We recently had some real discussions about this. It’s worth remembering that designing, building and testing a custom chip runs, for a relatively old density, order 100 million dollars. Add to that the fact it would only be useful for a small subset of the problems your users want to use the machine for and it becomes pretty clear why supercomputer centers go generic as much as feasible. Also bitcoin is a bit of a poor example because the memory requirement for each unit is incredibly small. That’s rarely the case for any kind of meaningful simulation.