AI Needs Water-Dependent Infrastructure
As we expand access to AI, how do we protect access to water?
Data Centre Electricity Is Projected to Double by 2030
Generating electricity can consume substantial water, and every facility draws on a specific grid and catchment.
2.3×
the 2030 projection against the 2024 estimate
88%
of surveyed organisations used AI in at least one business function in 2025, up from 78% in 2024
5
technology companies now spend more capital than the world invests in oil and gas production
US Data Centres, 2023: Most Water Sits Behind the Grid
To see the water, look at the cooling system and at the electricity behind it.
Withdrawal
Water taken from a surface or groundwater source.
Consumption
The portion not returned to the immediate water environment.
AI's Water Footprint Is a Range, Not a Number
The answer depends on the model, the workload, where it runs, how it is cooled and what electricity powers it.
The Bottle of Water
A 500 ml bottle for roughly 10 to 50 medium-length GPT-3 responses, depending on when and where it is deployed.
50 mla response, roughly: the bottle lasts 10
Where It Runs
The weather outside changes the water cooling takes; the grid's fuel mix changes the water behind the electricity.
When It Runs
Hot hours raise evaporative cooling, so work that can wait, such as training, can move to cooler ones. In a simulation, a scheduler that moved it across hours and sites used about 25% less water than other schedulers.
How It Is Cooled
Cooling towers evaporate water. Dry coolers evaporate none but use more electricity, and so can take more water behind the grid. Microsoft says its newest design, which evaporates none, avoids the need for more than 125 million litres a year per data centre.
Which Water It Uses
Reclaimed or recycled water can stand in for drinking water in cooling, though it evaporates all the same. Microsoft says it has expanded its use in Texas, Washington, California and Singapore.
AI Starts in a Chip Fab, and Fabs Depend on Ultrapure Water
Before a model answers a single question it needs chips, and chips are made where contamination would be unacceptable.
300 mm
- About 3,800 Lof ultrapure water per wafer at 2 nm, modelled
- About 100rinses for each chip during manufacturing
- 57dies of the largest size on this wafer
- 28Blackwell GPUs per wafer at most, two dies each, before any fail
- About 90 L or moreof ultrapure water per Blackwell GPU, an estimate
- 3.6 millionBlackwell dies, two to a GPU, shipped to the four largest US cloud providers by March 2025, NVIDIA says
Each dot: one tenth of the 28 nm level.
Electricity: 3.46×
per wafer from 28 to 2 nm, in imec's 2020 model.
Blackwell
NVIDIA's Blackwell GPUs are made on a custom TSMC 4NP process.
In February 2021, after months of scant rainfall in Taiwan, TSMC began buying trucked-in water for some of its fabs.
Treatment, purification and reuse decide where chip capacity can grow.
Lower Intensity Can Still Mean Higher Totals
Lower water per unit of computing can coexist with higher total consumption if activity grows faster.
Total demand is higher than where it started.
The Identity
Total water demand is water per unit of computing times the volume of computing. Efficiency shrinks the first; growth can outpace it.
So ask for
Absolute Volumes
not only intensity
Peak Demand
not only annual totals
Source and Season
not only the building
Climate Emissions Are Not Local Freshwater Demand
The water question is where demand is added, when it peaks and whether communities keep a reliable supply.
Global: climate
Emissions and Warming
Greenhouse gases alter Earth's energy balance.
More than 90%
of the excess heat is absorbed by the oceans
Local: catchment
Demand and Local Pressure
Withdrawal and consumption affect a source shared with people and ecosystems.
Thermal discharge is a third, site-specific question.
AI Is Becoming Infrastructure. It Draws on the Same Water.
AI is not one more customer on the network.
The Questions Are Bigger Than Litres
Hot weather can raise evaporative cooling needs precisely when water supplies are under pressure.
Usual level
Shared source
Aquifer or reservoir
- Homes Daily supply
- Farms Irrigation
- Businesses Local economy
- AI Workloads New, growing demand
More and more of each runs on AI
Dry-Season Capacity?
Can service stay reliable in the hottest weeks?
Who Pays?
Who funds the added capacity?
