Artificial Intelligences Energy Paradox 2025

Data centre consumption includes both AI and non-AI elements. AI processing, particularly for generative AI, is more energy-intensive due to large model complexity, longer training durations and substantial data processing. This increased energy intensity, however, is accompanied by the additional benefits that capabilities like generative AI can provide, including the ability to perform more complex work and to enable expanded value opportunities.1.3 Opportunities to reduce AI system electricity consumption Data centre demand over time FIGURE 4 200400600800100012001400 20230 2024 2025 2026 2027 2029 2028 2030 Non-AI demand (TWh) AI demand (TWh)Data centre demand (TWh): Non-AI versus AI Note: This is an extrapolated scenario that extends the IEA’s forecast from 2023 to 2026 through 2030 using a combination of 2021-2023 historical growth and their proposed growth rate from 2023-2026. Source: International Energy Agency (IEA); Goldman; Accenture. Enabling a more energy efficient AI system includes exploring opportunities within data centres to reduce electricity consumption. Accordingly, a non-exhaustive inventory of example strategies are explored below. Data management strategies Within AI’s first stage (planning and data collection), “digital decarbonization” techniques can address “dark data”, which occupies server space and consumes electricity without providing value. For some organizations, dark data may account for as much as 60-75% of stored data.11 Digital decarbonization strategies can identify and eliminate dark data, reducing storage and electricity consumption. Opportunities may also exist to repurpose dark data to generate value. Featured data management use case TABLE 1 Loughborough University: automotive industry collaboration: unlocking dark data for sustainable industrial maintenance Situation/context Approach Results “Dark data” remained in storage, underused due to poorly structured formats.A knowledge management system with data scraping and enrichment techniques was developed to integrate and structure dark data, organizing it into valuable datasets for decision-making, and waste categories for disposal.In total, 10-20% of dark data was transformed into actionable knowledge, improving fault analysis and maintenance, enhancing data reliability, reducing downtime, lowering the environmental footprint and highlighting waste data. Source: Community consultation. Artificial Intelligence’s Energy Paradox: Balancing Challenges and Opportunities 9