How Many Questions Can the World Afford to Ask AI? (www.chicagobooth.edu)

Researchers have quantified the energy and material costs associated with AI operations, determining a "question budget" that caps the number of inquiries users can make based on current and projected energy supplies. By applying Landauer's principle, they established a theoretical minimum energy cost for generating AI tokens and compared it to the staggering energy consumption of current AI chips, which use approximately 50 quintillion times more energy per token than the minimum. This highlights significant room for improvement in hardware efficiency and underscores the industry's nascent stage of development, analogous to the steam-engine era. The study also delves into the implications of this question budget, projecting that by 2028, the U.S. energy allocation could allow for about 2,200 questions per person daily, contrasting starkly with current usage levels. As AI becomes cheaper and more accessible, demand for resources like copper—vital for powering data centers—will skyrocket, potentially stressing global supply chains. The findings draw attention to the interplay between physical, economic, and measurement constraints within the AI ecosystem, suggesting that while optimization may enhance performance, it doesn't necessarily address underlying resource challenges. This multifaceted analysis signals an increasingly urgent need for sustainable practices in the rapid expansion of AI technologies.