The Hidden Costs of AI: Why "Weightless" Technology is Trashing the Planet

 The Thirst of Artificial Intelligence

 

Eight in ten Britons are polite to AI chatbots, but here's what they don't know: every extra word costs water⁷. A single average conversation with ChatGPT (GPT-3) consumes roughly 500ml — the equivalent of a standard water bottle⁴. With GPT-5, just 13 medium-length queries use the same amount⁴. Considering OpenAI processes around 2.5 billion prompts daily⁴, the scale is staggering. Yet, AI companies publish no real data on water consumption per query, leaving users completely in the dark

(Image 1)

This is a classic example of a negative consumption externality which is where a user of AI experiences private marginal benefit (like getting Chat to “help” you with your essay) but is completely unaware of the hidden environmental social costs. Ultimately, this leads to an overconsumption of AI prompts relative to the socially desired outcome (which is shown by the triangle shaded in red).

(Image 2)

 

The aggregate picture is alarming. Big Tech's combined water usage rose 60% between 2020 and 2023, with Microsoft alone seeing an 87% increase¹¹. Water is consumed both to cool data centre servers and at the power plants generating the electricity to run them. US AI data centres used enough water to supply 15 million households in 2022, and local communities bear the brunt⁵. For example, during GPT-4's training, a single Iowa data centre cluster consumed 6% of the district's entire monthly water supply³. Residents had no say, reflecting a negative production externality. In this case the firm (GenAI) experiences the marginal private benefit (training Chat to be better) and third parties (the residents in Iowa) incur the external costs (less access to water).

 

A Physical Backbone Built on Extraction

 

Building the physical backbone of AI, from data centres to semiconductors and microelectronics demands vast quantities of minerals, including silicon, lithium, gallium and graphite. Extracting these materials depletes groundwater, contaminates soils and accelerates deforestation. The uncomfortable truth is that none of those costs appear on any tech company's balance sheet!

This is a textbook case of negative externalities. The firms profiting from AI bear the private costs of production, such as servers, software, and salaries. However, the environmental destruction caused by mineral extraction is externalised onto local communities and ecosystems. The social cost of AI far exceeds its private cost and the gap between the two represents a market failure hiding in plain sight.

The scale of this failure is staggering. Producing a single 2 kg computer requires roughly 800 kg of raw material inputs and the rare earth elements powering AI chips are routinely extracted through environmentally destructive processes¹². Beyond extraction, AI hardware generates toxic e-waste containing mercury and lead, while data centre cooling is projected to consume water approaching six times Denmark's entire annual usage ¹². By 2030, AI expansion could produce up to five million tonnes of additional electronic waste and nearly quadruple cooling water consumption to 664 billion litres⁸. The "cloud" might sound weightless, but the infrastructure behind it is anything but. When markets fail to price in these costs, overproduction is inevitable, and somebody always pays.

(Image 3)

 

Energy Use and Carbon Emissions

 

The rapid expansion of artificial intelligence has increased electricity use, especially in training and deploying large-scale models. Firms bear the private cost of energy but not the full social cost of emissions, leading to overproduction. Why does this matter? Because these environmental costs are not reflected in prices, resulting in excessive energy use.

Empirical evidence highlights the scale of this issue. Training large natural language processing models requires substantial computational power and can generate emissions comparable to the lifetime emissions of multiple cars¹⁰. Data centres—critical infrastructure for AI—already account for a growing share of global electricity demand, a trend expected to accelerate with increasing AI adoption².The impact is particularly severe in regions reliant on fossil fuels. Because firms do not internalise these environmental costs, the outcome is inefficient, imposing broader societal costs such as climate change and environmental degradation.

This highlights a broader issue: technological progress does not always lead to efficient outcomes. When firms do not internalise environmental costs, innovation can shift part of the burden onto society through climate change and environmental damage.

