What does LLM inference really cost? Break down GPU hourly rates, utilization, and cost per million tokens to decide when to self-host.