Carnegie Mellon University
Wilton E. Scott Institute for Energy Innovation  ›  Key Initiatives  ›  Open Energy Outlook  ›  Energy Use for Artificial Intelligence: Expanding the Scope of Analysis
May 27, 2024

Energy Use for Artificial Intelligence: Expanding the Scope of Analysis

By Mike Blackhurst

Stakeholders have expressed concerns over the energy used to support the development of artificial intelligence and machine learning (AI/ML) tools and their applications. This blog summarizes recent estimates of the energy used directly by AI/ML. Drawing on parallels to end-use efficiency, the article suggests that emerging AI/ML applications will likely significantly but indirectly impact energy use in other sectors, requiring an expanded scope of analysis to capture AI/ML’s full energy impacts. Finally, the article summarizes how public agencies could better understand and plan for the energy implications of AI/ML.

What influences direct energy use for AI/ML?
The energy required to develop or train an AI/ML model is typically much higher than a single application (Desislavov et al. 2023). However, energy used for the development phase is bounded, whereas energy use for AI/ML applications scales with repeated inferences. Within each of these phases, direct energy use also varies based on algorithmic efficiency, model application, and the conversion efficiency of hardware (Desislavov et al. 2023; Luccioni et al. 2024). For example, a more accurate algorithm consumes more energy, all else equal. Similarly, a simpler application, such as text classification, consumes less energy than a more complex application, such as image classification.

For context, the following activities consume about the same amount of electricity: half of a million applications of text classification, one thousand applications of image classification, one cycle of a clothes washer or dishwasher, 10 miles driven by an electric car, and one year of lighting using an LED bulb (Luccioni et al. 2022; DOE 2024a/c).

Industry-wide estimates of the energy used for AI/ML are uncertain given the scarcity of primary data. Data servers use 10 to 50 times more electricity than conventional commercial buildings (DOE 2024b). Recent short-term forecasts estimate global electricity demand for AI/ML will increase to 100 TWh by 2026 or 2027, which is similar to the electricity used by one million U.S. homes in a year (de Vries 2023; IEA 2024, DOE 2024a).

While improved estimates of energy used by AI/ML directly are helpful, AI/ML is currently transitioning from development to a deeply uncertain applications phase. Sectors that incorporate AI/ML could also indirectly but significantly change their energy consumption. Studies of conventional end-use energy efficiency offer insight into potential indirect impacts. 

What is end-use energy efficiency and how does it impact energy use?
“End-use efficiency” refers to the efficiency of converting energy into energy services using technologies found in buildings (e.g., appliances), industry (e.g., a boiler), and transportation (e.g., a vehicle). Engineering models of end-use efficiency assume that increased technological efficiency leads to commensurate reductions in energy use. For example, an engineering model would indicate that a 50% improvement in a vehicle’s fuel economy would reduce energy consumption for transportation by 50%.

This simplified perspective, however, ignores behaviors that often accompany efficiency changes. Since energy is not free, efficiency improvements can also reduce the effective price of energy services. As a result, many studies find improved end-use efficiency increases the demand for energy services (Greening et al. 2000; Sorrell et al. 2009). Moreover, the monetary savings achieved through efficiency can be spent on other goods and services, many of which also require energy (Sorrell et al. 2020). Studies also find that repurposed time saved through efficiency can impact energy use (Sorrell et al. 2020; Mizobuchi and Yamagami 2022). A host of less tidy, so-called “irrational” behaviors further challenge both engineering and neoclassical models of efficiency (Sorrell et al. 2020; Frederiks et al. 2015).

This trade-off between potential technological efficiency gains and respective behavioral responses is often called “the rebound effect,” which describes the observation that energy use “rebounds” away from expected savings towards growth in energy services. To be clear, most studies find that the energy savings from technological efficiency exceed that of commensurate responses. In other words, a 50% efficiency improvement may lead to a 20% reduction in energy efficiency, not the full 50% predicted by engineering methods alone. However, the long-term impacts of end-use efficiency remain elusive (Azevedo 2014). Indeed, the literature on rebound appears to stretch over decades of incremental knowledge accumulation, likely reflecting the intermittent policy emphasis on efficiency, slow technology adoption, and the emergent but long-run impacts of end-use efficiency. 

How does end-use efficiency relate to AI/ML?
AI/ML can be considered an end use in and of itself, converting energy into information services. Viewed this way, significant efficiency gains to date have moderated energy demands for AI/ML (Desislavov et al. 2003). However, AI/ML could more broadly change the balance of labor, capital, and time used in existing economic sectors and also create entirely new economic activities. This broader view of AI/ML suggests its energy implications extend well beyond the energy used directly to develop and apply AI/ML. A full energy accounting would track energy use induced by all of AI/ML’s efficiency modalities.

Improving our understanding of how AI/ML will impact energy use requires us to ask tough questions. What sectors will use AI/ML and how? How will these uses change consumption, production, and prices? How does AI/ML shift the balance of energy used in production? What new industries will be created by AI/ML? What happens to legacy industries and assets? Since AI/ML has been cast as time-saving, it will also be important to track how AI/ML shifts time use. For example, AI/ML used for autonomous vehicles could eliminate time spent paying attention while driving. How much of the time will be “re-spent” on additional travel or other activities? Answers to these questions are largely empirical and not sufficiently informed by how much energy is used directly by AI/ML.

