The amount of transistor packing you could do has been slowing down considerably in the past 5 or so years. Current state of the art is around 45nm production. That's why you see more increases in cores and parallel processing than in clock speed. The good thing is that GPU work lends itself well to parallel processing.
However, the thing to consider with cell phones is that battery life and heat dissipation are major factors, thus I don't really see high-end ps3 type of graphics anytime soon.
That said, OpenGL ES2 hardware already includes both a vertex shader and a fragment shader, and can do multi-sampling almost for free. I would expect actual non-fixed pipeline stream processor-based GPUs in the next year or two, and that should give you what top of the line integrated GPUs give you today.
One thing to keep in mind is that most of these mobile chipsets are SoC, which mean high speed data paths between internal components, but sharing external RAM between them (very much like integrated GPUs today). I don't think putting GDDR memory in there will be done anytime soon, since you really want to keep them cheap.