IBM have already got something based on the PowerPC arch that actually delivers 208 (peak) TFLOPs per rack that you can buy today. It's called BlueGene/Q and it is capable of delivering over 85% of that peak figure on LINPACK, which seems to be well beyond the reach of the CPU-GPU boxes at the moment. Admittedly it sounds a bit feeble compared to the vaporware - but take a look at the Top500 & Green500 and you'll see a bunch of them at the top end of the table. ;)
We are at an interesting point in the HPC wheel of reincarnation. On the one hand with have BlueGene/Q with small cores and tightly coupled vector units on the same die, and on the other we have big CPUs with big discrete GPUs on the other. Meanwhile in the commodity space we have fairly chunky CPUs with (smallish) integrated GPUs, and I suspect it'll be this model that will dominate in the nearish future. It is stating the bleeding obvious though, NVidia and AMD roadmaps point that way, and Intel's MIC stuff is pretty much already there.
The killer micros are coming back to attack the discrete GPUs. :)