What are potential next steps?
Public agencies have an important role in asking and answering these questions given the outsized impacts energy decisions have on society. While myriad agencies collect helpful information, existing federal surveys are currently uncoordinated across outcomes of interest, lag well behind innovation cycles, and are limited to cross-sectional analyses. An inter-agency collaboration focused exclusively on energy use for AI/ML—potentially including partnerships with states, grid operators, and the private sector—could prove incredibly helpful.

Recent news suggests grid regulators and operators have been caught off guard by the sudden spike in electricity provisions requested by AI/ML companies (Halper 2024; Vincent 2024). Given expanding electricity infrastructure requires long lead times, improved understanding of moderate- to longer-term AI/ML applications is needed. To this end, the energy community has developed expertise in scenario analysis that could prove useful in planning for AI/ML. Energy scenarios reflect informed but still uncertain futures related to prices, policy, and technology, factors that also characterize the uncertainty underlying AI/ML applications. Scenarios that reflect reasonable but varying assumptions related to the uptake of AI/ML in various sectors, the creation of new industries, and the resulting energy balance would likely provide actionable insight into how to better coordinate AI/ML innovation with our shared energy system.

  1. Greening, L., Greene, D. L., & Difiglio, C. (2000). Energy efficiency and consumption—The rebound effect—A survey. Energy Policy, 28(6), 389–401. https://doi.org/10.1016/S0301-4215(00)00021-5
  2. Alternative Fuels Data Center: Charging Electric Vehicles at Home. (n.d.). Retrieved May 24, 2024, from https://afdc.energy.gov/fuels/electricity-charging-home
  3. Azevedo, I. M. L. (2014). Consumer End-Use Energy Efficiency and Rebound Effects. Annual Review of Environment and Resources, 39(Volume 39, 2014), 393–418. https://doi.org/10.1146/annurev-environ-021913-153558
  4. Data Centers and Servers. (2024). Energy.Gov. https://www.energy.gov/eere/buildings/data-centers-and-servers
  5. Desislavov, R., Martínez-Plumed, F., & Hernández-Orallo, J. (2023). Compute and Energy Consumption Trends in Deep Learning Inference. Sustainable Computing: Informatics and Systems, 38, 100857. https://doi.org/10.1016/j.suscom.2023.100857
  6. DOE (U.S. Department of Energy). (2024). Residential Energy Consumption Survey, 2020. U.S. Department of Energy. https://www.eia.gov/consumption/residential/
  7. Frederiks, E. R., Stenner, K., & Hobman, E. V. (2015). Household energy use: Applying behavioural economics to understand consumer decision-making and behaviour. Renewable and Sustainable Energy Reviews, 41, 1385–1394. https://doi.org/10.1016/j.rser.2014.09.026
  8. Halper, E. (2024, March 7). Amid explosive demand, America is running out of power. Washington Post. https://www.washingtonpost.com/business/2024/03/07/ai-data-centers-power/
  9. IEA (International Energy Agency). (2024). Electricity 2024. https://www.iea.org/reports/electricity-2024
  10. Luccioni, A. S., Jernite, Y., & Strubell, E. (2023). Power Hungry Processing: Watts Driving the Cost of AI Deployment? (arXiv:2311.16863). arXiv. http://arxiv.org/abs/2311.16863
  11. Luccioni, A. S., Viguier, S., & Ligozat, A.-L. (2022). Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model (arXiv:2211.02001). arXiv. http://arxiv.org/abs/2211.02001
  12. Mizobuchi, K., & Yamagami, H. (2022). Time rebound effect in households’ energy use: Theory and evidence. Cleaner and Responsible Consumption, 5, 100066. https://doi.org/10.1016/j.clrc.2022.100066
  13. Sorrell, S., Dimitropoulos, J., & Sommerville, M. (2009). Empirical estimates of the direct rebound effect: A review. Energy Policy, 37(4), 1356–1371. https://doi.org/10.1016/j.enpol.2008.11.026
  14. Sorrell, S., Gatersleben, B., & Druckman, A. (2020). The limits of energy sufficiency: A review of the evidence for rebound effects and negative spillovers from behavioural change. Energy Research & Social Science, 64, 101439. https://doi.org/10.1016/j.erss.2020.101439
  15. Thomas, B. A., & Azevedo, I. L. (2013). Estimating direct and indirect rebound effects for U.S. households with input–output analysis Part 1: Theoretical framework. Ecological Economics, 86, 199–210. https://doi.org/10.1016/j.ecolecon.2012.12.003
  16. Vincent, J. (2024, February 16). How much electricity does AI consume? The Verge. https://www.theverge.com/24066646/ai-electricity-energy-watts-generative-consumption
  17. Vries, A. de. (2023). The growing energy footprint of artificial intelligence. Joule, 7(10), 2191–2194. https://doi.org/10.1016/j.joule.2023.09.004