I don't think that Moore's law is what's driving down ML compute costs in particular, where there seems to be a lot of innovation going on in terms of hardware architecture and compilers (much of which is proprietary). Even just thinking about memory bandwidth, which historically has scaled much slower than compute: the $/second required to push 10+ TB of training data into some piece of hardware that can do useful work on it isn't going to fall by 100x in a decade.