How Does AI Affect Water in South Africa?

There is no shortage of alarming global headlines about artificial intelligence and water. Training GPT-3 was estimated to evaporate around 700,000 litres of clean freshwater in Microsoft's US data centres. The International Energy Agency puts total data-centre water consumption at roughly 560 billion litres in 2023, on a path to roughly double by 2030. AI's physical footprint, some argue, represents a new form of extraction.

But how much of this applies to South Africa, specifically? The answer is more nuanced, and in some ways more interesting, than the headlines suggest.

Before examining AI's role, it is worth being clear about the context. South Africa is a water-scarce country by global standards, with an average annual rainfall of around 465 mm and a high evaporation rate. The Department of Water and Sanitation's National Water and Sanitation Master Plan records that a large share of the country's water resources are already under stress. Gauteng, the economic engine of the country, is heading toward a water deficit, with scientists warning of near-term climate tipping points.

South Africa is projected to face a water deficit of roughly 17% by 2030, a gap of between 2.7 and 3.8 billion cubic metres. Per capita, South Africans already consume more water per day than the global average. Meanwhile, ageing infrastructure means non-revenue water sits at 47.3% nationally — nearly half of all treated water lost to leaks, faulty metering, illegal connections and billing failures before it reaches anyone.

This is the backdrop against which any new demand on water resources, including from the digital economy, must be assessed.

AI does not use water in the way that farming or manufacturing does. The water connection comes primarily through cooling. Data centres house thousands of servers that generate significant heat. To prevent equipment damage, that heat must be removed, and water is one of the most effective and widely used coolants.

Large-scale evaporative cooling towers work by allowing heat to escape through water vapour. This is efficient but consumes real water. Some modern facilities use closed-loop or air-cooled systems that reduce on-site water use, but may trade off in energy efficiency or require more electricity.

There is also an indirect water cost through electricity generation. Globally, the International Energy Agency estimated that in 2023, data centres consumed around 560 billion litres of water in total, of which roughly two-thirds (373 billion litres) was tied to indirect consumption from power generation rather than direct cooling. In South Africa, this indirect dimension is particularly significant: the national grid remains heavily dependent on coal-fired power, and Eskom's coal plants use water both for cooling and upstream in coal mining and processing. Power generation already accounts for approximately 37% of water use in the Upper Olifants catchment, one of the country's most water-stressed regions.

Here is where the global narrative diverges from local evidence.

South Africa's data centre market is growing, but it remains modest by global standards. Teraco, the continent's largest operator, publishes a total power capacity of 189 MW across its Johannesburg, Cape Town and Durban facilities — a fraction of the capacity operated by US or European hyperscalers. The country hosts established players including Teraco (the continent's largest colocation operator), Equinix, BCX, and NTT, among others. Teraco functions as a "data centre hotel". It provides the facility, power, and cooling infrastructure within which global cloud providers house their own equipment. Its water footprint is measurable and reported.

The most significant piece of local evidence comes from a Carnegie Mellon University Africa study, the first of its kind to model water usage efficiency for data centres across 41 African countries. It found that South Africa and Morocco consume less water per unit of AI computation than the United States average. Nine of eleven African countries studied came in below the global average. South Africa's relatively mild Highveld climate and the use of dry-cooling technology at some facilities contributes to this outcome.

Figures for South Africa's total data centre water consumption circulate widely, but the ones we could trace lead back to market-research models rather than to metered, audited consumption. The CMU dataset is the only source here that publishes its method. Everything else should be treated as indicative rather than definitive — and we have removed the projections we could not attribute.

To put the numbers in perspective: irrigated agriculture accounts for close to 60% of South Africa's freshwater use, and Eskom's power generation for roughly 2% of national consumption. Data centres, at current scale, represent a fraction of those volumes, though that fraction is set to grow.

Given South Africa's coal-heavy grid, electricity consumed by data centres carries a larger embedded water cost than it would in a country running on renewables or nuclear. Every kilowatt-hour drawn from Eskom's coal fleet requires water at multiple points, at the mine, at the power station cooling system, and in maintaining coal stockpiles.

Schneider Electric has noted that liquid cooling has become a critical requirement for AI workloads, but that evaporative cooling is particularly challenging in water-scarce countries like South Africa. This tension (AI needs more cooling; South Africa has less water to spare) is real and will intensify as AI workloads increase.

As the grid transitions toward renewables, the indirect water footprint of data centre electricity will shrink considerably. Wind and solar have a fraction of the water intensity of coal. The energy transition and the water question are closely connected.

There is another side to this story that deserves equal weight: AI may be one of the most useful tools available to address South Africa's water crisis, not merely contribute to it.

Leak detection: Dr Giresse Komba at Tshwane University of Technology developed a real-time machine-learning system that detects and localises leaks in distribution networks with 96% accuracy, against 85% and 81% for the SVM and ANN baselines it was measured against. Given that South Africa loses close to 50% of treated water through infrastructure failures, roughly R10 billion annually, scalable deployment of such technology could recover more water than AI data centres consume, many times over.

Demand forecasting: AI tools can analyse historical usage, weather patterns, and population data to predict where and when water demand will spike, allowing utilities to pre-position supply rather than react to crises.

Drought monitoring: Satellite imagery analysed by machine learning can detect drought conditions, reservoir levels, and catchment health earlier and more accurately than traditional monitoring. In disaster-prone regions like KwaZulu-Natal, earlier warning translates directly into better resource allocation.

Infrastructure management: AI can flag procurement anomalies and infrastructure risks, areas where South Africa's water sector faces serious governance and maintenance challenges. Pattern recognition on sensor data can predict pipe failures before they become bursts.

