From Paradox to Progress A Net Positive AI Energy Framework 2025

AI model efficiency: transparency on energy and carbon impact BOX 11 Challenge AI model training and inference consume large amounts of energy and water. Comprehensive data on their environmental impact has been limited, making efficiency measurement difficult. Solution –Google developed a methodology to measure Gemini inference model energy and water use, and carbon emissions –Enables users to understand environmental cost of each prompt Impact Near-term impacts realized or anticipated (less than one year) –A 33-fold reduction in energy and a 44-fold reduction in carbon per prompt –Transparent reporting of model-level environmental impactFurther impacts realized or anticipated (more than one year) –Offers replicable measurement approach to inform broader sustainability practices –Supports user awareness and responsible deployment Reviewing the levers in action Global energy efficiency metrics: Reports efficiency at the inference and model level, publishing energy-per-query improvements Public disclosure frameworks: Sustainability reports include AI and data centre ESG energy disclosures Benchmarking platforms: Contributes research to open benchmarking* Third-party verification: Limited data centre operations verification, but independent model audits emerging* Open data repositories: N/A *See Table 7 for relevant “transparent measurement and accountability” use case examples. From Paradox to Progress: A Net-Positive AI Energy Framework 31