Google released a technical paper detailing a new methodology for measuring the environmental footprint of artificial-intelligence inference and reported steep efficiency gains for its Gemini models. According to the study, the median Gemini Apps text prompt now consumes 0.24 watt-hours of electricity, emits 0.03 grams of carbon-dioxide equivalent and uses 0.26 milliliters of water—roughly the energy required to watch nine seconds of television. The figures represent a 33-fold drop in energy use and a 44-fold decline in carbon emissions per prompt between May 2024 and May 2025. Google credited the improvement to a mix of software optimizations such as mixture-of-experts routing and speculative decoding, smaller distilled model variants, and its custom Ironwood tensor-processing units, which it says are 30 times more energy-efficient than its first publicly available TPUs. The company’s infrastructure leaders, Amin Vahdat and Jeff Dean, said the analysis goes beyond chip-level power draw to include idle capacity, CPU and RAM overhead, and cooling systems across its data-center fleet, which operates at an average power-usage-effectiveness of 1.09. Google said it is publishing the methodology to encourage industry-wide consistency as AI workloads grow and to underpin its goal of running data centers on 24/7 carbon-free energy while replenishing 120 percent of the water they consume.