Protien folding simulation is such a large and basic need globally there ought to be enough large scale interest to make development of specialized ASICs to deal with these problems cost effective and exceedingly useful for all who need to do these simulations. A quick check of google shows such chips do in fact exist with unbelivable performance figures which kick the snot out countless tens of thousands of CPU/GPUs. There is no shortage of funding for medical research so it begs the question why waste CPU/GPU resources on folding simulations?
I still do seti and milkyway at home because there are no resources allocated for seti and milkyway at home is interesting to me personally.
First of all, protein folding is not the only thing they do, the Folding@HOME infrastructure is used by many for a variety of bio-molecular studies.
Secondly, custom ASIC-based machines like Anton and MDGRAPE (which are AFAIK the only such machines around these days) consist of much more than a custom-chip, they use specialized interconnects, memory, software, etc. and cost a lot. The MDGRAPRE-4, the coming version of the Riken-developed custom molecular simulation machine costs $10M + $4M (development + manufacturing) which poses serious financial limitations to it. Moreover, these specialized machines are only able to run a handful of molecular dynamics algorithms and while fast, they are nowhere near as versatile as general-purpose codes like AMBER, GROMACS, NAMD, etc. Although it is true that these specialized machines are a few orders of magnitude faster in terms of absolute performance (i.e time to solution and not Flops), due to their limitations and the way they are used, some researchers argue that they employ a "brute force" approach to molecular simulations which is not cost-effective from the point of view of science/$ delivered.
I personally wouldn't call machines like Anton and MDGRAPE a complete waste, they achieve impressive advances in hardware, software, and science results in a specific direction: pushing the limits of how fast can one run a single simulation. There are certainly other (some would say better) ways to get amazing results with general-purpose (super)-computers be it using massive clusters or cycles donated to folding Foldging@HOME.
Finally, let me explain why is there compute-resource shortage in the (bio-)molecular simulation filed which will remain for the foreseeable future no matter how much money do various governamental and non-governamental agencies pour into it. Molecular dynamics is extremely compute-intensive, a single iteration of the MD algorithm requires 10^8-10^10 Flops (not LINPACK Flops!), repeated for millions of times during a single simulation of a bio-molecular system (and such a simulation can take weeks even on a big machine). And that's still a few orders of magnitude short of what would be needed to simulate timescales at which biological processes take place. Therefore, any compute-resource available can be harnessed for molecular simulation research and Folding@HOME does a decent job at utilizing donated cycles.
Admittedly, there are some in the community who think that Folding@HOME is wasteful, but that's a topic for another discussion.
Disclaimer: I am involved in the development of the GROMACS open-source molecular simulation package which is in fact on of the computational engines used by Folding@HOME. Still, I believe I have not been biased in the way I presented Folding@HOME and molecular dynamics in general.