Water Consumption of African Data Centers in the Age of AI

AI chatbots, like ChatGPT, require a significant amount of computing power and energy. Each time you prompt a chatbot with a question, it will send a request to a data center where user input is processed and a generated response is sent back. On the surface, the process seems simple enough, but it comes with a hefty cost and strain on the world’s most valuable natural resource—water.

With the rapid acceleration of AI usage, there is an increased demand for data centers, which use water for cooling servers that generate substantial heat. For data centers to support the demand of AI, they could require millions of gallons of water a day, causing additional strain for communities where this resource is already in short supply. Even in developed countries, water scarcity is not uncommon. But in Africa, that scarcity is especially severe, as many countries grapple with extended periods of drought.

Despite the rapid expansion of data centers in Africa and the region's known challenges with water scarcity, there is little to no research about water efficiency for African data centers. In a new study, researchers at Carnegie Mellon University Africa looked at national weather and electricity generation data to estimate the water usage efficiency (WUE) for data centers across 41 African countries in five distinct climate regions. Using these water efficiency estimates, the researchers also estimated the water consumption of two large language models, Llama-3-70B and GPT-4, in 11 African countries, comparing the numbers with estimates in the U.S. and globally.

"Similar studies have been carried out in the U.S. and other Western countries to try to estimate the water consumption of these data centers, especially for generative AI use cases," said Noah Shumba, Upanzi Network Research Associate. "The goal of our research was to try to estimate how much water would be consumed if these data centers were to be placed in Africa."

Their paper, "A Water Efficiency Dataset for African Data Centers," was presented at the 2025 ACM COMPASS conference, which focuses on the application of computing for social impact, sustainability, and global development.

Researchers modeled WUE and AI water consumption both onsite, which refers to water directly used to cool down a data center, and offsite, which includes indirect water consumption by the generation of electricity used to supply each unit of data center energy. Africa's dry climate results in more onsite water consumption, but the significantly lower offsite water usage makes most countries less water-consuming overall.

When tasked with writing a 10-page report, researchers found that Llama-3-70B consumed 0.7 liters of water, or about three cups. This estimate is well aligned with Mistral AI's recent disclosure about the water consumption by its Large 2 123B model, which required 50 milliliters for each page of content. GPT-4 on the other hand was estimated to use up to 60 liters, roughly the amount to run a standard washing machine cycle. Additionally, it was determined that eight out of the 11 selected African countries used less water than the global average when performing the same task, and Morocco and South Africa used even less water than the U.S. average.

"Another surprising finding of the paper is that a lot of the water costs actually come from leakage in water transmission pipes," said Giulia Fanti, CMU associate professor of electrical and computer engineering and one of the study's authors. "Leakage in many African countries is estimated to be much higher than in the U.S., for instance. So if we want to invest in better AI infrastructure, sometimes that can be helped by just investing in better classical infrastructure."

Researchers also found a correlation between climate conditions and water consumption in Africa. Five different climate regions were identified: rainforest, savanna, desert, steppe, and Mediterranean regions. Countries that fell under the rainforest region (i.e., Republic of Congo, Rwanda, and Gabon) and the steppe climate (i.e., Ethiopia and Namibia) had higher or the same water consumption as the global average. This can potentially be attributed to the hot and humid conditions of the rainforest climate and dry conditions in the steppe regions, researchers report.

Another potentially compounding factor could be the high offsite water consumption of these countries, which can be attributed to water-intensive energy fuels like hydroelectric or thermo-electric power.

"We need to look at both the energy production side and the climate," said Opelo Tshekiso, Upanzi Network Research Associate. "When you combine the two and they fit the same narrative where both sides of the scale are putting pressure on the water supply, then that’s where you find countries reporting higher water consumption."

Moving forward, the data presented in this paper can encourage organizations looking to build data centers to consider both the centers’ placement and energy sources used to power them, thereby promoting sustainable water usage.

"I think what makes this area of study challenging is that it is rarely enough to think about just one cost or value metric," said Fanti. "There are so many trade-offs that impact the placement of data centers, including physical constraints like energy and water usage, and conceptual constraints like data locality or privacy requirements. I think it would be very interesting to see new frameworks and measurement studies evaluating the costs of data centers in different regions globally, while simultaneously taking into account the many ways in which a data center can impact a community."

Pengfei Li and Shaolei Ren from the University of California, Riverside also contributed to this work as co-authors of the study.