Google has released one of the most detailed disclosures yet on the environmental footprint of artificial intelligence usage. The report focuses on the impact of a single text prompt on its Gemini Apps assistant and outlines electricity consumption, carbon emissions, and water use, while comparing results with past external estimates.
Specific Environmental Footprint of Google Gemini
Data from researchers at Google revealed the median text prompt used in its generative AI tool Gemini in May 2025 required 0.24 watt-hours of electricity, produced 0.03 grams of carbon dioxide equivalent, and consumed 0.26 milliliters of water. This is roughly equivalent to 5 drops of water or 9 seconds of television viewing.
The company claims its methodology is more comprehensive than previous studies. It includes not only the power drawn by the chips running artificial intelligence inference but also the energy used by central processing units, memory, and idle capacity maintained for reliability, as well as cooling overheads measured by data center efficiency indicators.
Google contends that earlier estimates focused exclusively on machine calculations and excluded surrounding infrastructure. This overstated per prompt impacts. A technical paper accompanied the announcement. It detailed the accounting methods and asserted that the new figures represent a more accurate view of environmental costs.
The report also highlighted efficiency gains achieved within a year. The company explained that the energy per median text prompt fell by a factor of 33, while total carbon emissions per prompt decreased by 44 times. These are due to hardware improvements, software optimization, and serving system refinements across data centers.
Important limitations remain. Google notes that calculations only cover text-based inference within Gemini and do not extend to more resource-intensive multimodal prompts like image and video. It also does not account for model training, which is known to consume substantial energy. Google also withheld figures on the total volume of prompts processed.
Overall Environmental Impacts of Artificial Intelligence
Comparisons with outside assessments reveal significant differences. A 2024 study by the Electric Power Research Institute estimated that a typical ChatGPT prompt required 2.9 watt-hours of energy, while a conventional search engine usage consumed about 0.3 watt-hours, underscoring the magnitude of variation across published findings.
Nonetheless, even with per prompt efficiency improvements, aggregate demand remains an issue. The International Energy Agency forecasts that global data center electricity consumption could more than double by 2030, potentially reaching 945 terawatt hours. This is approximately equal to the current total electricity consumption of Japan.
The disclosure from Google can be placed against the backdrop of its rising carbon and overall environmental footprint. The total greenhouse gas emissions of the company have increased by 51 percent since 2019, despite some efficiency progress. It reached 11.5 million metric tons of carbon dioxide equivalent in 2024 due to AI workloads.
It is also worth noting that the report may overlook larger system dynamics. Improvements in per prompt do not negate surging demand, and some raise concerns about the reliance of Google on market-based carbon accounting and partial water disclosures. A call for greater transparency and further details on location-based energy impacts are needed.
The disclosure still represents both a significant step in transparency and a reminder of unresolved questions. Google has demonstrated that inference can be more efficient than widely assumed, but without full data on usage volumes, training costs, and global deployment, the environmental footprint of AI remains uncertain and contested.
FURTHER READINGS AND REFERENCES
- Electric Power Research Institute. 2024. Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption. Electric Power Research Institute. Available via PDF
- Elsworth, C., Huang, K., Patterson, D., Schneider, I., Sedivy, R., Goodman, S., Townsend, B., Ranganathan, P., Dean, J., Vahdat, A., Gomes, B., and Manyika, J. 2025. “Measuring the Environmental Impact of Delivering AI at Google Scale (Version 1).” arXiv. DOI: 48550/ARXIV.2508.15734
- Vahdat, A. and Dean, J. 22 August 2025. “How Much Energy Does Google’s AI Use? We Did the Math.” Google Cloud Blog. Google. Available online