I suspect that any GPU cycle not spent on inference will just be dedicated to training (which as I understand it can “soak up” essentially unlimited compute at constant value per token), and I’d not expect to see time-based billing until that changes.