Everybody has an opinion about local AI. Almost nobody has numbers. So we built a two-box local LLM stack for a client, put a power meter on it, and compared every token against the cloud.
