Water consumption: the big unknown with increasing AI use

The water consumption of artificial intelligence (AI) is a significant and growing problem: It has a major impact on the environment, particularly due to the immense cooling requirements of the data centers already needed to train AI models. Their energy consumption has long been at the center of a larger debate. In contrast, the sometimes more problematic water consumption of AI has so far played a less visible role.

The authors of a study by the German Informatics Society (GI) are therefore now taking a systematic look at the water consumption of AI systems throughout their entire life cycle for the first time. This ranges from chip production and data center operation to the disposal of the hardware. In the research report, they recommend technical, regulatory and social measures to reduce water consumption. There is an urgent need to "develop sustainable strategies for the future".

Estimates

An initial estimate of water demand dates from 2023, according to which global water use by AI applications could increase to between 4.2 and 6.6 billion cubic meters (m³) by 2027. This corresponds to more than four to six times the annual water consumption of Denmark. Using the GPT-3 language model with 175 billion parameters as an example, this study shows that its training in Microsoft's state-of-the-art US data centers is likely to have required around 5.4 million liters.

According to other studies, creating a ten-page report with Meta's Llama-3-70B consumes around 0.7 liters of water, while GPT-4 could consume up to 60 liters. Each email formulated by the AI model or 20 to 50 questions to an AI chatbot such as ChatGPT therefore requires around half a liter of water.

However, according to the GI study, these estimates are subject to considerable uncertainty: on the one hand, the water consumption of data centers varies depending on the cooling system used and the location. Secondly, due to a lack of data availability, the projections do not take into account the water consumption of the supply chain, particularly in chip production.

Standardized measurement methods for all AI-related water consumption are lacking. Many tech companies do not publish detailed consumption data, the scientists complain. Modeling is therefore often based on estimates.

"Water stress" already high worldwide

Particularly challenging: around half of the world's population is already affected by "water stress". This means that the available freshwater resources can no longer meet demand. By 2030, Spain, Italy, Belgium and Greece in particular, but also parts of Germany such as Brandenburg and regions in central Germany, are expected to be affected.

The construction of new data centers is exacerbating the problem, the authors explain. Many of these facilities are being built in areas that already suffer from or are threatened by water shortages. This could lead to conflicts of use between the increasing water requirements of digital infrastructure and other social and ecological demands.

Energy-efficient algorithms, adaptive training processes, specialized hardware with lower power requirements and the use of smaller, task-specific AI models could reduce water demand, the authors write. Other fields of action: resource-saving data center infrastructure such as water-saving cooling technologies, the establishment of a consistent circular economy in hardware production, the selection of data center locations taking into account local water availability and the establishment of binding transparency standards and new evaluation metrics.

The study concludes: "If AI is to make a contribution to overcoming global challenges, it must itself be designed in a sustainable and responsible way." There is an urgent need to continue the debate on the ecological transformation of digital infrastructure on a sound basis.

(mki)