Agreed. Although inference is good enough on the Mac, there is no way I am training on them at all.

It's just more efficient to offload training to cloud Nvidia GPUs