I doubt how good results that could achieve. I agree that 10~100 times improvement is feasible by optimizing the hardware. But the hardware in general need to be improved, yet the impenetrable barrier light speed is blocking on the way.
And more complete AI systems should require hundreds of thousands times the computation power. Really, this has the same issue as bitcoin.
I doubt it since GPU was already not a bad tool for this job. The generality of GPGPU helped a lot here.