Fifty API keys, one queue, ten GPU workers: the parts of a shared LLM that are not the model