Energy Efficiency May Redefine the AI Competition Over Hardware
A Stanford and Together AI study suggests that AI progress is increasingly tied to how much intelligence can be delivered per unit of electricity, challenging the focus on larger chips and data centers.
A joint study by Stanford University and the AI infrastructure firm Together AI introduced a metric called intelligence per watt, assessing how accurately a system answers queries for each unit of power consumed. Testing over a million real-world queries on roughly twenty compact models and eight hardware categories, the researchers observed that models running on consumer-grade silicon now correctly answer a large portion of single-turn chat and reasoning tasks.
Between 2023 and 2025, the efficiency of these models improved several-fold, and the proportion of queries they could handle rose markedly. At the same time, the price of using leading models such as OpenAI's GPT-4 fell dramatically, underscoring a steep efficiency curve. The authors contend that as more everyday AI demand shifts to efficient, edge-based hardware, the traditional focus on building ever-larger data centers and restricting high-end chips may lose its strategic edge. They suggest that future policy should consider grid capacity and energy efficiency alongside chip export controls to shape AI competitiveness.
In this story
