A recent report highlights the varying energy usage of the Gemini AI model, as illustrated by Dean, who notes that complex tasks, such as generating detailed synopses from multiple books, demand significantly more energy than typical prompts. In particular, reasoning tasks may also have a higher energy requirement due to their multi-step process.
This report focuses exclusively on text prompts; therefore, it does not account for the energy needs associated with generating images or videos. Other studies, including one from the MIT Technology Review’s Power Hungry series earlier this year, indicate that these tasks typically require even greater energy.
The report reveals a significant reduction in energy consumption for Gemini queries over time. Specifically, the median energy used for a prompt in May 2025 was 33 times lower than that in May 2024, a trend attributed to advancements in model efficiency and software optimizations made by Google.
Additionally, Google provides an estimate of greenhouse gas emissions per median prompt, calculated at 0.03 grams of carbon dioxide. This figure results from multiplying the total energy consumption by the average emissions per unit of electricity. Interestingly, the company employs a market-based approach to estimating emissions, which considers its electricity purchases from clean energy projects. Since 2010, Google has secured agreements for over 22 gigawatts from renewable sources like solar and wind, which has led to a reported emissions rate that is about one-third of the average grid emission levels in the regions where it operates.
The report also notes that AI data centers use water for cooling, estimating that each prompt consumes approximately 0.26 milliliters, or about five drops. Overall, the aim of the report is to provide users insight into the energy and water consumption associated with their AI interactions, suggesting that such resource use is comparable to common daily activities.
Source: https://www.technologyreview.com/2025/08/21/1122288/google-gemini-ai-energy/