E.M.’s contributions were funded solely by the Mellichamp Academic Initiative in Sustainability at the University of California, Santa Barbara. The authors also acknowledge substantive suggestions by Nuoa Lei of Amazon.
1AI refers to a broad class of computational methods designed to mimic aspects of human intelligence. These include recognizing patterns, interpreting language, making predictions, and even generating creative content. While the concept of AI has existed for decades, the field has evolved dramatically in recent years ‒ both in technological capability and in public visibility.
2Traditionally, most AI systems have been “narrow” or task-specific: optimizing ad recommendations, detecting fraud, forecasting demand, reducing data center cooling energy use, or analyzing satellite or medical imagery. These systems are typically built using machine learning algorithms that identify patterns in data and improve human and institutional performance.
3More recently, a new generation of AI systems known as generative AI has emerged. These models, especially large language models (LLMs) like ChatGPT, Copilot, Claude, or Gemini, are trained on extensive datasets and can produce fluent human-like text, images, code, or music. Generative AI is often viewed as a general-purpose technology, capable of performing a wide array of tasks with minimal customization. It has raised hopes about productivity gains and accelerated scientific breakthroughs, but it has also sparked concern about its environmental footprint.
4This article focuses on a critical dimension of AI’s rise: its relationship with energy. AI’s rapid growth is drawing attention to its energy use and associated emissions. This article outlines what we know, what remains uncertain, and why these questions matter for policymakers, researchers, and industry leaders working to align AI with sustainability goals. It also lays out a framework for organizing and studying the disparate effects on energy use that this new technology will likely engender.
5To make sense of AI’s impact on energy use, it is helpful to distinguish three categories of effects, as illustrated in Figure 1:
-
Direct effects of AI operations
-
Effects of applying AI to energy-related activities
-
Interactive systemic effects of AI on the broader economy
Figure 1: Dimensions for assessing the e ects of AI on energy use and emissions
6These include the electricity consumed when AI models are trained and then used in practical applications. Training is the process where an AI model learns from large datasets ‒ adjusting its internal parameters to perform a task effectively. This step is performed periodically using highly energy-intensive and powerful computing infrastructure. Once trained, the model is deployed and used to generate predictions or outputs ‒ a process known as inference. Inference is far less computationally intensive per task than training but happens continuously, often across millions of devices.
7Training and inference together along with the associated “overhead” energy use associated with data center cooling and power provision make up the direct operational energy use of AI data centers. Among the various effects of AI on energy systems, these are the most readily quantifiable today ‒ although data gaps and varying methodologies make estimates uncertain.1,2 As AI adoption grows, inference may become the dominant source of energy demand due to its persistent and distributed nature. This potential shift has important implications for how data center infrastructure is designed, located, and managed.
8AI may increasingly be embedded in tools and systems that influence how energy is produced, distributed, and consumed across the economy.3 These potential applications can span many sectors.4 In electricity systems, AI could be used to forecast demand, balance grids, and manage batteries, interactions with variable renewable generation sources.5 In buildings and industry, AI could optimize heating, cooling, and manufacturing processes to reduce energy waste. In transportation, AI may enable route optimization, electric vehicle charging management, and mobility-as-a-service platforms. In agriculture, AI could accelerate automation of industries, shifting costs from labor to capital and energy. AI could be used to improve irrigation scheduling and reduce energy-intensive fertilizer use. Finally, it could be used to educate citizens about climate change and the need for an energy transition.
9On the other hand, AI may be deployed to reduce the cost of oil and gas exploration and extraction ‒ potentially increasing the competitiveness and longevity of fossil fuel systems. If used to guide autonomous vehicles, it could increase passenger miles driven6 and siphon passengers from public transit. It could also result in the creation of new goods and services and more targeted marketing of those goods and services, thus increasing consumption.7 Finally, it could be deployed to disseminate misinformation about the energy transition and manufacture doubt about new technologies.8
10These applications could thus lead to either increases or reductions in energy demand in the aggregate, depending on how the technologies are designed, deployed, and governed. While harder to measure than direct operational effects, these application-driven impacts are likely to grow significantly as AI adoption expands.
