When it comes to total water use, AI data centers are a drop in the bucket

Editor's take

Large language model training and inference, while demanding, represent a relatively small fraction of the global water footprint attributed to data centers. The analysis highlights that while individual AI workloads are water-intensive, their current scale in aggregate is dwarfed by the water needs of broader data center operations, including cooling and general IT infrastructure.

This finding is significant because it shifts the focus of water conservation discussions within the AI industry. Instead of solely targeting the energy efficiency of specific LLMs like OpenAI's GPT-4 or Google's PaLM 2, the emphasis must broaden to the entire data center ecosystem. This impacts not only AI developers but also cloud providers like AWS and Microsoft Azure, as well as the communities where these facilities are located, which are often already facing water stress.

Future attention should be directed towards the water efficiency of cooling technologies used across all data center functions, not just AI-specific hardware. The development and adoption of closed-loop cooling systems and drought-resistant cooling methods will be critical. Furthermore, understanding the geographical distribution of AI workloads relative to water availability will become increasingly important as AI adoption scales.