“Approximately 95% of the tokens served on Baseten's dedicated inference platform are for custom models that have been modified by customers.”