700K
Liters to train GPT-3
1.7T
Liters consumed by data centers in 2022
+30%
Annual growth in AI water consumption
500mL
Water used per 20–50 ChatGPT prompts
Computers get very hot when they run AI. To cool them down, data centers pump enormous amounts of cold water past the servers. When you ask ChatGPT 20 questions, it drinks roughly a bottle of water. When companies train a big AI model from scratch, it can drink as much water as 17,000 Americans use in a day.
AI data centers use water in two ways: direct cooling (water flows through chillers to cool servers) and indirect evaporative cooling (water evaporates into the air to cool the building). Li et al. (2023, Nature) calculated that training GPT-3 consumed ~700,000 liters. Inference (running the model) consumes less per query but adds up at scale: 500mL per 20-50 ChatGPT prompts. Microsoft's global water consumption jumped 34% from 2021 to 2022, largely due to AI expansion.
Li, R. et al. (2023). "Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models." Nature. GPT-3 training: ~700,000L at Microsoft's Iowa data center (Water Usage Effectiveness [WUE] ~1.8 L/kWh × ~380 MWh compute). ChatGPT inference: ~500mL per 20-50 prompts based on WUE and estimated query compute. Microsoft 2022 ESG Report: water consumption increased from 4.196GL (2021) to 6.4GL (2022). AWS and Google publish water efficiency metrics but not total consumption. The key metric is WUE (Water Usage Effectiveness) — targets are sub-1.0 L/kWh for new facilities.
Primary: Li, R. et al. arXiv:2304.03271 (2023). Microsoft 2022 Environmental Sustainability Report (PDF). Google Environmental Report 2023. AWS Sustainability page. Metrics: WUE (Water Usage Effectiveness, L/kWh), PUE (Power Usage Effectiveness). Data center WUE benchmark: ASHRAE TC9.9, Green Grid standards. Global data center water: IEA "Data Centres and Data Transmission Networks" 2022. Compute estimates for GPT-3 training: Narayanan et al. (2021), ~3.14×10²³ FLOP. Convert FLOP to kWh using A100 efficiency: ~0.4 kWh/10¹¹ FLOP.
The Growing Thirst
AI data center water consumption is growing faster than the industry's efficiency improvements. And the data centers are often placed in water-stressed regions for cheap land and power.
Global Data Center Water Consumption (Trillion Liters)
IEA estimates + AI growth projections.
Microsoft Water Consumption by Year (Billion Liters)
From Microsoft ESG reports. 2022 jump driven by AI expansion.
Drought Paradox
Microsoft's primary AI training cluster in Iowa sits in a region experiencing increasing drought stress. Google's data centers in drought-prone Arizona and Nevada withdrew millions of gallons from stressed municipal water systems in 2022. The AI boom is competing directly with local water supply needs.
Sources
• Li, R. et al. (2023). "Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models." arXiv:2304.03271.
• Microsoft (2022). Environmental Sustainability Report. microsoft.com/en-us/corporate-responsibility/sustainability.
• IEA (2022). "Data Centres and Data Transmission Networks." iea.org.
• Google (2023). Environmental Report. sustainability.google.com.
• AP. "AI water use: How ChatGPT consumes water." July 2023.