Intelligence per Watt: A Unified Metric for the AI Era

- "Intelligence per Watt" is a proposed metric that attempts to measure AI system performance relative to energy consumption, similar to how "performance per watt" is used to evaluate traditional processors and hardware. - The metric aims to provide a unified way to compare AI models (like GPT, Claude, Llama) and hardware (NVIDIA GPUs, custom AI chips) on efficiency grounds, not just raw capability. - This matters because AI's exploding energy demand is becoming a major cost and environmental concern — datacenter electricity use is projected to double by 2026, and training a single large model can emit as much CO₂ as several cars over their lifetimes. - Current benchmarks (MMLU, HumanEval, etc.) measure only capability; there is no standard way to ask: "How much intelligence do I get per unit of electricity?" This site proposes filling that gap, aiming to give researchers, buyers, and regulators a common efficiency yardstick. - The concept is analogous to fuel economy (miles per gallon) for cars: without it, you might compare only top speed or horsepower, missing the cost and sustainability dimension entirely.