OpenAI inference cost reduction cut ChatGPT guest traffic from tens of thousands of Nvidia GPUs to just a couple hundred, ...