From Paradox to Progress A Net Positive AI Energy Framework 2025

2 Dark data and unconscious consumption: Many AI queries and training runs occur without visibility into their energy or carbon cost, creating wasteful computation and growing volumes of “dark data” that consume power through ongoing storage and cooling.20 AI’s hidden data impact FIGURE 2 Useful data Collected and unused data Compute demand Emissions footprint Energy storage The compute demands of frontier models and the concentration of energy use in hyperscale data centres intensify these dynamics. Unlike past digital shifts, today’s AI boom is unfolding under explicit resource constraints, presenting a responsibility and opportunity to design energy use deliberately. The risk of a net-positive divide The risk of a generative AI divide is a real concern. Countries with concentrated technology capabilities and large-scale data centres are pulling ahead, while others face structural barriers. This imbalance could deepen economic and innovation gaps, creating a two-speed world where AI-driven benefits are unevenly shared. Existing publications and articles21 show that computing power for advanced AI is increasingly concentrated in a few regions, raising concerns about equitable participation in the AI economy. Actions to avoid this could include: –Investing in south-north collaboration and shared infrastructure –Supporting regional innovation hubs tailored to local energy contexts –Ensuring equitable access to sustainable computing to prevent digital inequalityAs energy becomes a limiting factor, AI capacity may concentrate in regions or firms with surplus electricity and infrastructure. This dynamic risks deepening digital divides and creating asymmetries in access to innovation, potentially turning energy-rich areas into dominant hubs and leaving others behind.22 Barriers to achieving net-positive AI energy Achieving net-positive AI energy requires overcoming a complex set of challenges across six dimensions: Technical –Energy-intensive model training and deployment –Cooling and infrastructure inefficiencies –Supply-constrained hardware limitations and slow refresh cycles Measurement and transparency –Lack of standardized energy use metrics –Opaque or incomplete energy reporting –Fragmented data and benchmarking gaps Unlike past digital shifts, today’s AI boom is unfolding under explicit resource constraints, presenting a responsibility and opportunity to design energy use deliberately. From Paradox to Progress: A Net-Positive AI Energy Framework 8