I have observed some unusual behavior.

I wonder if it simply due to reprioritization of resources.

Presumably, there is some parameter that determines how long a model is allowed to use resources for, which would get tapered in preparation for a demand surge of another model.