11Effects of AI outside the energy sector include advances AI may enable in science and technology ‒ such as new materials or technology breakthroughs ‒ as well as broader economic and societal changes related to productivity, economic growth and structural changes in labor markets. While these effects may ultimately have the largest impact on energy use and emissions, they are also the hardest to predict and depend heavily on how AI is used and governed.
12The inverted triangle in Figure 1 represents conceptually the relative scale and uncertainty of each category for energy use and emissions: direct effects are best understood but likely smaller in impact; system-level effects could be transformational but are highly uncertain. Scenario analysis and nimble governance will be essential for navigating future pathways.
13To assess the energy implications of AI, it is important to situate it within the broader landscape of information technology (IT) electricity use. AI workloads are typically hosted in large-scale data centers that often also support conventional IT services, such as search, social media, e-commerce, and enterprise computing.
14Centralized IT can be broadly divided into two categories: compute data centers, which run AI and other cloud services, and networks, which transmit and route data. These categories may sometimes overlap, and their energy profiles can differ.
15According to Malmodin et al.,9 compute data centers and networks each consumed about 1% of global electricity in 2020, and IEA estimated that the share of compute centers rose to 1.5% by 2024.4 IEA also estimates that AI workloads running on “accelerated servers” accounted for about 15% of global compute datacenter electricity use in that year. End-user devices ‒ such as laptops, desktops, smartphones, and printers ‒ account for an additional 2% of global electricity use, a figure that has remained relatively steady. Cryptocurrency mining, dominated by Bitcoin, has fluctuated between 0.5% and 1%, with large year-to-year variations.
16Figure 2 summarizes approximate fractions of global electricity use for these components in 2020 and 2024, based on the references and assumptions cited in the caption. Total electricity consumption from all computing equipment was about 6% of all electricity use in 2024.
Figure 2: Approximate fractions of total global electricity use represented by di erent types of computing and network equipment in 2020 and 2024
Sources: Compute data centers and networks in 2020: Malmodin et al..9 Compute data centers in 2024: IEA4. Networks in 2024 assumed to match compute data center total from IEA (1.5%). End-user devices in 2020 from Malmodin et al.,9 with percentage assumed constant to 2024. Cryptocurrency (Bitcoin) totals from Cambridge Bitcoin Energy Consumption Index (https://ccaf.io/cbnsi/cbeci) divided by electricity consumption totals from IEA.10
17Looking ahead, electricity use from AI workloads will be shaped by changes in five interrelated factors:
-
Service demand: How much AI is adopted and how intensively it is used across different sectors.
-
Energy efficiency of computing: How many computations we can generate per kWh of electricity used, which depends on hardware, software, and model architecture.
-
Cooling and power delivery technologies: The supporting energy needed to cool the computers and keep data centers running, often called infrastructure energy use.
-
Deployment constraints: Physical or logistical bottlenecks that affect the pace and scale of AI and energy infrastructure expansion.
-
Policies: Government rules, regulations and subsidies affecting data center design, construction, and operational choices.
18a. Service Demand. Many current assessments of AI’s energy impact assume it will become a ubiquitous, economy-transforming technology. While this is one possible future, it is not by any means the only one. Assessing AI’s future energy effects requires considering a range of possible scenarios, including some that include more moderate growth in AI use.11
19Industry projections for AI compute demand are aggressive, but their realization depends on whether businesses generate strong returns on AI investments and on whether user concerns about accuracy and reliability can be adequately addressed.12,13,14 AI’s trajectory could also be affected by new technologies (e.g. DeepSeek) that deliver similar services at much lower training costs.15 It could be affected by redefining the service being delivered, moving from large-scale training relying on increasingly large computing systems to smaller models, smaller datasets, reasoning models, and narrower use cases. Finally, as more AI moves from the training phase (which is energy-intensive) to the more widespread deployment or inference phase, the way it draws power will also change ‒ potentially spreading energy demand across more locations and devices.