Who Carries the Risk?
Who faces restrictions if supply falls short?
Water Is Economic Infrastructure
We can change how efficiently our systems use water, but we cannot make their need for reliable water disappear.
- FoodTiming of supply, not only volume
- CitiesReliability and safety, not litres
- ManufacturingDowntime costs more than the water
- EnergyGeneration can use substantial water
- ComputingCooling, and water behind the grid
Resource base: catchments, aquifers, soils and ecosystems
A low water bill can coexist with a high cost of failure.
The Service Gap
These are service gaps with consequences for people's health, time and opportunity.
Drinking water
2.1 bn
people lack safely managed drinking water
1 in 4 people
Safely managed: on premises, available when needed, free from contamination.
Sanitation
3.4 bn
people lack safely managed sanitation
2 in 5 people
A separate service, measured separately.
2024 estimates.
Each dot: 1 in 20 people worldwide.
Exposure Is Economically Material
Water risk belongs in decisions about where and how we invest, including where AI infrastructure is built.
31%
of projected global GDP, $70 trillion, exposed to high water stress by 2050
A projection: exposure, not loss.
Climate Resilience Reaches People Through Water
We track warming through greenhouse gases; many of its effects reach us through water.
About 60%
of the adaptation responses documented since 2014 address water-related hazards
Drought
Storage, efficiency and reuse that hold through dry seasons.
Flood
Retention landscapes and drainage that protect communities.
Reliable Services
Safe water and sanitation that keep working under stress.
Design Can Change the Outcome
There are practical alternatives, and each one needs evaluating for local conditions.
Reuse
Treat and recirculate cooling water.
Appropriate Cooling
Avoid evaporation where local conditions allow.
Repair
Find and fix leaks, with AI to detect them.
Joint Planning
Plan water and electricity together.
Quincy, Washington
A city reuse utility treats and recirculates data-centre cooling water, reducing reliance on local groundwater, according to Microsoft.
Treated and recirculated
- Data Centre
- Treatment
- Cooling
Community potable network: a separate system
In The Water Foundation's own network, Rotreat builds modular membrane plants that recover water from industrial wastewater.
AI Helps When Insight Becomes Action
A model can flag a leak; an operator decides, a crew repairs, and only a measured outcome proves the value.
AI Does One Step in a Chain: Measure, Detect or Forecast, Decide, Act, Verify.
Test for a useful deployment
- Did it find actionable problems?
- Were they fixed?
- Did verified losses or interruptions fall?
- What computing and upkeep did it need?
In The Water Foundation's own network, Aqua Predict forecasts a location's long-term water availability from satellite climate data.
Less Loss. Better Service. Lower Avoidable Cost.
Sometimes the most valuable water investment is improving the system we already have.
- Source Extracted
- Treatment Plant Treated
- Pumping Pumped
- Leak Lost before delivery
- Customer Service delivered
- Delivered
- 79 L
- Lost
- 21 L
In developing countries, about 21 of every 100 litres of treated water leak from the pipes before reaching anyone.
Delivery
More treated water reaches the customer it was paid for.
Energy
Less pumping and treatment spent on water that is lost.
Reliability
Steadier service, and sometimes new supply can wait.
Invest in What Makes Growth Possible
How do we design energy, water and AI together, who benefits, who carries the risk, and how do we finance the result?
The Money Does Not yet Match the Need
The figures cover different scopes and are never added together, but they point to the same challenge: capable institutions, better projects and capital that can support them.
Current spending
About $164.6 bn
a year of water spending in developing countries, estimated for 2017
Public roughly 91%, private less than 2%
Investment need
Up to $7 tn
to mobilise for water infrastructure by 2030, globally: an upper estimate, not an annual sum
Project cash flow
Who Pays?
Over the project's lifetime: tariffs, public payments or contracts. Need alone does not make a project bankable.
Our approach
The Water Foundation is stepping up to fill this gap at a systems level, using systems level and first principles thinking.
- We bridge private and public sectors through blended finance instruments.
- From catalytic seed to patient private equity, we work across the entire capital spectrum.
