AI’s next infrastructure consideration could be water

AI is fuelling the need for computing density, specialised accelerators, and energy-hungry data centres

The AI infrastructure discussion so far has focused primarily on one issue: The source of energy.

AI (Photo credit: Unsplash)

Indeed, that is a valid point. AI is fuelling the need for computing density, specialised accelerators, and energy-hungry data centres. However, as the AI infrastructure of India grows, there is another issue of resource scarcity that should be taken into consideration equally, and that is water.

And this is no distant issue of sustainability; this is fast becoming an infrastructure planning issue.

The environmental impact of AI is usually examined via power usage and CO2 emissions. However, high-density computing creates an immense amount of heat, which has to be cooled constantly.

According to a WRI India Water Research report, AI's total water footprint will amount to 4.2 - 6.6 billion cubic metres per year by 2027. It is also mentioned that two-thirds of data centres constructed after 2022 are situated in water-stressed areas. Direct cooling, water required for power production, and semiconductor fabrication are among the aspects of water footprint associated with AI.

The Indian data centre market is growing fast, with AI set to play an important role in driving demand for high-performance computing.

According to a May 2026 analysis by WRI India, over 50% of India’s data centres are located in areas where there is water stress. In addition, the report established that 75% of India’s data centres are based in Maharashtra, Tamil Nadu, Karnataka, Telangana and Uttar Pradesh. Some of these data centres are located in water-stressed areas.

What makes this relevant is that water availability is essentially local in nature. Power production and distribution can increasingly happen over much broader networks. The same cannot be said of water availability.

What this means is that water balance needs to be taken into consideration when setting up a data centre in an area that is experiencing water stress.

The silver lining is that there is no trade-off between AI progress and sustainable water consumption.

There have been significant advancements in cooling technologies. Liquid cooling, direct-to-chip, immersion cooling, closed-loop, and other cooling technologies are capable of decreasing reliance on traditional cooling methods. Newer cooling technologies can help data centres manage rising computing needs more efficiently, while reducing reliance on traditional water-intensive systems.

It is essential to treat cooling technology as more than just a utility. India’s policy landscape too is taking steps to recognise this issue.

The government also said that the District Cooling Guidelines of the Bureau of Energy Efficiency suggest the use of treated water in cooling tower make-up water where feasible. As mentioned in the same reply, the Resilience, Innovation & Efficiency Working Group at the India AI Impact Summit focuses on efficiency and environmental resilience in AI infrastructure.

This is a significant step in itself. Water efficiency, which was more of a sustainability concern earlier, is now getting included in infrastructure discussions regarding AI.

For many years now, PUE or Power Usage Effectiveness, has been among the most common efficiency metrics in the industry. However, with the increasing intensity of AI operations, it cannot be the only way of measuring the efficiency of the infrastructure.

It is important for data centre operators and policymakers to have insight into WUE, or Water Usage Effectiveness, water stress in the area, cooling technology used, water source, and options for recycling or reuse of water.

The 2026 Water Research report argues for the same, calling for facility-level transparency regarding water usage and for the inclusion of water restrictions while developing infrastructure for AI systems.

This is because a facility-level efficient infrastructure design can simply shift environmental impact elsewhere. Thus, instead of optimising a single metric, it is important to optimise the water, energy, and carbon metrics as a whole.

It is important that the sustainability dialogue shifts beyond just the physical data centre. Workload provisioning, scheduling and management will affect the utilisation of underlying infrastructure. Resource planning, optimised workloads, virtualisation, and the right cloud architectures can go a long way in helping businesses avoid paying for, and using resources for, unnecessary compute capacity.

This is especially true in the case of India’s MSMEs and AI adopters. Scalability and security are essential for cloud infrastructure, but transparency with regard to resource utilisation, economics, and underlying infrastructure becomes critical.

India’s AI opportunity is too big to view sustainability as a trade-off between growth and prudence.

A better approach to take would be asking ourselves how to construct our infrastructure around AI keeping into consideration the resources that go into sustaining this infrastructure.

Power shall continue to be the biggest bottleneck, but in some regions of the world, water shall become equally crucial.

The answer does not lie in reducing the rate of uptake of AI. It lies in making resource efficiency an integral component of AI infrastructure right at the outset--through intelligent location choices, cooling systems, water reuse, measurement of efficiencies and cloud architecture.

AI is virtual at the level of application, but very physical at the level of infrastructure.

The next stage of India’s AI infrastructure expansion should thus be marked not only by the number of GPUs deployed or the compute capacities added, but by the smartness with which power, water and infrastructure is used to run the GPUs.

India's responsible AI infrastructure will eventually mean constructing around both digital scale and resource resilience.

(The views expressed are personal)

This article is authored by Rahul Takkallapally, co-founder, BharathCloud.