20b. Energy Efficiency of Computing. Historically, rising demand for computing services has been partially or mostly offset by efficiency gains.16 In the early stage of the AI boom, efficiency was not top of mind ‒ companies bought available hardware, regardless of efficiency. As constraints in deploying AI manifested, the industry began to focus on efficiency as one path to alleviating those constraints.17
21This pattern matches the 2000 to 2005 period, when electricity use by data centers in the US and globally roughly doubled,18 prompting a strong focus on efficiency. That effort led to slower growth in data center electricity use to 2010 and little growth from 2010 to 2018.19 It is not yet clear how much efficiency improvements will offset growth in AI service demand in the future.
22For AI systems, efficiency is complex. Historically, models improved performance with scale20, 21 and models trained on more data outperformed earlier “best in class” models. AI researchers often plowed efficiency gains back into training even bigger models, thus increasing electricity use.22 This trend held for the most recent explosion of AI development.23
23For the most part, costs and benefits did not factor into these training efforts, except in the narrow sense that the companies building the big models presumably kept to the budgets they set for those projects. However, business value ‒ not just performance ‒ will eventually drive future AI deployments. Whether scaling models continues to make economic sense is uncertain. Ultimately, the market will decide ‒ based on users’ willingness to pay for AI services and the cost to deliver them.
24How much potential is there to improve computing efficiency? Most people know “Moore’s Law”, the trend of shrinking transistors combined with architectural changes24, 25 that drove rapid and consistent improvements in performance and efficiency for decades.26 Before 2000, computing energy efficiency at peak output doubled every 1.6 years or so.16 After 2000, this rate of improvement slowed to a doubling every 2.6 years.27
25Shrinking transistors isn’t the only path to efficiency. As that approach has become harder, the industry has shifted to innovations in other areas. Leiserson and colleagues28 identified ways the industry could continue to push performance and efficiency through changes to hardware architecture, better software and algorithms, and special purpose computing, such as integrated hardware-software design. The potential to improve AI efficiency is vast ‒ current technology is far from computing’s physical limits.22 Companies will no doubt need to rethink computing technology from first principles as they approach these limits in coming decades.29
26Some have invoked the “Jevons paradox”30 to argue that efficiency gains inevitably lead to greater overall energy use. But what we’re seeing today is different: AI service demand is rising so rapidly that it outpaces efficiency gains ‒ not because of higher energy efficiency, but because a new technology is expanding quickly.
27c. Cooling and Power Delivery Technologies. There is large variation in the efficiency and water use of the equipment used to cool and power data centers, and changes over time in the characteristics and use of those technologies can have a significant effect on energy used by these facilities.31, 32, 33
28d. Deployment Constraints. Growth in ability to meet service demand can also be uncertain because of supply-chain constraints in producing and deploying AI nodes and supporting equipment. In the first part of the recent AI boom, people bought as many AI nodes as NVIDIA could produce, causing shortages that could persist if rapid demand growth continues. There are also constraints on the physical systems, like backup power generators for data centers, that may hamper the speed of AI deployment. If service demand growth moderates, these issues become less pressing. When growth in new technologies is rapid, it may affect the rate at which these technologies can be deployed.34
29e. Policies. Policy changes may also have an impact on the evolution of data centers in the years ahead.4 There has been relatively little focus on such policies beyond setting targets for infrastructure efficiency, encouraging more efficient power supplies, and mandating disclosure of infrastructure efficiencies and other sustainability metrics.35 The difficulty of assessing computing efficiency across diverse data centers has prevented much work on the compute side of the equation. There is increasing attention to utility rate design, cost allocation, and incentives related to data centers, as these facilities have become important contributors to demand growth in some countries and regions.36, 37
30The interplay among the five factors discussed in the previous section introduces substantial uncertainty in the pace and scale of AI expansion. Service demand could grow rapidly if businesses ultimately realize strong returns on AI investments ‒ or may moderate if concerns about cost, performance, and reliability slow adoption.
31At the same time, the expansion of AI compute for inference means energy use patterns may become more distributed, lessening the need for more centralized computing facilities that can result in significant local pressures. Efficiency improvements are possible ‒ from better chips to co-designed software ‒ but they may not outpace demand growth. Past trends, such as Moore’s Law, suggest rapid gains are possible, but future improvements depend on breakthroughs in multiple layers of the computing stack.
