- Type
- Environmental impact metric
- Main uses
- Data-centre cooling, power generation
- Typical scale
- Millilitres per query; millions of litres per training run
- Key studies
- Li et al. 2023, Google 2025 inference measurement
- Emerged
- 2021 (data-centre water research)
- Related
- AI and energy, AI data centres in Malaysia
- Type
- Environmental impact metric
- Main uses
- Data-centre cooling, power generation
- Typical scale
- Millilitres per query; millions of litres per training run
- Key studies
- Li et al. 2023, Google 2025 inference measurement
- Emerged
- 2021 (data-centre water research)
- Related
- AI and energy, AI data centres in Malaysia
Where the water goes
Data centres use water in two ways. On-site (scope 1) cooling — evaporative cooling towers and air-handling systems dump heat by evaporating water, and direct-to-chip liquid cooling loops top up losses. Off-site (scope 2) water is embedded in the electricity consumed, since thermal and hydroelectric power generation evaporates water at the plant. For a given facility the split depends on climate and design: dry or air-cooled sites shift the burden to power generation; evaporative sites consume more locally but use less electricity.[1][2]
AI adds intensity on top of ordinary cloud computing because training runs pack thousands of accelerators into dense racks for months, and inference at population scale means fleets of GPUs answering queries around the clock.
How large is it?
Published estimates vary by more than an order of magnitude, and the differences are mostly about boundaries:
- Researchers estimated US data-centre cooling consumed about 66 billion litres of water in 2023, under 1% of national consumption.[3]
- A widely cited 2023 paper calculated that training GPT-3 in Microsoft's US data centres could consume roughly 5.4 million litres, including about 700,000 litres evaporated on site.[2]
- In 2025 Google published the first per-query measurement for a production chatbot, reporting a median-length Gemini text prompt consumed about 0.26 millilitres of water — roughly five drops — alongside 0.24 watt-hours of energy, far below the "500 ml per query" figure that circulated earlier from extrapolation rather than measurement.[4]
- A UN-linked assessment projects data centres could consume on the order of 9.3 trillion litres globally by 2030 under current growth trajectories.[5]
Reducing it
Levers include closed-loop and air-cooled designs (trading water for electricity), waste-heat reuse for district heating or aquaculture, siting facilities in cool or water-abundant locations, reclaimed and non-potable water sources, and operational tuning — raising the permitted coolant temperature even slightly cuts evaporation. Providers have begun publishing water and energy per token so customers can compare models on environmental cost as well as price.[4]
>Key Takeaways
- Most AI water use is cooling and electricity generation, not the model itself.
- Per-query consumption is small (about five drops for a median Gemini prompt); facility and regional totals are what strain supplies.
- Water impact is inherently local — the same model is sustainable in one siting and contentious in another.
- Efficiency levers are mostly engineering choices: cooling design, coolant temperature, reclaimed sources, heat reuse.
- Malaysia's Johor and Penang clusters make water a live planning constraint, not a distant sustainability question.
See Also
🇲🇾 Relevance to Malaysia: Malaysia's data-centre boom — over half of it in Johor, with further clusters in Cyberjaya and Penang — makes water a front-line policy issue. Johor's facilities draw on the Sungai Johor and other cross-border sources that also supply Singapore, so consumption is watched by both governments; Singapore's regulators have explicitly conditioned new data-centre capacity partly on water and energy efficiency.[6]
Malaysia's tropical heat raises cooling load year-round, and evaporative cooling is the default cost-optimal choice — exactly the design that consumes local water. Penang's periodic water rationing (most recently in 2024) showed how quickly data-centre demand can collide with domestic supply. The government's response has included guidance toward reclaimed water and higher PUE and WUE efficiency expectations in data-centre approvals, while the Johor state government and MIDA screen proposals on utility impact.[6][7]
For Malaysian AI firms, water intensity is becoming a commercial factor: regional buyers and investors increasingly request environmental disclosure alongside model performance, and Malaysia's adherence to ASEAN's AI governance conversation on responsible infrastructure means sustainability reporting is likely to be normalised for large facilities in the Malaysia Digital corridor.[7]
References
- ↑Shehabi, A. et al. 2024 United States Data Center Energy Usage Report. https://escholarship.org/uc/item/32d6m0d1
- ↑Li, P. et al. Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models. https://arxiv.org/abs/2304.03271
- ↑Knowable Magazine. How much of a problem is AI's water use? https://knowablemagazine.org/content/article/technology/2026/how-much-water-do-ai-data-centers-use
- ↑Google Cloud. Measuring the environmental impact of AI inference. https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
- ↑Earth.org. Data centers could consume 9.3 trillion liters of water by 2030 (UN report). https://earth.org/9-3-trillion-liters-of-water-un-report-exposes-unfathomable-footprint-of-data-centers-as-ai-booms/
- ↑MIDA. Data centre investments in Malaysia. https://www.mida.gov.my/
- ↑MCMC / Malaysia Digital. https://www.mcmc.gov.my/