Google has released detailed data on the environmental impact of its Gemini AI prompts, revealing a median energy consumption of just 0.24 Wh per prompt—comparable to watching TV for nine seconds—and associated greenhouse gas emissions of 0.03 grams of CO2e, with water use of 0.26 milliliters.

This data shows a significant 33-fold reduction in energy use per prompt from May 2024 to May 2025, driven by model improvements and software optimizations.

Google's investments in cleaner energy sources, new infrastructure, and hardware and software efficiencies have contributed to these reductions, supporting their goal of achieving net-zero emissions by 2030.

The methodology accounts for full system energy consumption, including idle hardware, CPUs, RAM, and cooling infrastructure, providing a more accurate picture of AI's environmental footprint than previous estimates.

This comprehensive approach considers all components involved in AI queries, leading to higher but more precise estimates of energy use compared to earlier public or industry figures.

Despite the low median energy per prompt, the overall environmental impact depends on total query volume, which remains undisclosed, though Gemini reportedly has over 350 million monthly users.

While the report offers unprecedented transparency, it does not include total query volume, which is critical for estimating overall AI energy demand, and some details remain proprietary.

AI's environmental costs include high electricity and water demands for data centers, which can strain ecosystems, along with environmental impacts from manufacturing and transporting hardware.

Google emphasizes that their estimates are lower than many public claims, partly because they account for idle energy use and cooling overhead, which many other estimates overlook.

The company highlights the importance of transparency and industry standards to better understand and mitigate AI's environmental impact.

Google's full-stack approach, involving advanced architectures, efficient algorithms, and ultra-efficient data centers, aims to enhance sustainability and reduce resource consumption.

Their methodology, which includes active and idle hardware, results in higher but more accurate estimates of energy use than previous public or industry estimates.