Explanation
Yes, at the scale of large AI workloads, water use can be substantial. The footprint comes from the infrastructure running AI, including water consumed in cooling and electricity generation. There is no single water cost that applies to every AI system or request.
Researchers estimated that training GPT-3 in the US data-center conditions they examined could directly evaporate about 700,000 liters of freshwater. This was a modeled estimate for a particular training scenario, not a publicly metered total for all AI or a measurement of current chatbot requests.
Cooling design makes a major difference. Microsoft has described an AI-oriented design that recirculates cooling water without ongoing evaporation. That does not eliminate every water use associated with the facility or its electricity. Large-scale consumption is real, but universal claims about liters per prompt oversimplify it.
Key takeaway
AI can consume substantial water at scale, but its footprint depends on how and where it runs.