Thirsty AI Challenge for a Sustainable Future

The carbon footprint of AI has been increasingly under scrutiny by the public. However, the no less significant water (withdrawal and consumption) footprint of AI has mostly been overlooked. For instance, training the GPT-3 language model in Microsoft’s cutting-edge U.S. data centers amounts to directly evaporating 700,000 liters of clean freshwater, but that information has been hidden from view. More importantly, the AI demand at the world level is predicted to require a withdrawal of 4.2–6.6 bm 3 in 2027, which is higher than the whole annual water withdrawal for four to six Denmark’s or half of the UK. This is alarming given that there is currently a global freshwater crisis. In a response to the world’s water challenges, AI can and also should assume social responsibility and set an example by mitigating its own water footprint. In this paper, we propose a principled approach to measure the water footprint of AI, and we raise the question of spatial-temporal diversities to AI’s runtime water efficiency. Lastly, we emphasize that just like carbon footprint, water footprint must be addressed holistically to realize truly sustainable AI.