What is the environmental impact of AI?
Since the launch of ChatGPT in late 2022, a lot of people have started to use generative AI in a very short time: to generate text, images, music, videos, computer code, etc. The tools owe their success partially to their user-friendly interfaces that mask the particularly complex infrastructure behind them.
As Sam Altman, CEO of OpenAI (the company behind ChatGPT), puts it, it ‘feels like magic’. However, the reality is that for all their simplicity, these tools consume huge amounts of power, water and raw materials. Wim Casteels of AP Hogeschool exposes the biggest challenges.
“ not gpt-5, not a search engine, but we’ve been hard at work on some new stuff we think people will love! feels like magic to me. ”
Measuring ecological footprint
To measure the ecological footprint of a technology, you first need to identify in what phases there is an environmental impact. Pollution and consumption can occur at different stages of the AI life cycle, such as:
- extraction of resources
- production of server infrastructure
- AI model development
- AI model use
- waste treatment
We don’t know everything yet: the full ecological footprint of most phases is unclear because AI processes are complex and the companies developing them are not always transparent.
For now, most research is focused on the impact of the development and use of AI models, as those phases are more easily measured.
Why does AI use so much power?
AI models are getting bigger and bigger. According to empirical scaling laws, more data, parameters and computing power lead to better accuracy. As a result, AI models have grown enormously in recent years and require more and more computing power.
This is measured in FLOPS or ‘Floating Point Operations’, an elementary computer calculation. The graph illustrates this by showing the number of FLOPS required for the development of each model. Performing all these FLOPS required huge amounts of electricity.
How much electricity is needed to train GPT-4?
While very little information is available about most models, but there is information about GPT-3. This 2020 OpenAI model, with 175 billion parameters, made the first version of ChatGPT possible and launched the global popularity of generative AI. The development of GPT-3 used 1,287 MWh, equal to the annual consumption of 500 Flemish households. This also emitted 552 tonnes of CO2 for development, the equivalent of 120 petrol cars a year.
Compared to the newer AI models, these are rookie numbers: it is estimated that GPT-4 has 1.8 trillion parameters, while Google’s Gemini Ultra is said to be even larger.
How much does the use of AI consume?
How much power is needed for one AI prompt?
After development, using an AI system continues to consume power, even more than during development. For example, ChatGPT – with millions of daily users – exceeded amount of the power it consumed for training (1,287 MWh) after only a few weeks.
A typical ChatGPT prompt needs about 0.3 Watt-hours (Wh). This is less than turning on an LED lamp or using a laptop for several minutes but varies greatly depending on the length of the prompt.
Power consumption varies greatly per prompt: images require more energy than text – generating one image consumes as much energy as charging a smartphone (220 Watt-hours).
Which one is better, Google Search or ChatGPT?
According to Google, a query via their AI model costs the company 10 times more power than a traditional Google Search.
Lots of users = lots of prompts = lots of consumption
Most people ask GenAI more than one question, and then the real impact becomes clear: 400 million users send more than a billion messages to OpenAI’s ChatGPT every week. And OpenAI is not alone: there is also Anthropic’s Claude, Google’s Gemini and others like China’s DeepSeek
– which dethroned ChatGPT as the most popular app in the Android Play Store and iPhone App Store in January. Moreover, AI is increasingly being used in other applications, such as KBC’s AI assistant Kate.
AI consumes power in data centres
In 2022, global electricity consumption from data centres was estimated at 240-340 TWh, about 1-1.3% of global electricity demand. Just two years earlier, in 2020, data centres and data transmission accounted for 0.6% of total greenhouse gas emissions, comparable to the total emissions of the aviation industry.
These data centres are not only used for AI. AI is currently said to account for 10-20% of electricity use, a share that is expected to rise at lightning speed.
How much water does AI consume?
Less than half of the electricity used by AI in data centres is used to power the servers: just like a laptop has fans to cool down during use, server rooms need to be actively cooled by dedicated air conditioners. Cooling data centres packed with servers is quite a challenge: cooling devours around 30–40% of total power consumption.
