It's all just conjecture, your hypothesis about moving systems equally so.

But you seem adamant that there's no chance the providers serve slightly quantized models for subscription users during high loads, or otherwise tweak models for requests from those users.

It's tricky to prove either way, but the chance is not zero.