Is it worth paying for a pre-configured GPU for open-source LLM inference instead of using a traditional pay-per-token provider? The biggest advantage seems to be having no usage limits