32Many projections of future compute data center electricity use suggest continued rapid growth, doubling or tripling in the next five to ten years.4, 32, 38 All such projections (even those projecting more modest growth) should be viewed with caution given the underlying uncertainties and how fast things change in this industry.
33The deep uncertainty associated with future electricity demand for compute data centers reinforces the importance of scenario-based approaches.39, 40, 41, 42 Rather than projecting a single future, scenario analysis explores a range of possible outcomes shaped by technological, economic, policy, and behavioral dynamics. Scenarios help us anticipate both upside and downside risks ‒ highlighting, for instance, how AI could accelerate decarbonization in one path or entrench fossil fuel dependence in another.11
34Well-designed scenarios should reflect the full range of uncertainty, including future adoption rates, emissions intensities of power grids, regulatory responses, and possible rebound effects. As in other complex domains ‒ like finance or climate planning ‒ scenario thinking can help decision-makers stress-test assumptions and identify resilient, no-regrets strategies, and by including scenarios that diverge significantly from the conventional wisdom, potential upside and downside risks can at least be anticipated and assessed.
35Understanding AI’s climate impact requires an integrated view of its operations, applications to the energy sector, and interactive and systemic effects. Changes in energy use are a key factor affecting emissions, but there are other systemic effects that must also be considered.
36As outlined in this primer, direct energy use from AI operations ‒ training and inference ‒ is growing but still represented a relatively small share of global electricity demand as of 2024. However, most studies predict significant growth in data center electricity use in the years ahead.2, 4, 38 Its future impact on emissions will depend on demand trajectories, efficiency trends, deployment speed, and power grid decarbonization.
37Effects of applying AI to the energy system and to the economy as a whole could be more far-reaching for the climate. Crucially, these dynamics unfold in the context of growing global electricity demand and electrification across multiple sectors.10 AI and data centers are one but not the only large contributor to this growth ‒ but their local impacts can be significant, particularly where energy infrastructure is already under pressure.
38The IEA’s 2025 report on Energy and AI offers a valuable thought experiment as to why holistic scenarios can be useful.4 As shown in Figure 3, while the IEA estimates that direct energy use of data centers and AI operations will contribute to growth in emissions, they also illustrate how these emissions might be offset if AI is broadly adopted to support decarbonization of the energy sector. However, high-rebound scenarios ‒ where AI-driven efficiency gains lead to increased consumption ‒ could partially negate those benefits.
Figure 3: An exploratory analysis of AI impacts on global emissions in 2035
Source: Adapted from IEA (2025).4
39It is critical to move beyond high-level thought experiments toward designing test cases, controlled experiments, analysis frameworks, datasets, and scenario approaches to explore these three effects quantitatively.1 From there, the drivers of different trajectories can be identified, which could inform proactive measures to help steer AI in the directions most likely to deliver net emissions savings at the societal level, and away from directions that increase net emissions.
40Even given the immense uncertainties in making such estimates, the IEA’s analysis highlights how the overall climate impact of AI will depend not just on the technology itself, but on the broader context in which it is embedded. For example, AI’s net emissions effect will be shaped by how quickly electricity grids decarbonize, how demand of AI computing facilities shapes local grids, how AI is used across sectors, how much increases in service demand may offset efficiency gains, and whether local infrastructure can keep up with new loads.
41In short, the energy and emissions implications of AI are deeply context dependent. Meeting climate goals in an AI-driven future requires more than tracking electricity use ‒ it demands deliberate governance, forward-looking investment, an understanding of emissions intensity trends over time, and tools like scenario analysis that help us prepare for a range of possible futures.
42AI’s implications for energy and climate remain deeply uncertain ‒ especially beyond the next 5-10 years. While much attention focuses on the energy used to train and run models, the impacts stemming from how AI is applied across energy systems and how it drives broader social and economic change are equally important and potentially much larger than direct effects, leading to either net increases or decreases in energy use and emissions. At the same time, the technology itself is evolving rapidly, adding further unpredictability. Scenarios are critical tools for exploring this shifting landscape and understanding how AI’s expansion could influence energy systems, emissions, and climate goals.
RSS feed
