A dedicated server for ChatGPT inference must balance GPU power for token generation, network latency for user expe