BTW, how did you manage all of the throughput to the models and navigate the various throttling strategies for all the models you mentioned?