the answer, from a supercomputer user/programmer/scientist's perspective (for now) is a hybrid of the two. some portions of a parallel program are very easily ported to the GPU, and net a huge speedup right away, others require a complete rework from an algorithmic perspective which will take some time (in science, this means years).
if GPUs become standard on all nodes of major supercomputers, the memory per GPU core goes way up, and communication between GPU nodes, bypassing the CPU entirely, is standardized, then GPUs will have a shot at becoming more of a dominant force in supercomputing. looking forward to it!
If you look at a high enough level of abstraction, the current series of GP-GPUs are basically a large collection of vector processors with a fast interconnect between them -- one that mirrors the mechanism Cray used in its supercomputing system (X-MP, Y-MP). The main difference is one of scale.
These are currently limited in scope of applications, but they are very well suited to applications that were formerly the domain of specialized vector supercomputers, so I think that they will be big in supercomputing even if they don't grow in versatility. However, I think that they will, especially with Intel getting into the biz.
Intel may not very good at building graphics processors and drivers, but Intel is VERY good at building on-chip interconnects, small processors, and very fast on-die caches. That ought to make for some interesting competition, since nVidia and AMD aren't constrained by the x86 ISA, while Intel is.
