if there is GPU support on the machine and swap in the physics engine with the best performance (PAL is awesome).
Here's the thing: GPU-accelerated physics are just plain faster. Why? Because collision detection is a highly parallelizable problem. Guess what hardware we have that can help? The GPU. Another great part of using the GPU is that it frees the CPU to do more random crap (like AI or parsing the horribly slow scripting language).
AMD is working on both BulletPhysics and Havok so they can do GPU acceleration. But I have a feeling that PhysX performance will remain faster for a while: PhysX was designed to natively run on the GPU (technically, a GPU-like device), while these other libraries are not. Furthermore, nVidia has quite a head start in performance tuning, optimization and simple experience. In five years, that shouldn't matter, but I'm just saying that it will take a while.
So here is my message to AMD: If you want people to use your stuff, make something that works and let me test it out in my applications. You've released a demo of Havok with GPU acceleration. PhysX has been and continues to work with GPU acceleration on nVidia GPUs and will frequently outperform the software implementation. I'm all for open alternatives, but in this case, the open alternatives aren't good enough.