- We accelerate the transition from water innovation to global resilience by turning bold ideas into bankable infrastructure and lasting equity.
Who Decides, Who Benefits, Who Carries the Risk?
The benefits of AI can be spread across the world, while its water demand lands on one community and one source.
Equity
AI's benefits are global; its water demand is local.
Transparency
Absolute use, peak demand, source and season, per facility.
Governance
Allocation and drought rules agreed before construction.
Climate
Emissions are a separate, global question. Plan both together.
What a city can ask of a new data centre
- 01A water and energy balance
- 02Peak-condition scenarios
- 03Suitable sources
- 04A credible operating plan
- 05Transparent reporting
As We Expand Access to AI, How Do We Protect Access to Water?
Plan together. Finance reliability. Assign accountability.
Ways to Work With Us
Join a network of partners across the private and the public sector.
Sources
- IEA, 2025: Energy and AI: Energy demand from AI
- IEA, 2026: Key Questions on Energy and AI: Executive summary
- Stanford HAI, 2026: AI Index Report 2026
- LBNL, 2024: 2024 United States Data Center Energy Usage Report
- Li et al., 2025: Making AI Less “Thirsty”
- Islam et al., 2018: Exploiting Spatio-Temporal Diversity for Water Saving in Geo-Distributed Data Centers
- Microsoft, 2024: Sustainable by design: Next-generation datacenters consume zero water for cooling
- imec, 2026: Reducing the Environmental Impact of Wet Chemical Processes for Advanced Semiconductor Manufacturing
- imec, 2020: The environmental footprint of logic CMOS technologies
- NVIDIA: Blackwell architecture
- NVIDIA, 2025: GTC 2025 keynote deck
- CORDIS, 2021: The ultra-clean water making a splash in the semiconductor industry
- Focus Taiwan, 2021: TSMC buying trucked-in water in preparation for supply shortages
- CRS, 2025: Section 301 and China: Mature-Node Semiconductors
- TSMC: 2nm Technology
- ASML: TWINSCAN NXT:2050i
- NOAA NCEI, 2021: Monthly Ocean Heat Content and Temperature Anomalies Released
- Barnthouse and Coutant, 2022: Modernizing thermal discharge assessments for the 21st century
- FAO: Agricultural water management: water scarcity
- WRI, 2023: 25 Countries, Housing One-Quarter of the Population, Face Extremely High Water Stress
- Global Commission on the Economics of Water, 2024: The Economics of Water, chapter 3: Towards a new economics of water
- OECD, 2022: Financing a Water Secure Future, chapter 2: The Challenge of Financing Water-related Investments
- WHO/UNICEF JMP, 2025: Progress on household drinking water, sanitation and hygiene 2000-2024: special focus on inequalities
- WHO/UNICEF JMP, 2025: Progress on household drinking water, sanitation and hygiene 2000-2024, on UNICEF Data
- WHO, 2025: 1 in 4 people globally still lack access to safe drinking water
- IPCC, 2022: AR6 Working Group II, Chapter 4: Water
- Microsoft, 2025: Our 2025 Environmental Sustainability Report
- Microsoft, 2026: Building Community-First AI Infrastructure
- World Bank, 2024: Closing the $7 Trillion Gap: Three Lessons on Financing Water Investments from World Water Week
- Rotreat: rotreat.at
- Shu et al., 2024: Long-term water demand forecasting using artificial intelligence models in the Tuojiang River basin, China
- Aqua Predict: aqua-predict.com
- World Bank, 2006: The Challenge of Reducing Non-Revenue Water (NRW) in Developing Countries
- World Bank, 2018: What do private companies look for in a performance-based non-revenue water project?
- World Bank, 2024: Funding a Water-Secure Future
- World Bank, 2024: Funding a Water-Secure Future: An Assessment of Global Public Spending, Executive Summary
- IEA, 2025: Energy and AI: Executive summary
- Eurostat, 2025: 20% of EU enterprises use AI technologies
- Eurostat, 2026: Artificial intelligence by NACE Rev. 2 activity (isoc_eb_ain2)
- S&P Dow Jones Indices, 2026: S&P Global Water Index