Now is a great time because 3D hardware is coming out of the dark ages of being toy game acceleration technology, and morphing into highly-parallel general computing technology in its own right. The last hurdle is that the GPU vendors need to get out of the mindset of "how many shader instructions should we limit out card to?" and aim to create true Turing-complete computing devices.
We're already almost there. You just need to stop treating your 1024 shader instruction limit as a hardcoded limit, and redefine it as a 1024-instruction cache of instructions stored in main memory. Then my 1023-instruction shaders will run at full performance, and my 5000-instruction shaders might run much more slowly, but at least they will run and not give you an error or corrupt rendering data. You need to stop looking at video memory as a fixed-size resource, and integrate it seamlessly into the virtual memory page hierarchy that's existed in the computing world for more than 30 years. The GPU vendors need to overcome some hard technical problems and also some mental blocks.
In the long run, what will define a GPU -- as distinct from a CPU -- is its ability to process a large number of independent data streams (be they pixels or vertices or something completely arbitrary) in parallel, given guarantees that all input data (such as textures or vertex streams) are constant for the duration of their processing, and thus free of the kind of data hazards that force CPU algorithms to single-thread. There will also be a very different set of assumptions about GPU performance -- that floating point is probably much faster than a CPU, that mispredicted branches are probably much slower, and that cache misses are probably much more expensive.
