How AI Queries Affect Energy Consumption Levels
Artificial intelligence often feels weightless. It appears as text, images, or videos on our mobile phone and laptop screens. However, behind these seemingly instantaneous responses lie large data centres working intensively and consuming substantial quantities of electricity.
Recent measurements reveal a startling fact: a single video generated by AI can consume more than 100 times the electricity of a single image. This means the type of request (prompt) we submit significantly affects the energy footprint produced.
This finding comes from engineers at the University of Michigan (U-M), who have built a public measurement system to compare energy consumption, speed, and quality across various AI models performing the same tasks. In their study, the team tested 46 models and seven types of tasks, comprising a total of 1,858 combinations of hardware and software configurations.
The focus of measurement was on the inference stage — the moment when a trained model provides an answer to a user’s request. This stage consumes the most power. According to estimates from the ML.ENERGY Benchmark, inference accounts for 80–90 per cent of total computational requirements in the AI economy.
To illustrate the scale, the International Energy Agency (IEA) projects that electricity consumption by data centres in the United States will reach 183 terawatt-hours in 2024 and increase to 426 terawatt-hours by 2030. These figures demonstrate the importance of understanding energy efficiency in AI development.
In testing, the reasoning mode of problem-solving could consume up to 25 times more electricity per response compared to casual conversation. Why is this?
The key lies in tokens — small segments of text generated by the model. Each token triggers mathematical calculation processes and data movement within servers. In reasoning mode, the model often generates around ten times more tokens, thereby increasing electricity consumption and reducing the number of requests that can be processed simultaneously.
In essence, the longer the AI response, the greater the electricity consumed.