You said you need $80k in hardware for "good performance". I'm saying the local AI inference workflow will be a lot more flexible about performance than that, and can get away with something vastly cheaper and in line with what the user owns already.