even if you're using your own llm anything over 7B parameters is definitely costing you lol every time you run inference you get charged compute.