In a study conducted by Sasha Luchoni, a research associate at Hugging Face, and Boris Gamazaychikov, head of AI sustainability at Salesforce, the performance of 40 open AI models was analyzed, including developments from companies such as OpenAI, Google, Microsoft, and the Chinese DeepSeek. All models were tested on the same computer with identical queries—ranging from simple questions to complex mathematical tasks. The CodeCarbon tool was used for real-time monitoring of energy consumption.
The difference in energy consumption turned out to be colossal. For example, DeepSeek R1 required only 50 Wh per query when reasoning was disabled, while with reasoning activated, it consumed 308,186 Wh. The Microsoft Phi 4 model used 9,462 Wh during active reasoning and only 18 Wh when reasoning was turned off. OpenAI gpt-oss showed a less noticeable difference—8,504 Wh at maximum load and 5,313 Wh at minimum.
In August, Google presented internal research indicating that the average energy consumption per text prompt in the Gemini AI service is 0.24 Wh, which corresponds to less than nine seconds of television viewing. The company noted that this figure is "significantly lower than many public estimates."
This research helps to better understand the differences in energy consumption between models and to make informed choices about systems for specific tasks. As Luchoni pointed out, "not every query requires the use of the most powerful systems," and it is important to apply AI wisely to reduce the burden on energy networks.
The increase in energy consumption from AI raises concerns among both analysts and the general public. Over the past five years, wholesale electricity prices near data centers have risen by 267%, and the construction of new centers complicates the achievement of climate goals set by major tech companies.
With the growing use of reasoning models, the industry needs to consider their energy costs. Microsoft CEO Satya Nadella noted that the industry should obtain "public permission for energy consumption" and use AI to promote economic growth and societal well-being.
