My research was in this direction. We already know that these analog neural chips could be orders of magnitude faster than digital equivalents. There is also a lot of military research going on in this area for few decades. However architecture innovations are much faster on software level and dedicated hardware approaches have not been able to catch up. Once things slow down on software level, slowly hardware llms could become the norm.

I agree.