Stanford researchers introduce Intelligence per Watt metric

Researchers at Stanford University presented Intelligence per Watt, a metric measuring the intelligence efficiency of local language models compared to power draw. The study evaluated over 20 local language models across one million queries to assess how much inference demand can shift from cloud data centers to local hardware. They also previewed a companion study called Joules to GDP, which measures the economic value and expert time produced per joule of energy. Open article →