The rapid integration of Artificial Intelligence (AI) into global economies has driven transformative advancements in sectors such as healthcare and agriculture. However, this technological revolution incurs significant environmental costs, particularly through substantial energy consumption and greenhouse gas (GHG) emissions. The carbon footprint of AI, stemming from energy-intensive processes like hardware production, model training, and operational inference, poses a critical challenge to global sustainability. Drawing on scholarly sources, the scale of AI’s environmental impact, the factors intensifying these costs, and potential mitigation strategies are explored, underscoring the need for sustainable practices to prevent AI from exacerbating climate change.
The carbon footprint of AI is a composite of emissions generated across its lifecycle, encompassing the production of computing hardware, the training of AI models, and their operational use. Mitu and Mitu (2024) articulate that the production of computing technology, such as graphics processing units (GPUs) and data centre infrastructure, is highly energy-intensive, contributing nearly 36% of global industrial emissions (Chen et al. 2022). The extraction and processing of raw materials, often reliant on fossil fuels, further amplify this footprint (Panagiotopoulou et al. 2022). Training large-scale AI models, particularly deep learning models such as OpenAI’s GPT-4, significantly contributes to AI’s environmental footprint. O’Donnell and Crownhart (2025) estimate that training GPT-4 consumed approximately 50,000 megawatt-hours (MWh) of electricity, equivalent to powering a city like San Francisco for three days. Similarly, De Vries (2023) reports that models like GPT-3 and BLOOM required between 324 and 1,287 MWh during training, driven by the computational intensity of processing vast datasets and optimising billions of parameters. This substantial energy demand, often reliant on fossil fuel-based grids, highlights the critical need for sustainable practices in AI development to mitigate its environmental impact.
The increasing complexity of models exacerbates this issue, as computational power requirements grow exponentially, outpacing energy efficiency improvements (Mitu & Mitu 2024). The operational phase, or inference, where AI models generate real-time outputs, is increasingly recognised as a dominant energy consumer. De Vries (2023) estimates that ChatGPT’s inference phase requires 564 MWh daily, significantly surpassing its training energy use. O’Donnell and Crownhart (2025) highlight that inference accounts for 80–90% of AI’s computational power demand, driven by the proliferation of AI applications in daily life, from chatbots to image generators. For instance, a single ChatGPT query consumes approximately five times more electricity than a standard web search (Zewe 2025). As AI models become ubiquitous, with ChatGPT alone receiving one billion daily messages (O'Donnell & Crownhart 2025), the cumulative energy demand of inference is poised to dominate AI’s environmental impact.
Several factors amplify the environmental costs of Artificial Intelligence (AI). First, the reliance on fossil fuel-based energy sources for data centres significantly increases AI’s carbon footprint. O’Donnell and Crownhart (2025) note that data centres often use electricity with a carbon intensity 48% higher than the US average, driven by the high power density requirements of AI systems. The rapid expansion of data centre infrastructure outpaces the availability of renewable energy, necessitating fossil fuel-based power plants (Chen 2025). Second, the lack of transparency from AI companies hinders accurate assessment and mitigation of environmental impacts. Chen (2025) and De Vries (2023) argue that proprietary models, such as ChatGPT and Google’s Gemini, provide little data on energy consumption, treating such information as trade secrets. This opacity forces researchers to rely on estimates from open-source models or supply-chain analyses, which may underestimate the energy demands of larger proprietary systems (Chen 2025). Third, the short lifecycle of AI models contributes to inefficiency. Frequent releases of new models render previous versions obsolete, wasting the energy invested in their training, while the increasing complexity of newer models escalates energy consumption (De Vries 2023).
