Google has published a technical report that estimates the resource footprint of a text prompt to its Gemini artificial intelligence. The company reports a median prompt consumes 0.24 watt-hours of electricity, which it equates to running a standard microwave for roughly one second. The study also provides average estimates for greenhouse gas emissions and water use associated with a text query. Jeff Dean, Google’s chief scientist, said the team aimed to be comprehensive in the elements they included, and the publication is presented as one of the most transparent per-prompt measurements issued by a major technology company to date.
The report measures more than the specialized chips that run models. Google finds its custom tensor processing units account for 58 percent of the 0.24 watt-hours, while the host machine’s central processing unit and memory account for 25 percent, idle backup equipment for 10 percent and data center overhead, including cooling and power conversion, for the remaining 8 percent. Google emphasizes the median figure is not representative of all queries. More complex uses, such as ingesting dozens of books for a detailed synopsis or using reasoning models that take more internal steps, will require significantly more energy. The analysis covers only text prompts and does not include image or video generation, which prior work suggests can demand much higher energy. Google also reports the median prompt used 33 times more energy in May 2024 than in May 2025, attributing improvements to model and software optimizations.
The report includes a market-based greenhouse gas estimate of 0.03 grams of carbon dioxide per median prompt and a water-use estimate of 0.26 milliliters, or about five drops, per prompt. Google derived emissions using purchased clean energy credits and power purchase agreements totaling over 22 gigawatts, noting its on-paper emissions per unit are roughly one-third of the average grid where it operates. Researchers quoted in the article welcomed the disclosure as an important data point for understanding artificial intelligence energy demand, while also noting remaining gaps, including the company’s withheld daily query volume and the need for standardized, comparable energy metrics across providers.
