Three of India’s largest announced data-centre campuses have each made a different choice about water. Google says its 1 GW AI hub in Visakhapatnam will use “advanced air cooling”, after its own assessment found pressure on the local watershed. Reliance’s first 168 MW in Jamnagar, leased to Meta, is reported to cool with desalinated seawater. In Thane, Amazon says its new site will use air-cooled chillers and almost no water, while the city has offered it 12 million litres a day of treated sewage. Maharashtra has meanwhile told data centres to run on treated wastewater, and in October set up a committee under Anil Kakodkar to assess their water use.

Behind these choices is one question, usually framed as “closed loop or open loop”. This piece explains what the terms mean, works out the water and energy each uses per unit of computing, maps the announced campuses against the coast and the subsea cables that pull them there, and recommends a design for each kind of site. It follows this blog’s piece on where the gigawatt data-centre bets land.

What “open” and “closed” actually mean

Two things are easy to miss. First, “closed loop” often describes only the loop inside the building. A direct-to-chip liquid-cooled hall is closed at the rack, but the heat it collects still has to leave the site, and if it leaves through an open cooling tower the site still evaporates water. Microsoft’s “zero water for cooling” design, announced in 2024 and avoiding a projected 125 million litres a year per data centre, is closed at both ends, and Microsoft acknowledged that it raises energy use. Second, “open” covers two different things: a cooling tower, which consumes water by evaporating it, and once-through cooling, which borrows water and returns it warmer. Google’s Hamina centre in Finland pumps seawater through heat exchangers that cool a separate closed freshwater loop, and mixes the outflow so it returns at close to the sea’s temperature.

The trade: water on site, or electricity

The water an evaporative tower uses follows from physics. Evaporating a litre of water near 30°C absorbs about 2.43 megajoules, and a kWh is 3.6 megajoules. A tower serving a water-cooled chiller has to reject the IT heat plus the chiller’s own work, about 1.16 kWh of heat per kWh of IT load, and does about 85% of that by evaporation: about 1.46 litres. Because evaporation leaves salts behind, some water has to be bled off. At three cycles of concentration the total is about 2.2 litres per kWh of IT load, and at six cycles about 1.8.

Dry cooling saves that water but costs electricity. At the minimum efficiencies in India’s Energy Conservation Building Code (2017), a large water-cooled chiller plant uses about 0.20 kWh, tower fans and pumps included, to remove a kWh of heat; an air-cooled chiller uses about 0.33 kWh. That is a penalty of 0.135 kWh per kWh of IT load when compressors run every hour, which is the worst case. India’s electricity is still about three-quarters thermal, and NTPC’s coal plants consumed 2.68 litres per kWh in 2023-24, so the extra electricity carries about 0.27 litres of water at the power plant.

So even in the worst case, dry cooling saves about 88% of the water once the power plant’s share is counted, and the water it does use is drawn in the coal belt rather than from a city’s reservoirs. The cost is energy. For 100 MW of IT load running at 70% on average, an open tower would use about 3.7 million litres a day; dry cooling with compressors would instead use about 83 GWh a year more electricity, about Rs 58 crore at an assumed Rs 7 a unit. That works out to Rs 491–636 for every kilolitre of water saved. Seawater desalination in Chennai has been costed at Rs 30–49 a kilolitre, and tertiary-treated sewage for industry at about Rs 45.

The worst case is avoidable. Modern AI chips are increasingly cooled by liquid at the chip, and ASHRAE’s classes allow facility water at up to 40°C or 45°C. A dry cooler can supply water a few degrees above the outside air temperature, so at those temperatures it can run without compressors on all but the hottest hours. An Uptime Institute briefing in April 2026 found that well-designed dry and adiabatic systems can match evaporative ones on energy across ASHRAE climate zones 2 to 6. Most of lowland India is in the hotter zones 0 and 1, outside that range, and a July 2026 preprint found dry cooling alone infeasible at design conditions in the hottest US desert cities. That is why the design temperature of the liquid loop matters more in India than the label on the cooling plant.

