If they are overwhelmed with usage requests and can't buy GPUs fast enough, the model could be the same but the amount of compute spent on the answer could be r

If they are overwhelmed with usage requests and can't buy GPUs fast enough, the model could be the same but the amount of compute spent on the answer could be ramping down impacting the quality. It would be really cool to quantify how much compute spend per interaction. If this was visible, users could spend an arbitrary amount and modify how much they're willing to pay for 'better' responses. This is probably a better business model.