And that cooling process also requires large amounts of water to dissipate the heat – water that mostly evaporates. Research estimates that GPT-3, the AI model behind ChatGPT, consumes 500ml of water for every 10 to 50 prompts. As with energy, the same applies to water: the newer the model, the greater the consumption.
And raw materials?
To make AI servers, several metals are needed whose mining causes great ecological damage. Moreover, some metals, such as cobalt and tungsten, are so-called conflict minerals mined in conflict zones. The mining and trade of these minerals contribute to human rights violations and armed conflict.
Is sustainable AI possible?
The huge electricity requirements of AI systems have caused several technology giants such as Microsoft and Google to miss their own climate targets and fail to achieve the desired greenhouse gas emission reductions. These companies are now looking for alternative solutions to still achieve their climate ambitions.
Several of the approaches that are currently being explored:
- The use of green energy from renewable energy to power data centres. This has made Iceland an attractive location for data centres because the country has geothermal and hydroelectric energy. The cold climate also means less power and water are needed for cooling. But environmentalists question whether this is a good use of energy and whether it would not be better used for Iceland itself.
- Nuclear power. Although nuclear power does not emit greenhouse gases, the technology has other challenges such as safety risks and radioactive waste disposal. In Pennsylvania, Microsoft agreed to reopen the Three Mile Island nuclear power plant, a plant known for a partial nuclear meltdown in 1978.
- Building data centres in space. Ascend (Europe) and Starcloud (USA) are looking for an extraterrestrial solution: space. It has two main advantages: all-day solar power and extreme cold for cooling. But it isn’t cost-effective (for now?) to launch the necessary equipment, despite costs are steadily decreasing.
Global race for more computing power
Meanwhile, the power hunger of AI applications continues to grow unabated. US President Donald Trump announced the Stargate project with a planned $500 billion investment in AI data centres.
French President Emmanuel Macron presented nuclear overcapacity as France’s key asset at the 2025 AI Action Summit in Paris. In an implicit response to the American slogan ‘drill, baby, drill’, he described the French approach as ‘plug, baby, plug’.
China’s DeepSeek looks for the solutions by focusing on reducing power consumption and increasing efficiency, which sent shockwaves through the AI and financial world. The long-term impact of that increased efficiency remains unpredictable for now: Jensen Huang, CEO of chipmaker NVIDIA, expects demand for servers to further increase as a result (see: Jevons paradox).
An incomplete picture
We know that AI consumes a lot of energy and a lot of water and requires even more resources, but we don’t have the full picture yet:
- There is too little information about server production, industrial waste from data centres and associated environmental hazards.
- Using AI to make polluting industries more efficient can paradoxically lead to more pollution, just like in oil and gas exploration did.
- AI infrastructure also contributes to the growing mountain of e-waste, estimated at up to 5 million tonnes by 2030.
Thoughtful decisions and innovative solutions
Above all, the environmental impact of AI calls for a broad public debate. Only through thoughtful decisions and innovative solutions can we make technological advances that are not at the expense of the environment:
- Businesses, governments and consumers must work together to accelerate the energy transition and to develop sustainable alternatives.
- When developing and using AI systems, the environmental impact needs to be factored in, so that can we ensure that AI contributes to a sustainable future, instead of undermining it.
Sustainability as a guiding principle at AP University of Applied Sciences and Arts Antwerp
AP University of Applied Sciences and Arts Antwerp takes these guidelines to heart: sustainability is a key guiding principle. For our IT courses, this means the following:
- In the Sustainable IT course, students gain insight into the environmental impact of the IT applications they develop.
- For IT students specialising in AI in the Bachelor IT & Artificial Intelligence, the AI & Society course takes a deeper look at the impact of AI on society.
- The MDI research group focuses on sustainable AI applications by developing specialised systems that use a relatively small AI model with a limited ecological footprint. An example is the AI4Care project, in which we are developing a specialised AI tool to be used for support services for young people.
