Partially appliedApplied Not applied Note: This key applies across all boxes.Emerging design for efficiency use case examples TABLE 2 Source: AI Energy Impact public use case database. The supercomputer: a sovereign, green AI data centre for regulated sectors BOX 1 Challenge AI growth in regulated sectors is outpacing the available computing power. Building sovereign- grade, high-performance infrastructure remains complex and limited, constraining access to responsible and energy-efficient computing power. Solution –Modular, high-performance AI data centres with advanced infrastructure management and liquid cooling –3D modelling of thermal and electrical flows for predictive optimization and scalability –Operating on renewable energy to ensure long-term efficiency and low-carbon performance Impact Near-term impacts realized or anticipated (less than one year) –50% faster deployment compared to traditional builds –High computational efficiency –20% energy savingsFurther impacts realized or anticipated (more than one year) –Additional efficiency and carbon reductions from ongoing optimization –Scalable, low-carbon design supporting the expansion of AI infrastructure Reviewing the levers in action Energy-efficient hardware: Advanced liquid-cooled racks Green data centres: Modular, renewable- powered Model optimization: Opportunity to build on infrastructure-driven energy reductions through software efficiency improvements* Life cycle impact tracking: Opportunity to integrate real-time life cycle assessment (LCA) to reduce embodied carbon* Heat recovery systems: Opportunity to deploy to offset local heating demand* *See Table 2 for relevant “design for efficiency” use case examples. Sources: Kahil, H., et. Al. (2025). Reinforcement learning for data center energy efficiency optimization: A systematic literature review and research roadmap. Applied Energy, Vol 389, 125734. https://www.sciencedirect.com/science/article/pii/ S0306261925004647?; Makin, Y., Maliakkal, R. (2025). Sustainable AI Training via Hardware-Software Co-Design on NVIDIA, AMD, and Emerging GPU Architectures. Cornell University. https://arxiv.org/abs/2508.13163?; Lescuyer, L. (2024). Revealing full data center environmental faces thanks to life cycle analysis. https://www.datacenterdynamics.com/en/opinions/revealing- full-data-center-environmental-faces-thanks-to-life-cycle-analysis/?; IRENA. (2025). Waste heat recovery from data centres. https://www.irena.org/Innovation-landscape-for-smart-electrification/Power-to-heat-and-cooling/31-Waste-heat-recovery- from-data-centres?.Energy- efficient hardwareGoogle (AI chips over two generations): More efficient tensor processing unit (TPU) chip design has led to a threefold improvement in the carbon-efficiency of AI workloads Model optimizationGlobal software company (CodeGen 2.5 Optimization): Re-engineered architecture and TPU tuning cut training energy 40% with no performance loss, through improved model design and hardware tuning Green data centresCrusoe and Redwood Materials (renewable powered modular AI): Data centre with second-life electric vehicle (EV) battery storage, anticipating carbon dioxide (CO2) and operational cost savings Life cycle impact trackingBrussels Environment: Redesigned digital services with life cycle principles, cutting webpage carbon intensity by approximately 80% and demonstrating efficiency gains through sustainable design Heat recovery systemsNordic District: Repurposed compute waste heat for heating, supplying 40% of the district’s needs From Paradox to Progress: A Net-Positive AI Energy Framework 14

From Paradox to Progress A Net Positive AI Energy Framework 2025