More likely, there's a work queue and sometimes it gets a bit backed up or falls a little behind while autoscaling resources kick in. Having fully dedicated capacity for something that's handled asynchronously, fluctuates with time, and can tolerate many minutes of processing delay would be a waste of money.
It's Amazon, to me it seems more likely they are tuning their ML with human input than they somehow need an hour to spin up AWS for their shiny new baby.
