Behind the prompt: The environmental cost of using AI By: Imaan Ghoor

I use ChatGPT almost daily. Sometimes I ask it to help me reword an awkward sentence. Other times I use it to summarise documents or give me quick insights on topics I don’t have time to dig into. It’s efficient, and it feels like magic. But what I hadn’t really thought about until recently was what powers that magic.

Every time I hit "send" on a prompt, somewhere in the world, massive servers spin up. These are housed in data centers: huge buildings filled with racks of computers, all working together to produce a response in seconds. But they don’t run on magic. They run on electricity, and they need to be cooled constantly often with water that could otherwise be used for farming, drinking, or keeping rivers alive. That’s where the environmental story of AI starts to get uncomfortable.

By the end of 2024, developers in the U.S. had approved over 1,240 data centers, nearly four times the number that existed in 2010. These facilities are power-hungry. Together, they’re expected to consume around 239 terawatt-hours (TWh) of electricity annually, which is roughly the same as what the entire state of Florida used in 2023 (Business Insider, 2025). Globally, estimates suggest data centers already use between 1 and 1.5% of all electricity, with projections suggesting that number could double by 2030. This rise is being driven largely by the explosion in AI use (IEA, n.d). 

This global trend has real implications for countries like South Africa, which is both energy and water constrained. In 2023, almost 6,950 hours of load shedding occurred in South Africa. This was significantly higher than the previous year, which reported 2,400 hours of power cuts in the country (Statista, n.d.).  At the same time, South Africa described as a water-scarce country with deeply entrenched challenges: our water infrastructure suffers from poor management, inadequate maintenance, limited investment, and inefficient monitoring systems. In Gauteng an estimated 46% of water is lost due to leaking and failing infrastructure (Department of Water and Sanitation, 2025).

It’s tempting to believe that these centers are green-after all, tech companies often talk about their renewable energy commitments. But the reality is that many still rely on fossil fuels, either directly or through the grid mix. And even when companies buy renewable energy credits or install solar panels, it doesn't always match when and where the energy is actually used. The result is a growing burden on already stretched electricity grids and in some cases, communities are footing the bill.

According to an investigation by Business Insider, the public health costs of AI’s power demands in the U.S. alone could range from $5.7 billion to $9.2 billion annually. That includes hundreds of premature deaths and hundreds of thousands of asthma cases, largely from fossil-fuel-related air pollution (Business Insider, 2025). What’s even more worrying is that many of these centers are being built in areas with existing energy constraints leading to new infrastructure needs, like power lines and substations, that often take years and billions to build (Spencer and Singh, 2024).

Then there’s the issue of water. This one surprised me more than anything else. To keep servers from overheating, most data centers rely on water-based cooling systems. That means that while AI seems digital and intangible, its footprint is very physical. A single 1 MW data center can use up to 6.6 million gallons of water per year for cooling. That’s about 25 million litres (Purite, 2023).

In a water-scarce country like South Africa, this kind of demand poses significant risks. Analysis done by the Water Resources Group projects a 17% gap between supply and demand by 2030 if no action is taken (Department of Water Affairs, n.d.). In 2022 cities like Nelson Mandela Bay and Gqeberha were on the brink of “Day Zero” water shortages (Makoni, 2022). And yet, infrastructure such as data centers is growing without consistent transparency around their local water use.

In 2022, Google reported that its data centers consumed 4.3 billion gallons (roughly 16 billion litres) of water. Microsoft wasn’t far behind, using over 1 billion litres that year. In some areas, this demand increased total water use by more than 30%. In others, data centers were consuming as much as a quarter of the city’s entire water supply (TechRadar, 2022).

Closer to home, Amazon Web Services (AWS) has built multiple data centers in Cape Town—a water-stressed city still recovering from the 2018 “Day Zero” crisis. In December 2022, AWS announced its commitment to become “water positive” by 2030, pledging to return more water to communities than it consumes. Cape Town is one of the first 20 global locations where AWS uses recycled water for cooling, and it sources sustainable water through rainwater harvesting and reuse systems (Goldstuck, 2022).

In addition, AWS has partnered with The Nature Conservancy in the Greater Cape Town Water Fund to remove water-hungry invasive plant species from catchment areas. To date, more than 47,000 acres have been cleared, restoring native fynbos and returning an estimated 8.9 billion litres of water to the regional supply annually (Goldstuck, 2022). Once all of AWS’s global water replenishment projects are completed, they are expected to return 2.4 billion litres of water per year to communities (Goldstuck, 2022). However, there are no publicly disclosed figures on exactly how much water or electricity Amazon data facilities use in South Africa.

The thing is, water is not an unlimited resource, especially in the places where many of these data centers are being built. In England, regulators are already warning that rising demand for water from AI infrastructure could push the country towards a 5 billion litre per day shortfall by 2055 (Business Insider, 2025). In Oregon, Google was criticised for drawing huge volumes of water in a city already dealing with drought (Oregon Tech, 2022). 

And this is only the beginning. By 2027, AI-related water usage is expected to hit 6.6 billion cubic meters (Hao, 2024). Global AI data centers are projected to use six times more water than Denmark consumes annually (Li et al., 2025).  So even though using AI feels lightweight, its infrastructure is heavy and it’s growing rapidly. This made me think about the smaller, everyday uses. Each ChatGPT query uses a fraction of a watt, maybe 2.9 watt-hours per prompt, which is ten times more than a Google search (Times of India, 2024). 

