I was under the impression that doing things directly in custom silicon was always more efficient than emulation (all other things held fixed) both from a performance and power consumption perspective. The problem of course is that it costs orders of magnitude more money to implement, to say nothing of the difficulty in designing/testing the thing.

It's basically the same progression we saw in bitcoin mining: Custom software on a general purpose CPUs, then on GPUs, then programmed on FPGAs, then actual ASICs that are straight up custom silicon to do nothing but check those hashes.