Honestly, the primary reason you would ditch a H100 is because the ML models have moved onto a different data type, e.g. nobody uses fp8 anymore and everyone jumped onto fp4 or some custom block float thing.