Key Point 1: Google has publicly disclosed the energy, water resource, and carbon emissions consumption data for each text query of Gemini AI for the first time, showcasing the highest transparency in the industry.
Key Point 2: A single text query with Gemini consumes an average of 0.24 watt-hours, produces 0.03 grams of carbon dioxide, and uses 0.26 milliliters of water, which is far lower than previous estimates, improving efficiency by 33 times.
Key Point 3: Experts point out that Google's report does not include key information such as indirect water usage and regional carbon emissions; the full picture remains to be clarified.
How energy-intensive is using AI?
Google recently publicly disclosed the energy consumption, carbon emissions, and water resource usage data for each text query of its Gemini AI application, becoming the most transparent case among large tech companies.
According to Google's latest technology report, the median energy consumption for a single text query in Gemini Apps is only 0.24 watt-hours (equivalent to one second of microwave operation), with carbon emissions at 0.03 grams and water usage at only 0.26 milliliters (about five drops of water), far below past external estimates of AI data centers.
Google emphasizes that the energy consumption for a single AI text reasoning is equivalent to less than nine seconds of watching television.
Google emphasizes that these figures reflect a significant improvement in Gemini's performance over the past year. Compared to May 2024, the energy consumption for similar queries in 2025 has decreased by 33 times, and carbon emissions have dropped by 44 times, primarily due to model architecture, hardware-software collaboration optimization, and increased data center operational efficiency.
Google's report not only calculates the power consumption of AI chips (TPUs, Tensor Processing Units) themselves but also includes the power used by the main CPU and memory (accounting for 25%), backup idle devices (accounting for 10%), and indirect energy consumption from data center operations such as cooling and power conversion (accounting for 8%). Google Chief Scientist Jeff Dean stated that this comprehensive calculation method is closer to the actual energy consumption of AI operating at a global scale.
Google also employs a 'market benchmark' to calculate carbon emissions, based on its agreements for purchasing electricity from clean energy sources such as solar, wind, geothermal, and advanced nuclear energy, resulting in a carbon emission rate per unit of electricity that is only one-third of that from the general power grid. Regarding water resources, Google only accounts for the direct water used for cooling data centers, excluding indirect water usage such as for power generation; the report indicates that each query consumes approximately 0.26 milliliters of water.
Google's initiative has been recognized by academia, as it contributes to the standardization of resource consumption comparisons in the AI industry. However, experts including Shaolei Ren, a professor at the University of California, Riverside, and Alex de Vries-Gao, founder of 'Digiconomist,' pointed out that Google's report only reveals 'the tip of the iceberg,' as it does not include indirect water usage (like power generation cooling) and regional carbon emissions (calculated based on local power grid energy composition), and does not disclose key data such as the number of queries, making it difficult to fully assess the overall environmental impact of Gemini.
Alex de Vries-Gao pointed out that AI currently accounts for 20% of global data center electricity consumption, and it is expected to rise to nearly 50% by the end of this year. He emphasized that the growth rate of AI's energy consumption far exceeds that of Bitcoin mining, posing a greater threat to climate goals: 'The funds that Bitcoin miners invest are insignificant compared to the investments of tech giants like Google and Microsoft in AI. This trend is rapidly intensifying and poses a greater threat.'
He also emphasized that the high energy consumption of AI has significantly increased greenhouse gas emissions for companies like Google, with emissions rising 48% since 2019, making its net-zero goal for 2030 even more difficult.
Further Reading: The strongest card of the year has arrived! The 'Citi Uni-Open Co-branded Card' has been launched, analyzing the three major considerations behind Luo Zhixian.
A new ride-hailing platform has appeared in Taiwan! What is Bolt? Five strategies to challenge Uber, possibly triggering a battle of incentives?
Sources: Google, MIT, Wired
The initial draft of this article was prepared by AI, organized and edited by Li Xiantai.