Calculate the water consumption of AI model training and inference. Estimate cooling water usage for data centers running large language models and ML workloads.
Artificial intelligence is transforming industries worldwide, but its environmental cost extends far beyond electricity. Training and running large AI models requires enormous amounts of water for cooling the data centers that house the powerful GPUs and TPUs needed for computation. A single training run of a large language model like GPT-4 can consume millions of liters of water, equivalent to the annual water usage of hundreds of households.
Data centers use water in two primary ways: direct evaporative cooling, where water is evaporated to remove heat from server rooms, and indirect cooling through electricity generation at power plants. The water footprint varies dramatically depending on the data center's location, cooling technology, local climate, and the energy grid's water intensity. Facilities in arid regions or those relying on water-cooled power plants have significantly higher water footprints.
This calculator helps you estimate the water consumption associated with AI workloads, from training large models to daily inference queries. By understanding the hidden water cost of AI, organizations and individuals can make more informed decisions about sustainable AI deployment, model selection, and data center siting.
Understanding AI's water footprint is essential as the world deploys more AI systems while facing growing water scarcity. This calculator helps researchers, businesses, and policymakers quantify the hidden water cost of AI and make more sustainable technology decisions. Keep these notes focused on your operational context. Tie the context to the calculator’s intended domain. Use this clarification to avoid ambiguous interpretation.
Total Water = GPU_Hours × Power_per_GPU × PUE × Water_Usage_Effectiveness (WUE) + Indirect_Water_from_Electricity. WUE is measured in liters per kWh and varies by cooling method and climate. Indirect water accounts for water used in electricity generation (coal: ~1.9 L/kWh, natural gas: ~0.7 L/kWh, nuclear: ~2.3 L/kWh).
Result: 185,000 liters total water
Training a medium model for 10,000 A100 GPU-hours at 400W per GPU with a PUE of 1.2 and evaporative cooling in a temperate climate uses approximately 48,000 kWh. At a WUE of 1.8 L/kWh for direct cooling plus 0.7 L/kWh for grid electricity, the total water footprint is about 185,000 liters.
While much attention has been paid to the electricity consumption of AI, water usage is an equally critical but often overlooked environmental impact. Data centers are among the largest industrial consumers of water in many regions. Google's data centers alone consumed approximately 5.6 billion gallons of water in 2022, a 20% increase from the previous year, driven largely by AI workloads.
The relationship between AI computation and water consumption is complex. Every kilowatt-hour of electricity used by GPUs generates heat that must be removed. Evaporative cooling towers, the most common method, work by evaporating water to absorb heat — similar to how sweat cools your body. This process is energy-efficient but water-intensive.
Different AI tasks have vastly different water footprints. Training a large language model from scratch is the most water-intensive, potentially consuming enough water to fill an Olympic swimming pool. Fine-tuning uses 10-100x less water, while individual inference queries each use only milliliters. However, inference at scale — billions of queries per day — can rival training-level consumption over time.
Image generation models like DALL-E or Stable Diffusion have moderate water footprints per query, roughly 2-5x more than text queries due to higher computational requirements. Video generation and real-time AI applications push consumption even higher.
The AI industry is beginning to address its water impact through several strategies. Liquid cooling systems that circulate coolant directly over processors can reduce water consumption by 20-40%. Some facilities are experimenting with immersion cooling, submerging entire servers in non-conductive fluid. Using treated wastewater or seawater for cooling keeps potable water available for communities. Companies can also reduce water usage by optimizing model architectures, using more efficient hardware, and scheduling intensive workloads during cooler periods.
A single ChatGPT query uses approximately 3-10 mL of water when accounting for both direct data center cooling and indirect water from electricity generation. This may seem small, but with billions of queries per day, it adds up to millions of liters daily.
Data centers generate enormous amounts of heat from thousands of servers running 24/7. Evaporative cooling is one of the most energy-efficient ways to remove this heat, but it consumes large volumes of water. Some data centers use over 1 million gallons of water per day.
Power Usage Effectiveness (PUE) measures total data center energy divided by IT equipment energy. A PUE of 1.2 means 20% of energy goes to cooling and overhead. Higher PUE means more energy (and thus water) is needed for cooling.
Large language models like GPT-4 and Gemini Ultra have the highest water footprints due to their massive training requirements, often consuming millions of liters during training. Smaller models like GPT-3.5 or open-source alternatives use significantly less.
Yes, through air-cooled systems, liquid cooling directly on chips, using recycled or non-potable water, locating in cooler climates, and improving PUE. Some companies like Microsoft are committing to being water-positive by 2030.
Data centers in hot, arid climates need more cooling and consume more water. A facility in Arizona might use 3-5x more water than one in a cool Scandinavian location where outside air can provide most of the cooling needs.