(Image 4)

 

Policy Solutions

 

After establishing the need for intervention, here are a few policies we suggest addressing the real costs of artificial intelligence:

·   Marginal pricing of AI usage: Introducing a free quota of AI prompts means after that you pay a fee per additional prompt, limiting excessive use. Students will reserve their quota for high value use ONLY if they want to maximise their benefit from AI usage. The risk, however, is that this policy is regressive; wealthier students could afford to continue their high-value usage as they can just keep paying. 

 

·   Institutional restrictions: Universities across the nation could ban AI platforms on their university networks (e.g. blocking them on the eduroam network) to change student habits. This stops the consumption externality at its source and prevents students from using AI on campus! Students will switch back to lectures notes and resources provided by universities that provide a deeper understanding.

 

·   Pigouvian Taxes and Innovation: A Pigouvian tax on AI energy and water use, forces firms to face the true social cost of production. We know green innovation is achievable, since Microsoft's new direct-to-chip liquid cooling systems use a closed-loop design requiring no evaporation, saving over 125,000 cubic metres of water per data centre annually⁶. The technology exists; however, firms lack the incentive to implement it. The tax will create the incentive

 

Conclusion

 

In the future when you are considering using AI for a task, have a think about the externalities associated and if it is necessary to use an AI tool for your objective. 

 

Also, did you notice anything odd about the images? - You may have assumed that only image 4 is AI generated. However, all 3 images and the graph are generated by Gemini 3!

 

Bibliography

1. Grantham Research Institute (2025) What direct risks does AI pose to the climate and environment? London: London School of Economics and Political Science. Available at: https://www.lse.ac.uk/granthaminstitute/explainers/what-direct-risks-does-ai-pose-to-the-climate-and-environment/ 
2. International Energy Agency (2024) Electricity 2024. Paris: IEA.
3. Kleinman, Z. and Wheeler, B. (2025) Concern the UK’s AI ambitions could lead to water shortages, BBC News. Available at: https://www.bbc.co.uk/news/articles/ce85wx9jjndo 
4. Lo, L. (2025) AI has a hidden water cost – here’s how to calculate yours, The Conversation. Available at: https://theconversation.com/ai-has-a-hidden-water-cost-heres-how-to-calculate-yours-263252 
5. Mendez, L., Baxter, S. and Ivanova, N. (2026) AI water usage statistics, WifiTalents. Available at: https://wifitalents.com/ai-water-usage-statistics/ 
6. Microsoft (2025) 2025 Environmental Sustainability Report. Available at: https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/microsoft/msc/documents/presentations/CSR/2025-Microsoft-Environmental-Sustainability-Report.pdf 
7. Milton, J. (2025) Here’s why saying ‘please’ and ‘thank you’ costs 158,000,000 bottles of water a day, Metro. Available at: https://metro.co.uk/2025/08/08/saying-please-thank-you-costs-158-000-000-bottles-water-a-day-23091100/ 
8. Öko-Institut (2025) Environmental Impacts of Artificial Intelligence. Hamburg: Greenpeace Germany. Available at: https://www.oeko.de/en/publications/environmental-impacts-of-artificial-intelligence/ 
9. Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L.M., Rothchild, D., So, D., Texier, M. and Dean, J. (2021) ‘Carbon emissions and large neural network training’, arXiv. Available at: https://arxiv.org/abs/2104.10350 
10. Strubell, E., Ganesh, A. and McCallum, A. (2019) ‘Energy and policy considerations for deep learning in NLP’, arXiv. Available at: https://arxiv.org/abs/1906.02243 
11. Surfshark (2025) Big Tech’s water usage has risen 60% since 2020. Available at: https://surfshark.com/research/chart/big-tech-water-usage United  Nations 
12. Environment Programme (2024) AI has an environmental problem. Here's what the world can do about that. Nairobi: UNEP. Available at: https://www.unep.org/news-and-stories/story/ai-has-environmental-problem-heres-what-world-can-do-about 

Image 1-4: Generated by Google Gemini 3