They're performing the resets when the cluster load goes down and they have sufficient capacity to handle the reset.

Capacity is one thing but inference still costs money.