Skills support: South Africa faces a significant shortage of technical expertise in the water sector. AI-powered dashboards, automated alerts, and decision-support tools can help less-experienced operators manage complex water systems more effectively.

These are not theoretical benefits. The tools exist. The gap is deployment, data availability, institutional capacity, and political will.

It is worth being explicit about the distinction between what the global AI water literature says and what is known specifically about South Africa.

Global estimates — that training a large language model uses hundreds of thousands of litres, or that AI could account for 4.2 to 6.6 billion cubic metres of water withdrawal by 2027 — are largely derived from hyperscale operations in the United States, China, and Europe. These figures involve significant modelling assumptions and, in many cases, extrapolations that have not been peer-reviewed.

The CMU Africa study provides the most rigorous Africa-specific analysis to date. Its estimates for South Africa are more modest than global averages. Writing a 10-page report was modelled at about 0.7 litres on Llama-3-70B and up to about 60 litres on GPT-4, depending on where the computation happens and how efficiently the data centre is cooled. These are estimates, not measurements from South African facilities specifically.

The transparency problem is real: most South African data centres do not publicly disclose water consumption data. Teraco is a notable exception. Without better disclosure, it is genuinely difficult to know what the current footprint is, let alone project it accurately.

The South African government has begun engaging with this question. On 31 May 2024 the Department of Communications and Digital Technologies gazetted the National Data and Cloud Policy, which sets a cloud-first direction and actively promotes data centre development. What it does not do is set a water efficiency standard. The policy is not yet legally binding, but it signals a direction of travel.

More targeted regulation around water disclosure and efficiency standards for data centres would help close the evidence gap. South Africa does not yet have an equivalent to the EU AI Act's transparency requirements on energy demand.

At current scale: not significantly, but the trajectory matters.

Today, South Africa's AI and data centre infrastructure consumes a real but relatively small share of national water resources, considerably less, per unit of computation, than the global average. The indirect water cost through coal-fired electricity is arguably more significant than direct cooling water, and this angle is frequently overlooked in public discussions.

The more pressing water story in South Africa remains infrastructure decay, governance failures, and an accelerating climate deficit, problems that AI tools are genuinely well-positioned to help address.

However, the picture will change. South Africa's data centre market was valued at about $2.16 billion in 2024, and is expected to reach $3.4 billion by 2030. AI workloads are more energy- and heat-intensive than general computing. Planned data centre expansions could more than double current power demand in the sector. Without mandatory water efficiency standards, improved cooling technology, and accelerating grid decarbonisation, the water footprint of South Africa's AI infrastructure will grow, in a country that has very little margin to spare.

The question is not whether AI affects water. It does. The question is whether South Africa can grow its digital economy in a way that is honest about that cost, invests in the technologies that reduce it, and deploys AI where it can genuinely help, starting with the 47.3% of treated water that currently disappears before it reaches a tap.

1. ENS Africa — AI and water systems: Conservation tool or resource drain? https://www.ensafrica.com/news/detail/10304/ai-and-water-systems-conservation-tool-or-res
2. ENS Africa — Cooling the hype: The real questions behind South Africa's data centre boom https://www.ensafrica.com/news/detail/12152/cooling-the-hype-the-real-questions-behind-so
3. Mordor Intelligence — South Africa Data Center Water Consumption Market (2025–2030) https://www.mordorintelligence.com/industry-reports/study-of-data-center-water-consumption-in-south-africa
4. ACM SIGCAS/SIGCHI 2025 — A Water Efficiency Dataset for African Data Centers (Carnegie Mellon University Africa) https://dl.acm.org/doi/10.1145/3715335.3735483
5. Explain.co.za — Can South Africa's AI data centres spare enough water? https://explain.co.za/2026/07/24/south-africa-wants-to-capitalise-on-the-ai-boom-can-it-spare-the-water-needed-to-run-thirsty-data-centres/
6. Daily Maverick — The hidden impacts of AI data centres on water, climate and future power costs https://www.dailymaverick.co.za/article/the-hidden-impacts-of-ai-data-centres-on-water-climate-and-future-power-costs/
7. Daily Maverick — Addressing the water crisis in South Africa https://www.dailymaverick.co.za/article/2024-04-11-addressing-the-water-crisis-in-south-africa-scarcity-climate-change-and-the-spread-of-disease/
8. Ai-Impact.co.za — How AI can help solve South Africa's water crisis https://ai-impact.co.za/how-ai-can-help-solve-south-africas-water-crisis/
9. Plumbing Africa — Smart leak detection breakthrough boosts water management https://plumbingafrica.co.za/smart-leak-detection-breakthrough-boosts-water-management-2/
10. Engineering News — Liquid-cooled AI data centres challenging in water-scarce South Africa https://www.engineeringnews.co.za/article/liquid-cooled-ai-data-centres-crucial-but-challenging-in-water-scarce-south-africa-2024-07-22
11. Centre for Environmental Rights — Water costs of coal-fired power https://cer.org.za/news/water-costs-impacts-of-coal-fired-power-grossly-underestimated-in-electricity-planning
12. Frontiers in Environmental Science — Tackling rural water scarcity in South Africa https://www.frontiersin.org/journals/environmental-science/articles/10.3389/fenvs.2025.1550738/full
13. DBSA — How can South Africa overcome its water challenges? https://www.dbsa.org/article/how-can-south-africa-overcome-its-water-challenges-and-avoid-african-water-crisis
14. Climate Scorecard — South Africa Country Water Resources https://www.climatescorecard.org/2026/06/south-africa-country-water-resources/
15. Heinrich Böll Stiftung — AI wants our water https://eu.boell.org/en/2025/11/10/ai-wants-our-water