say using A10G ~$1.2/hr and with full utilization on vllm 112 reqs/min => per req ~$0.00018 versus gpt-3.5 turbo $0.002 per 1k token