AI Data Centre Water Use: Why the Hottest Day Matters More Than the Average
An efficient campus can still have an unfinished cooling system
Imagine an operator preparing to bring a new AI data centre into service. The servers are installed. The cooling equipment has arrived. The annual water estimate looks manageable.
Then the discussion moves to a hot afternoon. Can the connected supply deliver enough water, at the required quality and pressure, while other customers need the same service? Can the wastewater system accept what leaves the cooling equipment? What happens if the cooling design needs more electricity precisely when power demand is also high?
Those questions expose a different problem from the one suggested by a headline about annual litres. The operating constraint is not necessarily the amount of water used across a year. It is whether electricity, cooling, water supply, treatment and discharge can work together when the facility needs them.
This is not an argument that every data centre faces a water shortage. It is an argument for testing local operating readiness before treating an efficiency claim as proof that a project can perform.
Start with the heat, not the water headline
AI computers use electricity and produce heat. Cooling equipment moves that heat away from the chips so they can keep operating. But collecting heat inside the building is only the beginning. The heat must eventually reach equipment that releases it outside.
A liquid cooling system might circulate coolant through the servers, transfer the collected heat through a heat exchanger, and return that coolant to the chips. A separate circuit can carry the heat to outdoor equipment. If that equipment relies on evaporation, replacement water is still required.
That is why a closed coolant loop does not automatically mean zero site water consumption. The same liquid can circulate repeatedly near the chips while another part of the system evaporates water to release heat.
The useful system map therefore runs from electricity to chips, from chips to cooling equipment, and from cooling equipment to outdoor heat rejection. Water supply, treatment and wastewater connections attach to that chain wherever the chosen design requires them. Each connection must be tested, not assumed.
Annual averages hide a different operating question
The Interstate Commission on the Potomac River Basin reported that data centres in the Washington Metropolitan Area consumed about four million gallons per day on average in 2025, compared with about fifteen million on peak days.
These are regional consumption estimates, not the water intake of one campus. They do not prove that a shortage occurred. They do show why an average alone cannot describe the demand that a service may need to accommodate. The commission’s regional analysis makes the difference between average and peak demand visible.
For an individual facility, the question is narrower. How does its cooling demand change with weather and computing load? What capacity is actually available through its connection? Which other customers rely on that supply, and what operating conditions have the utility agreed to support?
A reassuring regional total cannot answer those questions. Neither can a company wide annual average.
Follow the water across the measurement boundary
Withdrawal means water taken from a source, such as a river, reservoir, well or utility system. Consumption is the portion not returned within the boundary being measured, often because it evaporates. A withdrawal figure should not be presented as though every litre disappears.
Direct and indirect water use also answer different questions. Direct use happens at the data centre. Indirect use can occur at power stations generating its electricity.
Berkeley Lab estimated that United States data centres directly consumed about 66 billion litres in 2023, with nearly 800 billion litres of indirect consumption associated with electricity generation. These are historical, modelled estimates for data centres overall, not measured 2026 totals or AI alone. They are not a complete account of water used to manufacture chips and equipment. The 2024 Berkeley Lab report provides the underlying scope.
The planning implication is straightforward: reducing the number on the campus water meter does not, by itself, establish the effect across the connected system.
A cooling choice can move the constraint
Evaporative cooling releases heat by turning some water into vapour. Dry cooling releases heat to air without depending on that evaporation step. Hybrid systems can use different approaches as conditions change.
Compare those choices against the same cooling task, local weather and computing load. Examine electricity demand alongside water demand. A design that reduces site evaporation may need more electricity than an evaporative alternative under comparable conditions. The wider water effect then depends partly on how that electricity is generated.
Microsoft’s December 2024 design announcement illustrates the distinction. It described a design intended to avoid evaporative cooling water while retaining administrative water uses and acknowledging an energy tradeoff. Its pilot and operating dates were future targets in that announcement, not evidence of completed fleetwide deployment.
A design announcement is useful evidence of direction. Operating performance still needs to be demonstrated at an installation, within a stated boundary and under relevant conditions.
Reclaimed water needs an operating pathway
Reclaimed water is wastewater treated for another suitable use. It can replace drinking water as an input to cooling, but the source change does not remove the need for treatment equipment, pipelines, connection capacity and maintenance.
