The compute costs must be extraordinary. As I understand it, inference needs, very roughly, something like 200 GB of RAM and multiple high-end GPUs/vector copr

The compute costs must be extraordinary. As I understand it, inference needs, very roughly, something like 200 GB of RAM and multiple high-end GPUs/vector coprocessors, to give output not much faster than a person can type. A VM on AWS of that class costs several dollars an hour. If there are 10,000 simultaneous users running inference (a low estimate, surely) then ChatGPT is probably costing Microsoft/OpenAI tens of thousands of dollars per hour, or millions per week? They have deep pockets for what is some great advertising, I imagine. But this couldn't possibly continue for free indefinitely with how much people are using it. Does anyone have more details on just what kind of resources GPT-4 needs to run inference?