How much water India’s data centres use is not known

The figures in circulation imply anything from under 1 to more than 11 litres per kWh. The physics puts evaporative cooling at about 1.8–2.2 and dry cooling near zero on site. The gap is a measurement problem: the Bureau of Indian Standards has adopted the international Water Usage Effectiveness metric (IS/ISO/IEC 30134-9:2022), but no state or central rule requires operators to report it.

Where the campuses are, and why the coast matters

India’s data centres cluster at its subsea cable landings. Mumbai held about half of operational capacity in the first half of 2026 (890 MW of 1,789 MW, per Cushman & Wakefield), and Chennai is second. Visakhapatnam is becoming a third landing hub, with Google’s America–India Connect cable due to land in the city where its campus is being built. Proximity to cables matters most for cloud and content services that serve users; AI training is far less sensitive to latency, which is why gigawatt campuses are also going inland to Hyderabad and Greater Noida.

Capacity and cooling as announced or reported; none of these operators publishes water use for an Indian site. Distances are straight lines between city centres, not fibre routes.

The coast changes the cooling choice. Every coastal campus in the table has seawater next to it, which can cool a closed loop through a heat exchanger without any fresh water, or be desalinated for a tower. Desalination takes about 3.5 kWh per cubic metre, so feeding a tower with it adds about 0.0077 kWh per kWh of IT load, about a twentieth of dry cooling’s worst-case penalty. The constraints are permits and the sea itself: intake and outfall structures need coastal regulation zone clearance, coastal power plants are typically bound to discharge water no more than 7°C above the sea, and pipes foul and corrode. An IIT Bombay study modelled cooling a 100 MW centre with cold water from the deep ocean and reported energy savings of 79%, while noting that the long pipelines are costly to build and run. The same coast is also where the water stress is acute: Mumbai cut supply by 10% from 15 May 2026 with its lakes at 23% of capacity, and Chennai’s reservoirs ran almost dry in 2019.

Recommendation

Neither loop wins everywhere. On the coast, an open seawater loop serving a closed inner loop is the efficient answer, and the rule should be that no data centre cools with drinking water. Inland, the closed loop wins, but only if the building is designed for warm-water liquid cooling, so that dry coolers are not running compressors through an Indian summer. The one design that should not be built in either place is an open tower fed from a city’s potable supply.

What would fix this

• States hosting data centres should require annual reporting of Water Usage Effectiveness under IS/ISO/IEC 30134-9, split by source (potable, groundwater, treated sewage, seawater, desalinated), together with PUE, for every site above 5 MW.
• “Closed loop” and “zero water” claims in project approvals should state how heat leaves the site, not only how it leaves the rack.
• Coastal states should set out a clearance route for seawater intake and outfall for data centres, with a discharge temperature limit and monitoring, so that seawater cooling is a practical alternative to municipal water.
• Maharashtra’s Kakodkar committee, due to report within three months of its October order, should publish the water allocation for each data-centre hub and the source it comes from.

These are proposals from this blog, not recommendations any agency has adopted.

What this piece could not establish

No Indian operator publishes site-level water or cooling energy, so the comparison rests on physics and code-minimum efficiencies rather than measured plants; real towers and chillers can do better than the code minimums. Hourly weather data for Indian cities could not be obtained for this piece, so it does not estimate how many hours a year dry coolers could run without compressors in each city. The Rs 30–49 desalination and Rs 45 treated-sewage costs are older Chennai estimates, and the Rs 7 electricity tariff is an assumption. Maharashtra’s 4,040 MW Navi Mumbai figure and the water demand attached to it come from a syndicated report of a state assessment that could not be opened. Several sources, including government releases, could not be opened directly and were read through search-engine text and news reports.

Sources and caveats

About this article: Researched, written and edited by Umashankar Triplicane Dwarakanathan, with AI research assistance; every figure is meant to trace to the primary source cited. See the Editorial Policy for how sourcing, AI use and corrections work.