February 12, 2011 - Epoch AI, a nonprofit artificial intelligence research organization, has released a report for the ChatGPT New research on energy consumption suggests that the level of energy consumption may be much lower than previously generally recognized.
Previously, a widely cited statistic showed that ChatGPT answering a question consumes about 3 watt-hours ofquantity of electric charge or currentThis figure is 10 times higher than that of Google searches. However, researchers at Epoch AI believe this figure is overestimated. By analyzing OpenAI's latest default model, GPT-4o, Epoch AI found thatChatGPT's average query energy consumption is approximately 0.3 watt-hours, which is below the energy consumption level of many household appliances.
In an interview with TechCrunch, Joshua You, a data analyst at Epoch AI, said, "ChatGPT's energy consumption is nothing compared to using regular appliances, regulating the temperature of a room, or driving a car."
1AI notes that with the rapid growth of the AI industry, its energy consumption and environmental impact have become a public concern. Just last week, more than 100 organizations jointly issued an open letter calling on the AI industry and regulators to ensure that new AI data centers do not overconsume natural resources and avoid relying on non-renewable energy sources.
Joshua Yoo noted that the reason he undertook this analysis.It's because of the outdated nature of previous studiesFor example. For example, the study that came up with the 3-watt-hour energy consumption estimate assumed that OpenAI was using older, less efficient chips to run its models. I've seen a lot of public discussion that correctly recognizes that AI will consume a lot of energy in the coming years, but doesn't accurately describe the current energy consumption of AI," You said. Additionally, some of my colleagues have noted that the estimate of 3 watt-hours of power consumption per query is based on older research and, from preliminary calculations, appears to be too high."
While Epoch AI's 0.3 watt-hour figure is an estimate, as OpenAI has not yet published enough details to make an accurate calculation, the study is still significant. In addition, the analysis does not take into account the power consumed by additional features such as ChatGPT's image generation or input processing. You also acknowledged that ChatGPT queries with "long inputs" such as long file attachments may consume more power than normal questions.
However, Yo expects thatChatGPT's base energy consumption may increase in the futureAI will become more advanced. He noted, "AI will become more advanced, it may take more energy to train these AIs, and future AIs will probably be used more frequently and handle more complex, more varied tasks than people do now with ChatGPT."
Additionally, the focus of OpenAI and the AI industry as a whole is shifting toward inference models. Unlike models such as GPT-4o, which are capable of near-instantaneous response, theReasoning models take seconds or even minutes to "think" about the answer., this process consumes more computational power, which leads to higher energy consumption.
Inference models will increasingly take on tasks that older models can't do, and generate more data to do so, all of which will require more data center support," You said. While OpenAI has begun to roll out more energy-efficient inference models, such as o3-mini, it seems unlikely that these efficiency gains will offset the additional energy requirements of the inference models' "thinking" process and the increased use of AI globally.
