Making AI less ‘thirsty’

A new study of the University of Colorado Riverside and the University of Texas Arlington uncovers the ‘secret water footprint of AI models’. As an example ChatGPT ‘drinks’ a 500ml bottle of fresh water for every simple conversation with about 20 to 50 questions. Considering the high popularity of OpenAI’s ChatGPT with billions of users, the daily consumption of fresh water is enormous. The paper shows how the ‘when’ and ‘where’ of the training have significant impact on the water footprint. Training GPT 3 in Texas Microsoft’s state of the art US data centre is estimated to consume 700,000 litres of freshwater. While the numbers may be probably 3 times higher in Microsoft’s Asian data centers. Following the example of carbon footprint reduction, a migration to public clouds could help to significantly reduce the water consumption.

The motivation behind the research is to bring this highly important subject ‘on the radar’ of the public as well as AI developer. But also, to request more transparency from the industry. Freshwater scarcity is among the most pressing global challenges. A ‘landmark report’ on water economics predicts that the demand of water could exceed the available freshwater supply by 40% by the end of the decade. Both, the carbon as well as the water footprint, require a holistic approach to enable truly sustainable AI.

Microsoft declined to comment on the study. OpenAI did not respond a request by Euronews.