Loudoun Water reports a reclaimed network with twenty miles of pipelines and more than 750 million gallons delivered in 2025. That volume covers all its reclaimed water customers, not data centres alone. Its reclaimed water program demonstrates a real service pathway, not universal availability at every proposed campus.
If reclaimed water evaporates during cooling, it is still consumed. The planning team should also ask where that treated water would otherwise have gone. Diverting water that would have returned to a river can change the downstream balance. That is a question to investigate, not a presumption that reuse is harmful.
The practical test is whether the connected source can supply suitable water when the cooling system needs it, with its wider effects understood.
The outlet can become the missing dependency
In an evaporative cooling system, dissolved material remains as water evaporates. Operators drain some circulating water, called blowdown, to control the concentration. That outgoing flow needs an acceptable receiving or treatment route.
Commissioning creates another interface. Cleaning, passivation and flushing can produce discharges that differ from normal operation. The receiving utility needs to understand their timing and characteristics before equipment enters service.
The Congressional Research Service describes wastewater and treatment challenges associated with data centre operations and commissioning. This does not establish that a particular project has been delayed. It explains why an intake connection alone is not a complete water plan.
If an essential supply or discharge arrangement remains unresolved, installed equipment can still be unable to perform its intended job. The possible consequences include additional treatment investment, revised commissioning sequences, equipment changes and a later operating date. These are project risks to test, not quantified losses established for a named campus.
Better intensity does not guarantee lower demand
Water Usage Effectiveness, or WUE, expresses water use relative to an energy denominator. It can support comparison when the scope and calculation are consistent. It does not report total litres by itself.
Microsoft reported average WUE of 0.27 litres per kilowatt hour for 2025. That is company reported, and the cited water update does not fully specify its calculation boundary. The number should not be used alone to rank different operators.
Consider an illustrative index. Activity starts at 100 and then doubles. Water use per unit falls by 25 percent. Total water use becomes 200 multiplied by 0.75, or 150. Efficiency improves while total demand rises by 50 percent. This is arithmetic, not a forecast.
Replenishment claims answer another question again. A beneficial project elsewhere does not automatically deliver water through a campus pipe on a hot afternoon. Location, timing and the outcome being counted still matter.
What capable operators should do before committing
Strong planning begins with one integrated operating profile rather than separate equipment, water and sustainability approvals.
Establish the cooling requirement under relevant weather and computing loads. Agree the peak assumptions with the supplying utility. Document the required water quality, connection capacity and response to interruption.
Compare cooling alternatives on equivalent terms, including electricity, treatment and conveyance. Confirm that proposed reclaimed infrastructure is connected and in service, rather than simply planned. Align commissioning discharges with receiving limits and responsibilities.
Then assign ownership to the interfaces. Someone must be accountable for proving that supply, equipment and discharge arrangements work together. Track total volume alongside intensity, using consistent boundaries.
Suppliers should be evaluated against those operating jobs. Ecolab, for example, offers coolant distribution, monitoring and water management technology for data centre cooling. Its integrated cooling offering establishes a direct connection to the problem. Customer adoption, measured operating improvements and commercial materiality remain separate evidence questions. An offering is not proof of a customer outcome.
What to watch over the next 12 to 36 months
Three conditional paths matter. Wider dry cooling adoption could shift more of the constraint toward electricity requirements. Reclaimed connections could reduce potable demand while creating treatment and pipeline dependencies. Clustered development could increase the importance of utility expansion and peak capacity planning.
Watch commissioned capability rather than announcements. Useful indicators include actual reclaimed connections, disclosed peak demand, operating performance in hot weather, wastewater acceptance conditions and utility capital programs.
Evidence that those services work together would weaken the case that water is limiting a particular site. Persistent gaps between demand and deliverable service would strengthen it. The direction of the constraint must be observed locally, not inferred from a global average.
The Sarga II insight
The AI water problem is often framed as an efficiency contest. The more useful operating question is whether the complete cooling system can perform when conditions are demanding.
Follow the heat until it leaves the building. Follow the water until it returns or is consumed. Connect both journeys to electricity, treatment and local service capacity.
A small annual water number is not the same as a resilient operating system. The strongest projects will prove the connections before they depend on them.
Comments