Cache once, judge many times. Cache the list once, run every judgement pass against it at a fraction of input price. This is the cheapest shape for the feature.

Multiply by the fleet. Cents per user times the fleet is a budget line. And the nightly clustering job, not the model, is where the compute goes.