And it's highly circumstancial, as LLM efficiency keeps improving as the tech matures.