Yeah at this point the goal is to see how to maximize for inference. For training it is impossible from the get go to compete with the frontier labs anyway. Im trying to calculate (even amortized over 2 years) the daily cost of running the equivalent rig that can get close to a single claude agent performance. (without needing a 6-digit gpu).

Really the only reason to have a local setup is for 24/7 on-demand high-volume inference that can't tolerate enormous cold starts.