Google's Gemini AI: Measuring Environmental Impact and Driving Efficiency

This document outlines a comprehensive study by Google into the environmental footprint of its AI serving infrastructure, specifically focusing on the Gemini AI assistant. As AI adoption accelerates, understanding and mitigating its environmental impact—including energy consumption, carbon emissions, and water usage—is crucial.

Key Findings and Methodology

Google introduces a full-stack measurement methodology to accurately assess these impacts in a production environment. This approach accounts for the entire AI serving infrastructure, from active AI accelerators to host systems, idle capacity, and data center overhead.

• Comprehensive Measurement: The study employs a detailed methodology that includes active AI accelerators, host CPU/DRAM, idle machines, and data center overhead (PUE).
• Gemini Apps Focus: The research applies this methodology to Google's Gemini Apps product, providing first-party data on AI serving environmental metrics.
• Reduced Impact: Findings indicate that a median Gemini Apps text prompt consumes significantly less energy (0.24 Wh) and water (0.26 mL) than many previous public estimates.
• Efficiency Gains: Over one year, Google achieved substantial efficiency improvements, leading to a 33x reduction in per-prompt energy consumption and a 44x reduction in carbon emissions for Gemini Apps.

The paper emphasizes the importance of standardized, comprehensive measurement boundaries for accurate comparison and to drive environmental responsibility across the AI industry.