Calculate illustrative water and energy estimates for AI model queries. Understand the uncertainties and context behind these numbers.
Last updated: March 2026
Enter a regional CO₂ intensity or leave default (0.42 kg/kWh).
AI water footprint refers to the freshwater consumed by computing infrastructure over its lifecycle. Some usage is directly for cooling, but the contribution per query varies greatly by cooling method, PUE, workload, and location. Per-query numbers are illustrative ranges, not exact values.
The water footprint also includes water embedded in electricity generation, hardware manufacturing, and facility operations. Because conditions vary widely between providers and regions, use the model comparison ranges and CO₂ input to explore different scenarios rather than treating outputs as precise measurements.
*CO₂ intensity varies widely by region and grid mix; use the input above to reflect local conditions. Default shown is 0.42 kg/kWh.
Note: These estimates are based on publicly available research and industry reports. Actual consumption varies based on data center efficiency (PUE), cooling technology, geographic location, and electricity grid carbon intensity.
Calculate annual water footprint for a moderate ChatGPT user:
This is an illustrative result using a rounded per-query average; real values vary by provider and location.
AI models require massive computational power, which generates heat. Data centers use water-based evaporative cooling because it's more efficient than air cooling for high-density GPU clusters. The water evaporates to dissipate heat, making it a consumptive use.
Yes, most of it. Evaporative cooling systems release water vapor to the atmosphere. While this eventually returns as precipitation through the water cycle, it's often lost to the local watershed and may fall hundreds of miles away.
Large language models like GPT-4 and image generation models like DALL-E consume the most water per query due to their computational intensity. Training these models initially consumes even more—potentially millions of liters for a single training run.
Yes, through multiple strategies: building data centers in cooler climates (reducing cooling needs), using closed-loop cooling systems (recycling water), improving hardware efficiency (reducing heat generation), and locating in areas with sustainable water supplies.
Data centers overall account for ~1% of global electricity use. For comparison, agriculture uses ~70% of global freshwater. However, AI's water footprint is growing rapidly and is concentrated in specific regions, making local impact significant.
Mixed. Hardware efficiency is improving (newer GPUs perform more operations per watt), but AI usage is exploding. The total water footprint of AI is growing despite per-query efficiency gains. Industry is investing in better cooling technology.
Using renewable electricity reduces CO₂ emissions but doesn't necessarily reduce water consumption. Wind and solar don't require water for generation, but cooling the AI hardware still does. Some renewable sources like hydro can have their own water impacts.
Yes: use AI selectively rather than for every query, choose smaller models when possible (GPT-3.5 vs GPT-4), cache and reuse responses, and support companies committed to water efficiency and transparent environmental reporting.
| Model | Water (mL) | Energy (kWh) |
|---|---|---|
| GPT-4 query | 200–800 | 0.005–0.05 |
| ChatGPT (GPT-3.5) | 5–50 | 0.0001–0.005 |
| Image generation (DALL-E) | 400–1,200 | 0.01–0.06 |
| Google search | 1–15 | 0.00005–0.001 |
Data sources: University of California research (2023), Microsoft sustainability reports, and industry benchmarks.
Related Tools
Compare bag environmental impact.
Compare tree types' impact.
Calculate coffee carbon impact.
Calculate crypto energy use.
Compare hand drying methods.
Calculate plastic usage impact.