The rise of commercially available artificial intelligence (AI) systems has generated alarming headlines. Every time ChatGPT drafts an email, it is akin to emptying a bottle of water, or every response from a chatbot equates to turning on a light bulb for about 20 minutes. As technology becomes more integrated into our daily lives, researchers are trying to quantify the environmental impact of AI. While concerns about AI’s energy demand and growing omnipresence are increasing, successful efforts by policymakers to track or regulate the industry’s footprint have been scarce. This is mainly due to the lack of relevant data and reporting mechanisms from companies in the tech and energy sectors.
What are the energy costs of AI and where do they come from?
Environmental comparisons used in news articles highlight only one aspect of AI’s energy demands. The ecological impact of the technology spans the months, and even years, needed to train and deploy an AI model. Different development stages require specialized hardware with varying, but generally large, consumption needs. During the training stage, models are fed curated, though often extensive, datasets to “learn” according to their algorithms. The graphics processing units (GPUs) used in this process typically run 24/7, generating high energy requirements (generally backed by non-renewable sources) that quickly escalate depending on the model’s complexity. For example, researchers from Google and UC Berkeley estimated that training OpenAI’s ChatGPT-3 model consumed enough electricity to power around 120 US households for a year. Although the lack of access to this type of data has made it difficult to calculate GPT-4’s electrical demands, researchers estimate it probably consumed 50 times more electricity. An estimate based on data from the International Energy Agency (IEA) suggested that ChatGPT uses nearly 10 times more electricity than a normal Google search. These numbers will only grow as more users adopt generative AI programs over traditional search engines. In fact, researchers estimate that if Google processed 9 billion AI-powered searches a day, they could require 23-30 times the energy needed for normal searches. Some experts, like the AI and Climate lead at Hugging Face, maintain an “AI Energy Score” dashboard, which reports an approximate difference of 62,000 times between the highest and lowest energy needs among different use cases and models on the dashboard. Energy demands for generating text, images, and videos vary widely, with video generation being the most energy-intensive; recently, the MIT Technology Review found that an AI model used approximately 3.4 million joules, the amount of energy needed to run a microwave for over an hour, to generate a 5-second video.
Much of the processing of these commands is done through data centers. The electric consumption of data centers was 4.4% of US electric demand in 2023 and could grow to 6% by 2026. Globally, these centers represent 1-2% of the world’s energy needs, but given the increasing demand for AI, some expect this could reach 21% by 2030. This does not account for water use, which is needed to cool the hardware. Some UK researchers estimate that global water use for data processing could reach half of the country’s water consumption by 2027. Additional indirect energy costs are incurred in maintaining large building infrastructures, which include computing hardware, storage systems, and networking equipment. There are also end-of-life costs. A GPU is typically used for four years before being discarded or reused. Currently, there is little information on how these devices are disposed of and what the environmental impact of this accumulation of waste will be. The lack of data makes AI’s energy consumption less transparent.
These estimates are further complicated by the fact that tech and electric companies rarely disclose energy or water consumption data and are not obligated to do so. Most estimates are uncertain because they lack proprietary information. Although Google, Microsoft, and Meta refused to share the energy needs of their AI model requests, OpenAI CEO Sam Altman recently wrote that an average query uses 0.34 watt-hours, or what a light bulb uses in a couple of minutes, and 0.000085 gallons of water. However, these figures do not include the extensive energy needs of training the model and only reflect energy needs on a per-query basis. There are improvements in AI energy consumption efficiency. This level of consumption is not lost on industry leaders. While there has been some growth in efficiency, these improvements have begun to stall. Large data centers present an opportunity for greater energy efficiency, as illustrated in the construction of “hyperscale” centers. Other examples include hardware-level improvements, such as power capping, or limiting the amount of energy fed to processors and GPUs; this technique has been shown to decrease energy consumption by up to 15% with minimal effects on the user experience. Other efficiencies also arise from building smaller models, pruning or quantizing the algorithm architecture design, transitioning to renewable energy sources, increasing collaboration between AI companies, or using AI to identify potential improvements for itself and other models. For example, Google DeepMind reported that it achieved a 30% reduction in data center energy consumption using AI to better predict cooling needs. However, these improvements still require additional AI use and, therefore, greater energy consumption, even with new efficiencies considered. With lower energy costs for AI systems, the increasing demand could lead to a rebound effect, where more efficient technology and its lower production costs generate greater demand and adoption, resulting in increased consumption again. This cycle is sometimes known as the Jevons paradox, a term coined initially when the steam engine became more efficient.
Unfortunately, the impact of increasing energy demands will not be felt uniformly. Some regions and states are already being drained of resources much more than others. “Data Center Alley” in Virginia offers an example; it houses over 300 data centers, the most in the United States. Residents are fighting against the continuous expansions of data centers in the region, leading to higher energy bills, increasing water demand, and worsening air quality due to the diesel backup generators of the facilities. Further south, the NAACP recently sued Elon Musk‘s xAI for allegedly operating turbines for a data center in South Memphis without proper permits. The NAACP emphasized that the toxic emissions from the turbines were directed towards predominantly Black neighborhoods that bear the brunt of environmental racism.
To ensure that the growing adoption of AI does not exacerbate environmental degradation, policymakers, companies, and communities will need a better understanding of where to focus their attention and resources, starting with better data on emissions and energy consumption.
Scientists from MIT found that companies using independent auditors for assurance decreased their total emissions by 7.5% annually, even when starting with higher emissions than companies without assurance. However, these estimates may not be detailed enough. For example, the IEA estimate includes activity in data centers, without a direct focus on AI applications. Unless tech companies reveal what percentage of their energy use is related to AI, ambiguity will persist. Last year, Senator Ed Markey (D-MA) introduced the “Artificial Intelligence Environmental Impacts Act of 2024” to better measure AI’s impact. The legislation would require the Environmental Protection Agency (EPA) to conduct a comprehensive study on the issue, as well as convene a consortium of stakeholders through the National Institute of Standards and Technology (NIST) and create a voluntary reporting system. This would culminate in an interagency report presented to Congress to inform findings and provide policy recommendations. The Government Accountability Office (GAO) also published a report on the environmental and human effects of generative AI in April 2025, detailing six possible policy options, including encouraging developers to share model details about the infrastructure used for training and using generative AI, as well as providing government incentives for more resource-efficient models and training methods. At the state level, at least 60 bills have been introduced nationwide to address the impact of data centers, but there has been little significant change. Harvard researchers found that the carbon intensity of data centers was 48% higher than the national average, in part because data centers are often built in areas with dirtier electric grids. Large companies, like Meta and Microsoft, have turned to alternative energy sources to power their data centers, focusing on nuclear energy. However, the MIT Technology Review emphasized that nuclear energy represents only 20% of US electricity production, while clean but intermittent technologies, such as wind and solar, cannot fully power continuously operating data centers. Given the many uncertainties, it is the right time for the US to support additional research initiatives on methods to reduce AI’s environmental impact while continuing to improve transparency and establish standards in this important area. Failure to do so could not only accelerate the climate crisis but also slow the adoption of AI applications and delay their benefits.
