Published: 29 June 2025, 08:30 BST | Updated: 29 June 2025, 08:30 BST
By Stuart Kerr, Data Journalist, Live AI WireAs artificial intelligence systems grow more sophisticated, their environmental footprint is becoming impossible to ignore. Behind every ChatGPT conversation and AI-generated image lies an often-overlooked reality: training these models demands staggering amounts of energy and water, with consequences for both climate change and local communities.
The Energy Toll of Intelligence
Training large language models like GPT-4 requires enough electricity to power small towns. The 2019 training of GPT-3 consumed 1,287 MWh – equivalent to 120 American households' annual consumption – while emitting 552 tonnes of CO₂ according to research by Strubell et al. published in arXiv. More recent models are even more demanding, though tech companies guard exact figures closely.
"The computational intensity of generative AI dwarfs traditional computing workloads," explains Dr. Noman Bashir of MIT's Climate and Sustainability Consortium. "We're seeing power densities seven to eight times higher than conventional data centre operations."
The environmental impact extends beyond initial training. Each ChatGPT query consumes roughly seven times more energy than a Google search according to Hugging Face researcher Dr. Sasha Luccioni's 2023 analysis. With billions of interactions daily, these incremental costs accumulate rapidly.
The Dirty Secret of Clean Energy Claims
Tech giants proudly announce commitments to renewable energy, but reality often falls short. While Google's Finland data centre operates on 97% carbon-free energy, its Asian facilities remain heavily dependent on coal power.
"Renewable energy credits create a facade of sustainability," Luccioni notes. "Many companies purchase clean energy from distant grids while their local operations still draw from fossil fuels."
The numbers reveal troubling trends:
Microsoft's emissions rose 29% since 2020
Google's carbon footprint grew 48% since 2019
Both companies attribute these increases to data centre expansions for AI development.
Water: The Overlooked Resource
In drought-stricken regions, AI's thirst is becoming controversial. Google's data centres in Oregon used over a quarter of The Dalles' water supply in 2023 for cooling systems. Similar concerns have emerged in Chile and Uruguay where new facilities are planned.
"Water-intensive cooling creates environmental justice issues," warns Dr. Shaolei Ren of UC Riverside. "The burden falls disproportionately on communities hosting these data centres."
Pathways to Sustainable AI
The industry is exploring solutions:
Efficient Models: Hugging Face's BLOOM language model emitted just 25 tonnes of CO₂ compared to GPT-3's 552 tonnes
Hardware Innovations: MIT's Clover software reduces emissions by 90% by shifting workloads to low-carbon energy sources
Renewable Investments: Solar-powered Australian data centres show promise for reducing grid dependence
However, Dr. Roy Schwartz of Hebrew University cautions: "The obsession with ever-larger models directly contradicts sustainability goals. We're prioritising scale over efficiency."
The Transparency Imperative
The EU's AI Act now mandates emissions disclosures for large models, but enforcement remains patchy. "True progress requires radical transparency," argues Dr. Vijay Gadepally of MIT Lincoln Lab. "We need open data on training methods, energy sources, and hardware efficiency."
As AI becomes embedded in daily life, its environmental impact can no longer be an afterthought. The challenge ahead lies in balancing technological progress with ecological responsibility – before the costs become unsustainable.
Sources:
Strubell, E. et al. (2019). Energy and Policy Considerations for Deep Learning. arXiv:1906.02243
Luccioni, S. et al. (2023). Estimating ChatGPT's Energy Footprint. arXiv:2306.03241
International Energy Agency (2024). Data Centres and Energy Demand
Microsoft (2024). Sustainability Report
About the Author:
Stuart Kerr is a data journalist specialising in AI's societal impacts. Follow his reporting at @liveaiwire.
