If you think cost-efficient scrypt ASICs are possible right now, please, buy a gun and one bullet and take yourself out of the gene pool.
Why wouldn't it be? It is after all possible on a GPU, and designing a custom circuit for a workload is much much easier (albeit a lot more expensive!) than implementing it on a GPU (which is specialized for a very specific task, insanely parallell vector calculations).
So you doubt that mtrlt can do the trick?
If course it is possible, I don't think anyone disputes that. But the performance difference between CPU and GPU is not going to be anything near what it is for hashing.