Despite these challenges, Artificial Intelligence (AI) offers opportunities for environmental sustainability if managed responsibly. Mitu and Mitu (2024) advocate transitioning data centres to renewable energy sources, such as solar or wind, to reduce operational emissions. Chen (2025) highlights that renewable-powered data centres can significantly lower AI’s carbon footprint, supporting broader sustainability goals. Additionally, optimising AI algorithms through techniques like model pruning, quantisation, and knowledge distillation can reduce energy consumption during training and inference (Mitu & Mitu 2024). Wang et al. (2024) provide empirical evidence of AI’s potential to enhance sustainability, demonstrating that AI reduces ecological footprints and carbon emissions while promoting energy transitions across 67 countries. AI’s second-order effects, such as improving efficiency in agriculture and manufacturing, can lead to resource conservation and lower emissions. However, Wang et al. (2024) caution that these benefits are context-dependent, with industrialised nations facing a rebound effect where efficiency gains increase overall resource use. Policy interventions are crucial to align AI development with sustainability goals. Mitu and Mitu (2024) call for international collaboration to establish regulations ensuring transparency in energy reporting and incentivising green AI practices. The European Union’s 2023 Energy Efficiency Directive (European Parliament and Council, 2023), requiring data centres to report annual energy consumption, represents a step towards this objective (Chen 2025). Wang et al. (2024) recommend long-term AI development strategies, citing initiatives like China’s Next Generation Artificial Intelligence Development Plan, which balances innovation with environmental objectives.
The environmental costs of AI are substantial, driven by the energy demands of hardware production, model training, and inference, compounded by reliance on fossil fuels, lack of transparency, and rapid model obsolescence. Without intervention, AI’s carbon footprint risks undermining its societal benefits, accelerating climate change. However, through renewable energy adoption, algorithmic optimisation, and robust policy frameworks, AI can be harnessed to support a low-carbon future. As Wang et al. (2024) and Mitu and Mitu (2024) suggest, AI’s dual role as both a contributor to and mitigator of environmental degradation necessitates a balanced approach. Policymakers, industry leaders, and researchers must collaborate to ensure that AI’s transformative potential is realised without compromising the planet’s sustainability.
References:
1. Chen, Sophia. 2025. ‘How Much Energy Will AI Really Consume? The Good, the Bad and the Unknown’. Nature 639(8053): 22–24. ^ Back
2. Chen, Yuxuan, Yan, Wei, Zhang, Hua, Liu, Ying, Jiang, Zhigang, and Zhang, Xumei. 2022. ‘A Data-Driven Design Approach for Carbon Emission Prediction of Machining’. Proceedings of the International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, Vol. 86212, V002T02A062. American Society of Mechanical Engineers. ^ Back
3. De Vries, Alex. (2023). ‘The Growing Energy Footprint of Artificial Intelligence’. Joule, 7(10), 2191–2194. ^ Back
4. European Parliament and Council of the European Union. 2023, September 13. ‘Directive (EU) 2023/1791 on Energy Efficiency and Amending Regulation (EU) 2023/955 (Recast)’. Official Journal of the European Union, L 231, 20.9.2023, pp. 1–111. http://data.europa.eu/eli/dir/2023/1791/oj ^ Back
5. James O'Donnell and Casey Crownhart. 2025, May 20. We did the math on AI’s energy footprint. Here’s the story you haven’t heard. MIT Technology Review. https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/ ^ Back
6. Mitu, Narcis Eduard, and George Teodor Mitu. 2024. ‘The Hidden Cost of AI: Carbon Footprint and Mitigation Strategies’. Available at SSRN 5036344. ^ Back
7. Panagiotopoulou, Vasiliki Christina, Stavropoulos, Panagiotis, and Chryssolouris, George. 2022. ‘A Critical Review on the Environmental Impact of Manufacturing: A Holistic Perspective’. The International Journal of Advanced Manufacturing Technology, pp.1–23. ^ Back
8. Wang, Qiang, Yuanfan Li, and Rongrong Li. 2024. ‘Ecological Footprints, Carbon Emissions, and Energy Transitions: The Impact of Artificial Intelligence (AI)’. Humanities and Social Sciences Communications 11(1): 1–18. ^ Back
9. Zewe, Adam. 2025, January 17. ‘Explained: Generative AI’s Environmental Impact’. MIT News. https://news.mit.edu/2025/explained-generative-ai-environmental-impact-0117 ^ Back
