From Paradox to Progress A Net Positive AI Energy Framework 2025

Making AI itself more efficient The first action driver, “design for efficiency”, embeds sustainability into AI models, hardware and infrastructure from the start. It supports both large- scale systems and smaller, frugal AI36 approaches by improving efficiency across algorithms, hardware use and system design, enabling models to achieve their purpose with minimal energy and material use. Learn more in the Forum’s insight report, Nature Positive: Role of the Technology Sector. Efficiency must also reflect how energy is consumed. Training is brief but power-intensive, while widespread inference can exceed it over time. These patterns vary across AI types and determine where optimization yields the most impact. Since AI requires significant electricity, water, land and materials, building efficiency early enables sustainable growth. This, in turn, ultimately supports this paper’s third action driver, shaping demand wisely, by promoting the use of appropriately scaled models for each application. Key levers include: –Energy-efficient hardware (low-power chips, accelerators, neuromorphic processors) –Model optimization (sparse models, quantization, pruning, federated learning) –Green data centres (renewable-powered, optimized water cooling, modular design) –Life cycle impact tracking (carbon, water, materials) –Heat recovery systems (waste heat reuse) Use case insights and takeaways: Across 37% of use cases within the current (but expanding) inventory, organizations are applying diverse levers to improve AI efficiency. However, few measure or disclose life cycle impacts consistently. Progress remains uneven, and transparency is essential.Strategic recommendations: –Embed energy key performance indicators (KPIs) into AI procurement and design. –Align hardware refresh and innovation with sustainability goals. –Track energy and carbon intensity throughout the AI life cycle.2.1 Design for efficiency Data centre electricity consumption is set to more than double to around 945 terawatt-hours (TWh) by 2030… rising to around 1,200TWh by 2035 (base case). International Energy Agency (IEA). (2025). Energy and AI. Design for efficiency – key levers FIGURE 5 Energy-efficient hardwareGreen data centresModel optimizationLife cycle impact trackingHeat recovery systemsDesign for efficiency The first action driver, “design for efficiency”, embeds sustainability into AI models, hardware and infrastructure from the start. From Paradox to Progress: A Net-Positive AI Energy Framework 13