EVERY CHATGPT QUERY COSTS MORE THAN YOU THINK: QUANTIFYING THE HIDDEN WATER FOOTPRINT OF AI
Keywords:
: Artificial Intelligence; Water Footprint; Data Centers; ChatGPT; Generative AI; Cooling Technologies; Geographic Load Balancing; Scope-1 Water Usage; Scope-2 Water Usage; Carbon-Water Trade-off; Sustainability Metrics.Abstract
Artificial intelligence (AI) has environmental impacts that are often
overlooked. This paper examines the water footprint of generative AI models
by analyzing water consumption associated with data center cooling and
electricity generation. It is estimated that approximately 5.4 million liters of
water were consumed during the training of GPT-3. In addition, a single
ChatGPT query requires about 500 ml of water for every 10–50 responses.
Water consumption varies significantly depending on the training location:
approximately 15.29 million liters would be required in Washington compared
with 3.68 million liters in Virginia. Temporal analysis showed no significant
correlation between carbon and water efficiency (Pearson r = 0.06). Potential
technical solutions include geographic load balancing, advanced cooling
