AI's environmental numbers, and why everyone's are different
Per-prompt water estimates for AI range from 0.26 mL to about 519 mL. That's a factor of 2,000. Neither is wrong — they're measuring different things.
Google's 0.26 mL figure covers on-site cooling water for a median Gemini text prompt, including idle chips and data centre overhead. It excludes water consumed generating the electricity. The UC Riverside team's 519 mL figure is for a 100-word email and includes that indirect water. Google's own narrow, chip-only version of its methodology gives 0.12 mL. Sam Altman has cited ~0.39 mL for ChatGPT.
So before repeating any per-query number, check what's inside the boundary.
The aggregate picture
Data centres used about 415 TWh in 2024 — roughly 1.5% of global electricity. AI-specific facilities used about 155 TWh in 2025, or about 0.5% of world electricity. That's the total-stock view.
The growth view looks different. In 2025 data centre demand grew 17%, AI-focused facilities grew 50%, and global electricity demand grew 3%. Work that against global generation and AI accounted for something like 5–6% of all new electricity demand last year. (That last figure is my arithmetic from the IEA's growth rates and Ember's generation total, not a number the IEA publishes.)
Same year. Same data. Roughly a tenfold difference depending on which denominator you pick.
Water
Berkeley Lab estimated US data centres consumed 17.4 billion gallons directly in 2023, plus about 211 billion gallons indirectly through electricity generation. Together that's under 1% of US water consumption.
Two caveats. The indirect figure is contested — deriving it from USGS thermoelectric factors instead gives roughly half, partly because Berkeley Lab counts evaporation from hydroelectric reservoirs. And water doesn't move between basins, so a national percentage tells you very little about any specific place. Google's data centres use roughly a third of the municipal water in The Dalles, Oregon.
Concentration and efficiency both matter
Data centres are about 21% of Ireland's electricity and around 26% of Virginia's. Roughly two-thirds of data centres built since 2022 sit in water-stressed regions.
At the same time, the IEA finds energy per AI task falling by at least an order of magnitude annually. Google reports a 33× energy drop and 44× carbon drop per median prompt over twelve months. Predictions that AI would hit 20% of world energy by 2025 missed by a factor of about 40.
What nobody has
No company publishes query volumes, so per-prompt figures can't be multiplied up to totals. Best estimates put all text queries at about 2% of AI data centre electricity — the other 98% isn't publicly broken down anywhere.
Small share of the total. Large share of the growth. Very concentrated locally. Improving fast per unit of work. All four are true at once, and which one you lead with is a choice, not a fact.
Sources are all public. Go argue with the primary documents rather than with me.
Google methodology: https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
Li et al., Communications of the ACM 68:54–63 (2025): https://arxiv.org/pdf/2304.03271
IEA, Key Questions on Energy and AI: https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
Berkeley Lab, 2024 US Data Center Energy Usage Report: https://escholarship.org/uc/item/32d6m0d1
Our World in Data summary: https://ourworldindata.org/how-much-energy-do-data-centers-and-artificial-intelligence-use
#AI #Energy #DataCenters #Sustainability #ClimateData
Aaron Linson
@TechnicallBlind@caneandable.social
Accessibility Project Lead & CPACC certified by day, multi-laptop setup wrangler by night. 💻 Deep into AI coding, prompt engineering, and mastering the 5-string banjo. 🪕 Let's talk inclusive tech, digital automation, and good audio gear. 🎙️
caneandable.social