References
- Kaplan, J., McCandlish, S. Henighan, T. et al. (2020) Scaling Laws for Neural Language Models, Cornell University
- Varoquaux, G., Sasha Luccioni, A. & Whittaker, M. (2024) Hype Sustainability, and the Price of the Bigger-is-Better Paradigm in AI, Cornell University
- Dastin, J. & Nelis, S. (2023) Focus: For tech giants, AI like Bing and Bard poses billion-dollar search problem, Reuters
- You, J. (2025) How much energy does ChatGPT use?, EPOCH AI, Gradient Updates
- Kant, R. (2025) OpenAI's weekly active users surpass 400 million, Reuters
- Zeff, M. (2025) DeepSeek displaces ChatGPT as the App Store's top app, TechCrunch
- De Cleene, D. (2024) En de KBC-werknemer van het jaar is... "Kate": hoe de AI-assistente de bank naar nieuwe hoogtes stuwt, De Morgen
- Data Centres and Data Transmission Networks (2023) Tracking Clean Energy Process, IEA
- Powering Intelligence: Analyzing Artificial Intelligence and Data Center Energy Consumption (2024) White Paper EPRI
- AI is poised to drive 160% increase in data center power demand (2024) Goldman Sachs
- Pengfei, L. Yang, J., Islam, M. A. & Ren, S. (2025) Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models, Cornell University
- Rathi, A. & Bass, D., Microsoft's AI Push Imperils Climate Goal as Carbon Emissions Jump 30% (2024), Bloomburg
- Metz, R., Google's Emissions Shot Up 48% Over Five Years Due to AI (2024), Bloomburg
- Adalbjornsson, T. (2019) Iceland’s data centers are booming—here’s why that’s a problem, MIT Technology Review
- Brinson, K. (2024) Hungry for Energy, Amazon, Google and Microsoft Turn to Nuclear Power, The New York Times
- Crownhart, C. (2024) Why Microsoft made a deal to help restart Three Mile Island, MIT Technology Review
- Advanced Space Cloud for European Net Zero Emission and Data Sovereignty, ASCEND
- Starcloud
- Smith-Goodson P. & Kimball, M. (2025) The Stargate Project: Trump Touts $500 Billion Bid For AI Dominance, Forbes
- Caulcutt, C. (2025) ‘Plug, baby, plug’: Macron pushes for French nuclear-powered AI, Politico
- Metz, C., What to Know About DeepSeek and How It Is Upending A.I., The New York Times
- Masood, A. & Bhattacharya, A. (2024) Microsoft is building a data center in a tiny Indian village. Locals allege it’s dumping industrial waste, Rest of World
- Greenpeace Staff (2020) Greenpeace Report: Oil in the Cloud, Greenpeace.org
- Sellman, M. (2024) 'Thirsty’ ChatGPT uses four times more water than previously thought, The Times
- Tzackor, A. (2024). Generative AI could create millions of tons of e-waste by 2030. MIT Technology Review
Learn more about sustainability in IT
Introduction to AI for the Common Good
1-day course - Brussels - FARI
Leuven-Bonn-Delft Workshop on AI and Sustainability
workshop - Leuven & online - KU Leuven, University of Bonn, TU Delft
Best practices for sustainable GenAI
learning community meet-up - Brussels - Knowledge Centre Data & Society, sustAIn.brussels
Wim Casteels
Wim Casteels obtained his PhD in physics at the University of Antwerp, after which he spent several years researching the quantum behaviour of atoms and light particles in Antwerp and at the Université Paris Diderot in Paris. He then shifted his focus more towards data and Artificial Intelligence and worked as a data scientist at Argenta’s Data Analytics Office. After gaining 2 more years of academic experience as a senior researcher at the imec research group IDLab, he started working at the AP Hogeschool in early 2022 where he both researches and teaches about AI and building data applications.
Expertise: Data Science, Machine Learning, AI deployment, AI in education, Learning Analytics, Deep Learning, AI4GOOD
Related courses
HPC Café
Seminarie - Online, Karlsruhe - KIT, HammerHAI, Training geselecteerd in de Belgische AI Factory Antenna (AIFA)
SAIAR Summer Studio: From latent to physical space
Zomerschool - Kortrijk - Howest Hogeschool, SAIARlab