A paper published in late 2024 estimates that generating between five and 50 ChatGPT prompts can require around 500 ml of fresh water. The exact amount varies depending on the server’s location and the time of year. This figure also accounts for indirect water use, including the cooling of power plants that provide electricity to data centers (Li et al., 2025).  That adds up, especially when you think about how often we use AI for things that aren’t really necessary, like asking it to write a birthday caption or joke reply.

This isn’t to say we should stop using AI, that’s neither realistic nor the goal. But we need to ask: how sustainable is this rapid growth? And what are the trade-offs for countries already under pressure like ours? There are ways to improve things. Some companies are investing in air-cooled or closed-loop systems that dramatically reduce water use. Others are exploring immersion cooling or even using wastewater or seawater instead of fresh drinking water. Google has committed to replenishing more water than it uses and improving watershed health in water-stressed communities as part of its broader sustainability goals (Google, 2022).

But these are still the exception, not the rule. Most data centers are still drawing on municipal water supplies, and most still rely on electricity generated at least partly from fossil fuels. And most of us, as users, aren’t really aware of it.

This matters. It matters because AI isn’t just some abstract force. It’s built on real-world resources such as power grids, water pipes, cooling towers, lithium batteries. Every improvement in convenience comes with a cost. We might not feel it today, but someone somewhere is paying for it through higher water bills, or power cuts, or polluted air.

I’m not anti-AI. I use it. But I think we all need to be more honest about what we’re using and what it costs. Tech companies should be required to report on their energy and water use publicly. Governments need to regulate where these centers can be built, and how they impact local communities. And we, as users, should be more thoughtful about how often and for what we use these tools.

Not every prompt is worth the environmental toll. And if we want AI to be truly smart, it should be sustainable too.

-----------------------------------------------------------

References 

Business Insider. (2025). The 'AI boom' pits neighbor against neighbor. https://www.businessinsider.com/data-centers-northern-virginia-noise-air-pollution-cost-2025-5

Business Insider. (2025). Tallying the true costs of AI. https://www.businessinsider.com/ai-data-center-cost-2025-6

Business Insider. (2025). AI runs on dirty power — and the public pays the price. https://www.businessinsider.com/ai-runs-dirty-power-and-the-public-pays-the-price-2025-6

Business Insider. (2025). The true cost of data centers. https://www.businessinsider.com/ai-data-center-development-true-cost-environmental-impact-2025-6

Department of Water Affairs. (n.d.). Strategic water partners network South Africa closing the water gap by 2030. https://www.sab.co.za/sites/g/files/seuoyk1916/files/sab/22.-swpn-water-project-final-low-res-min.pdf

Department of Water and Sanitation. (2025). Water Conservation & Demand Management https://ws.dws.gov.za/wsks/DefaultList.aspx?SubjectAreaID=6&DataTopicDetailID=12&DisplayTypeId=1&PerspectiveID=0&LvlID=10&DataTopicID=24

Goldstuck, A. (2022). Future cloud won’t hold water. Amazon Web Services aims to be “water-positive” in its data centres, and Cape Town will be among the first. https://gadget.co.za/future-cloud

Google. (2022). Our commitment to climate-conscious data center cooling. The Keyword. https://blog.google/outreach-initiatives/sustainability/our-commitment-to-climate-conscious-data-center-cooling/

Hao, K. (2024). AI is taking water from the desert. The Atlantic. https://www.theatlantic.com/technology/archive/2024/03/ai-water-usage-data-centers-desert/677622/

International Energy Agency [IEA]. (n.d.). The transformative potential of AI depends on energy. https://www.iea.org/reports/energy-and-ai/executive-summary

Khan, N. Inside Tech World (2025). https://insidetechworld.com/artificial-intelligence/chatgpt-water-consumption-and-environmental-impact/

Li, P., Yang, J., Islam, M. A., & Ren, S. (2025). Making AI less “thirsty”: Uncovering and addressing the secret water footprint of AI models. arXiv. https://arxiv.org/pdf/2304.03271

Makoni, M. (2022). After Cape Town, another place in South Africa is fighting off a ‘Day Zero’ water crisis. https://www.downtoearth.org.in/africa/after-cape-town-another-place-in-south-africa-is-fighting-off-a-day-zero-water-crisis-84639

Oregon Tech. (2022). Google’s water use is soaring in The Dalles, records show, with two more data centers to come. https://www.oregonlive.com/silicon-forest/2022/12/googles-water-use-is-soaring-in-the-dalles-records-show-with-two-more-data-centers-to-come.html

Purite. (2023). Water usage breakdown in data centres: Essential facility areas. https://purite.com/water-usage-breakdown-in-data-centres-essential-facility-areas/

Spencer, T., & Siddarth, S. (2024). What the data centre and AI boom could mean for the energy sector. International Energy Agency. https://www.iea.org/commentaries/what-the-data-centre-and-ai-boom-could-mean-for-the-energy-sector

Statista.  (n.d.). Total number of load shedding hours in South Africa from 2015 to 2023. https://www.statista.com/statistics/1459776/number-of-load-shedding-hours-in-south-africa/

TechRadar. (2022). Google's data center water usage is revealed – and it's a lot. https://www.techradar.com/news/googles-data-center-water-usage-is-revealed-and-its-a-lot

Times of India. (2024). ChatGPT, Google Search need power to run: Here’s how much water and electricity are used to answer questions. https://timesofindia.indiatimes.com/technology/tech-news/chatgpt-google-search-need-power-to-run-heres-how-much-water-and-electricity-are-used-to-answer-questions/articleshow/111382705.cms