Google’s new study shows that a Gemini AI text prompt uses just 0.24 Wh of energy, 0.03 gCO₂e, and five drops of water—figures far lower than public estimates—while highlighting the importance of comprehensive, transparent measurement and ongoing efficiency improvements in scaling AI sustainably. (Source: Image by RR)

Methodology Accounts for Idle Chips, CPU Usage, and Data Center Cooling Overhead

Google has released new research showing the real-world environmental footprint of AI inference, offering one of the most detailed looks yet at how much energy, emissions, and water AI prompts consume. According to the company’s methodology, the median Gemini Apps text prompt uses just 0.24 watt-hours of energy, emits 0.03 grams of CO₂ equivalent, and consumes 0.26 milliliters of water — about five drops. That’s far lower than many public estimates, and equivalent to less than nine seconds of TV watching. By measuring across the full system, Google argues it provides the most accurate accounting yet of AI’s environmental impact.

The study, as reported in cloud.google.com, emphasizes that inference efficiency is becoming more important as AI scales. Over the past year, Google says it cut the energy use of the median Gemini prompt by 33x and reduced its carbon footprint by 44x, thanks to innovations in hardware, software, and data center operations. This includes co-designed TPUs, more efficient model architectures like Mixture-of-Experts, quantization techniques, and inference optimizations such as speculative decoding. Google also highlighted that its data centers operate at an industry-leading power usage effectiveness (PUE) of 1.09, while pursuing 24/7 carbon-free energy and replenishing more freshwater than consumed.

A key aspect of Google’s methodology is comprehensiveness. Unlike many studies that only count the energy used by active TPUs and GPUs, Google’s approach includes idle machine energy, CPU and RAM usage, and data center overhead like cooling. This provides a fuller picture of AI’s operating costs at global production scale. Using a limited “active chips only” method, the footprint per Gemini text prompt would appear just 0.10 Wh and 0.12 mL of water — but Google says such figures significantly underestimate true impact.

By publishing its methodology and results openly, Google aims to set a standard for AI resource accounting and drive industry-wide consistency. The company’s “full-stack” approach integrates model design, algorithm optimization, hardware co-development, and efficient data center operations. With AI demand accelerating, Google frames these gains as essential for scaling AI responsibly. At the same time, the company acknowledges there’s more to do — and that transparent measurement will be key in reducing AI’s environmental costs while maximizing its scientific and economic benefits.

read more at cloud.google.com