The CUDA shaker?
What really matters for most HPC isn't CUDA, it isn't even OpenCL, it's what sits on top of those - the higher level numerical libraries such as BLAS, LAPACK and PETSc.
Real hardware muscle is fine, but it must provide the means to solve huge sparse matrices - as are abound in fluid dynamics and tectonic acoustics modelling - in a way that requires very little altering of source code. If I'm an expert in turbulence closure schemes for finite volume CFD (for example), it should be clear that I probably won't have the time - or unused grey matter! - to learn OpenCL or CUDA. I hedge my bets on well-established techniques which rely on software libraries that aren't going to disappear any time soon.
So where am I going with this? AMD is missing the boat on this. If the hardware has the capability, one way to completely floor Nvidia in the HPC arena would be to release GPU acceleration patches for the open source numerical libraries that almost every scientist and his white-coated dog use.
Overnight, I can bet you scientists looking to buy clusters will be looking for AMD kit.