curious how the token/$ on trainuim compare to nvidia chips and amd. is the future "chipless" inference?