Artificial Intelligences Energy Paradox 2025

Electricity consumption of AI1 Model deployment is AI’s most energy-intensive stage (accounting for approximately 60%) – innovative strategies can mitigate consumption. The AI lifecycle begins with planning and data collection, during which data is gathered, processed and stored.8 Next, the model development phase includes design, problem analysis and data preparation. Model training then optimizes the model through iterative data exposure. Model deployment subsequently opens the model for real-world application. Lastly, monitoring and maintenance support ongoing refinement.Further research is needed to estimate consumption for stages 1 and 5, however estimates exist for stages 2-4. Within these three stages, model deployment is the most energy-intensive (approximately 60-70% of combined electricity consumption), but will likely continue growing in the long term. Model training is the next most energy-intensive, accounting for 20- 40% of consumption, followed by model development at up to 10%.9 These estimates however, will likely vary across differing AI model types.1.1 The AI life cycle Electricity consumption across the AI life cycle FIGURE 2 2 341 5Stage 1: Planning and data collections on nature* Stage 2: Model development 10% Stage 3: Model training 30%Stage 4: Deployment 60%Stage 5: Monitoring and maintenance* *Insufficient data available for estimation Source: Electric Power Research Institute (EPRI). (2024). Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption. International Energy Agency (IEA). (2023). Tracking Data Centres and Data Transmission Networks. https://www.iea.org/energy-system/buildings/data-centres-and-data- transmission-networks; D. Patterson et al. (2022). The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink. Computer, vol. 55, no. 7, pp. 18-28. https://ieeexplore.ieee.org/document/9810097. Artificial Intelligence’s Energy Paradox: Balancing Challenges and Opportunities 7