Google Publishes First Detailed Data on Energy Use of AI Prompts
In a pioneering move toward transparency, Google has released an in-depth technical report detailing how much energy its Gemini AI applications consume per user prompt. Published on August 21, 2025, via MIT Technology Review, this disclosure represents the most comprehensive public estimate yet from a major AI company regarding the environmental footprint of individual AI interactions.
Measuring AI’s Energy Footprint
Google’s report reveals that the median query energy consumption for Gemini—a middle-ground prompt representative of typical usage—is approximately 0.24 watt-hours. To contextualize, this is about the amount of electricity needed to run a common microwave oven for one second. Beyond just electricity, Google also provided estimates for associated water consumption and carbon emissions linked to each text prompt submitted to Gemini.
This level of transparency marks a significant step forward. Prior attempts to gauge AI’s energy demands were hindered by lack of access to proprietary infrastructure and operational details integral to accurate measurement. As AI integration accelerates in society, interest in understanding the environmental impact of such technologies has intensified, making Google’s disclosure a breakthrough for researchers and analysts seeking clear data.
Energy Use Breakdown
Jeff Dean, Google’s chief scientist, emphasized that their analysis is notably comprehensive. The overall electricity usage per prompt does not stem solely from the AI’s specialized chips but includes full infrastructural support. Custom-designed Tensor Processing Units (TPUs), Google’s proprietary AI hardware analogous to GPUs, contribute 58% of the total electricity used.
The remainder breaks down into several key components:
- CPUs and memory of the host machines: 25% of total energy consumption
- Idle backup machines kept ready for failover: 10%
- Data center overhead, including cooling and power conversion: 8%
This holistic approach recognizes that the AI hardware only tells part of the story. The supporting systems and infrastructure are significant contributors to the overall energy footprint per AI prompt.
Contextualizing Energy Consumption
Casey Crownhart’s coverage in MIT Technology Review highlights that Google’s median estimate does not reflect every single query but rather a typical use case scenario. Some tasks, such as providing detailed synopses of numerous books or employing reasoning models that require multiple computational steps, consume substantially more energy.
Moreover, Google’s study was limited to text-based prompts. Generating images or videos through AI—tasks known to use more intensive resources—were outside the scope of this report.
Encouragingly, the company notes that the energy required to answer Gemini prompts has improved dramatically over time, with May 2025’s median prompt consuming 33 times less energy than one from May 2024. This progress is attributed to advances in AI model efficiency and software optimizations.
Environmental Impact Estimates
Google extended its analysis to estimate associated greenhouse gas emissions and water usage per prompt. The median Gemini prompt releases roughly 0.03 grams of carbon dioxide equivalent. This figure is derived using a market-based emissions factor that incorporates Google’s power purchase agreements with renewable sources, including solar, wind, geothermal, and advanced nuclear energy, totaling over 22 gigawatts since 2010. As a result, Google’s effective emissions per kilowatt-hour are about one-third of the average emissions of the grids where its data centers operate.
In terms of water consumption, Gemini queries use approximately 0.26 milliliters—around five drops of water—primarily for cooling servers.
Industry and Research Perspectives
Experts see this publication as an invaluable resource. Mosharaf Chowdhury and Jae-Won Chung, leaders of the University of Michigan’s ML.Energy leaderboard—which tracks AI energy costs—commended Google’s transparency and comprehensive analysis. Such data, only accessible to high-scale industry players, is vital to building realistic models of AI’s environmental footprint.
However, some key details remain undisclosed, such as the daily query volume processed by Gemini, which would allow estimates of total energy consumption. Researchers like Sasha Luccioni of Hugging Face underscore the need for more standardized reporting, akin to an Energy Star rating for AI tools, to enable clear and comparable energy assessments across models and platforms.
A Step Toward Environmental Accountability in AI
Google’s disclosure reflects growing public and academic demand for accountability concerning AI’s resource demands. Jeff Dean summarized the company’s goal: to provide users a practical understanding of energy use that places AI operations in perspective. He noted that the energy and water consumption per prompt are comparable to everyday activities like watching a few seconds of television or using a handful of drops of water.
As AI continues to permeate numerous sectors, such transparent analyses provide critical insight into its sustainability and highlight the ongoing need for efficiency improvements and responsible energy sourcing.
For more in-depth coverage on AI energy consumption and sustainability, read MIT Technology Review’s full technical report and related series on AI and environmental impact.
