Behind the immediacy of generative artificial intelligence lies an unprecedented consumption of natural resources. Although interacting with language models may seem like an intangible process taking place in the cloud, the training and use of these algorithms require a big physical infrastructure of data centres, the environmental cost of which is worth highlighting.
Of all the GPT model versions, it is the third version (GPT-3) that marks a key milestone, as it is one of the models with the most robust and reliable environmental and sustainability metrics published, closely followed by the GPT-4 model.
Training this model took approximately one month, during which it consumed 1,280 MWh of electricity, equivalent to the annual consumption of around 370 Spanish households. At the same time, this process generated approximately 550 tonnes of CO₂, comparable to the carbon footprint produced by 150 fuel-powered vehicles over a year and required around 700,000 litres of water for evaporative cooling of the servers, equivalent to a continuous 28-day shower. At user level, a single query to GPT-3 is equivalent to charging a smartphone by 10 per cent, using one tablespoon of water and running a combustion engine car for a few seconds, always taking into account the variability in processing time and the length of the response generated.
However, assessing the impact of the latest generations (such as GPT-5 or GPT-6) faces a major obstacle: a lack of corporate transparency and the absence of official, audited data. As the models have grown in complexity, the developers have chosen not to disclose sustainability metrics. Consequently, it falls to the scientific community to estimate the impact using indirect engineering models. Based on these calculations, it is projected that training systems such as GPT-4 increased electricity consumption to around 70,000 MWh and the direct water footprint to tens of millions of litres of water, about 15 Olympic-sized swimming pools.
On the other hand, the physics of supercomputing presents new technological challenges. The high thermal density of the latest-generation chips makes it impossible to use traditional evaporative cooling. This necessitates a shift towards alternative liquid-cooling architectures, ranging from sealed loops connected directly to the processor’s cold plates (Direct-to-Chip) to the complete immersion of the hardware in tanks containing non-conductive dielectric fluids. This would mean less water consumption, but would result in other associated costs.
For all the reasons mentioned above, the sustainability of artificial intelligence requires the adoption of algorithmic optimisation strategies (Green-IN). Techniques such as transfer learning, which allows pre-trained models to be reused, and parameter pruning, which removes redundant connections in models, have demonstrated significant reductions in sustainability metrics.
References
- Aczel, M., Chamanara, S., Matin, M., Farsi, A., Marwala, T., & Madani, K. (2026). Environmental cost of AI’s energy use: Carbon, water and land footprints. United Nations University Institute for Water, Environment and Health (UNU-INWEH).
- Jegham, N., Abdelatti, M., Koh, C. Y., Elmoubarki, L., & Hendawi, A. (2025). How hungry is AI? Benchmarking energy, water, and carbon footprint of LLM inference (arXiv:2505.09598). arXiv.
- Li, P., Yang, J., Islam, M. A., & Ren, S. (2025). Making AI less “thirsty”. Communications of the ACM, 68(7), 54–61.
- Moreau, I. (2026). AI energy consumption statistics. Gitnux.
- Morrison, J., Na, C., Fernandez, J., Dettmers, T., Strubell, E., & Dodge, J. (2025). Holistically evaluating the environmental impact of creating language models (arXiv:2503.05804). arXiv.
Researcher at the Data Science and Big Data Lab of Pablo de Olavide University (Seville, Spain), where he is pursuing a PhD in Computer Science. He holds a Bachelor’s Degree in Computer Engineering in Information Systems (2022) and a Master’s Degree in Computer Engineering (2024) from the same university. His research focuses on efficient machine learning and deep learning models for real-world sustainability challenges, such as energy efficiency, pollution reduction and frugal artificial intelligence. Since 2023/2024 he has taught Programming Fundamentals, Software Engineering and Digital Forensics at Pablo de Olavide University.