You're assuming AMD could even pay for it if they wanted to. They're cash poor atm and I doubt Nvidia would even license it to them if they did try to buy rights to it. Doing so would require giving AMD access to low level CUDA technology and software, something Nvidia absolutely does not want to do. AMD already eats their lunch in OpenCL (Nvidia's own fault since it's an artificial handicap) so there's no way in hell Nvidia would give them a way to suddenly be able to run CUDA too. It would put a big dent in their TESLA profits if consumer level cards from the competition suddenly had the ability to suddenly run CUDA code.
Nvidia pitched physx to both intel and AMD, both declined.