Intelligence per Watt: A Unified Metric for the AI Era (www.intelligence-per-watt.ai)

A new metric called Intelligence per Watt (IPW) has been proposed to guide the shift from centralized AI models in data centers to more efficient, distributed systems. As AI demand skyrockets—reportedly increasing by 1,300 times in token processing—the traditional performance-per-watt approach no longer suffices. IPW aims to measure task accuracy relative to power usage, facilitating better workload redistribution from cloud solutions to edge devices. Local models, particularly those with fewer parameters, can now handle everyday tasks like email drafting and document summarization with impressive cost and efficiency, yielding a 5.3x improvement in IPW through a combination of model and hardware advancements. The significance of this metric lies in its potential to reshape AI's economic landscape and enhance national competitiveness. By introducing the concept of Gross Domestic Intelligence (GDI), which correlates intelligence-per-watt with accessible power, the research reveals disparities in AI resource allocation between nations like the U.S. and China. This shift could enable local architectures to harness existing devices, boosting AI capacity by two to four times without requiring new data centers. Furthermore, the study emphasizes that economically relevant AI tasks can generate substantial value per watt, incentivizing advancements in both model efficiency and practical application across industry sectors.