This tradeoff will be great for self hosted LLMs, because they don’t need large scale batching usually, and less great for cloud providers that do.