I feel like the better pricing strategy would be something similar to what the geospatial API platforms like Google Maps do, with their explicit no-caching or time-limited-caching clauses. E.g. you're actually prohibited from retaining results from say a geocode beyond 30 days IIRC.
Amazon made this explicit with their Geospatial API pricing ( https://aws.amazon.com/location/pricing/ - "Places" tab) - where the pricing for being able to store a result is 8x higher.
>>paragr+(OP)
This really doesn't work in the context of AI training though. Sure, it would make reuse between models a lot harder, perhaps, but the general idea still holds, once a model is trained, the model is for forever.