aceMaking AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models
The growing carbon footprint of artificial intelligence (AI) has been undergoing public scrutiny. Nonetheless, the equally important water (withdrawal and consumption) footprint of AI has largely remained under the radar. For example, training the GPT-3 language model in Microsoft's state-of-the-art U.S. data centers can directly evaporate 700,000 liters of clean freshwater, but such information has been kept a secret. More critically, the global AI demand is projected to account for 4.2-6.6 billion cubic meters…
Training and running large AI models consumes millions of liters of mostly drinkable freshwater that companies do not disclose: an estimated 5.4 million liters to train GPT-3 alone, with global AI projected to withdraw 4.2 to 6.6 billion cubic meters of water in 2027.
All figures are modeled from published averages (Microsoft's annual water and power efficiency per site, national power-plant water factors, and an assumed 0.004 kWh per response), not metered measurements. GPT-3's training location is not public, and chip and server manufacturing water is excluded for lack of data. Ren, the senior author, submitted the paper to Truvace.
Evidence
- Peer-reviewedCommunications of the ACM2025-06-27
How should this claim be treated?
Truvace Impact Record TRV-2026-2359, v2: “Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models.” Truvace, 2026-11-04. /record/TRV-2026-2359 (accessed at citation time). sha256 ed61b224cc6d9c26…
Calibration history
Every change to this record since certification, in the open.
Rewritten from the paper's full text (previously generated from the abstract only): added the per-site training estimates, per-response water use, method, caveats and the authors' recommendations.
Certified into the record
How to verify without trusting this page
Fetch the canonical text of any version from /api/record/TRV-2026-2359 and hash it yourself — for example shasum -a 256 on the saved canonical field. The result must equal content_hash, and each version’s text ends with prev:followed by the prior version’s hash (version 1 chains to 64 zeros). If a single character of any version had been altered since certification, the chain would not reproduce.