Analog AI/approximate computing? There's research into it, worth looking at if you're interested.

The idea is you tolerate some loss/degradation (which neural networks do) but gain orders of magnitude power efficiency.