Overview

• Google published a technical paper estimating a median Gemini Apps text prompt at 0.24 Wh of energy, 0.03 gCO2e, and 0.26 mL of water for inference based on a May 2025 analysis.
• The company reports 33x lower energy use and 44x lower total carbon per median prompt over the past year, alongside improved response quality.
• Its full‑stack method includes actual chip utilization, idle capacity, host CPUs and memory, data center overhead via PUE, and cooling water, with Google citing a fleet‑wide PUE of 1.09 and goals for 24/7 carbon‑free energy and 120% water replenishment.
• Google attributes gains to model and system optimizations such as Mixture‑of‑Experts, Accurate Quantized Training, speculative decoding, distillation into Gemini Flash and Flash‑Lite, and custom Ironwood TPUs said to be 30x more efficient than its first TPU.
• External experts argue the figures omit indirect water tied to electricity generation and rely on market‑based emissions accounting; Google says it reports the median prompt to reflect typical user behavior.