The environmental cost of AI is a polarising subject. Some reports are clearly exaggerated, but we still need to be mindful of the harms that can result to communities exposed to them. India’s draft data centre policy rewards efficiency but says nothing about mitigating these concerns. There is still time to fix that.

This July, several hundred residents of Thane formed a human chain outside a 53-acre plot in Balkum that had been earmarked for a hyperscale data centre. They were agitating against the noise and heat it would generate, and the water it would consume in a city that is already struggling to supply its residents. This, as far as I can tell, was the first time Indians have taken to the streets against a data centre. It will not be the last.

The Real Cost of a Prompt

The arguments they were making are not new. Claims that each question we put to ChatGPT consumes half a litre of water and uses 10 times the electricity of a Google search have been around since 2023. But the soundness of these allegations have been questioned. The figures, it is said, come from outdated publications that misrepresent current facts. In August last year, Google stated that a median text prompt to Gemini consumes 0.24 watt-hours of electricity and 0.26 millilitres (about five drops) of water. If this is correct, then each prompt costs roughly what a Google search does—far less than what is alleged. But Google’s figure is a median for text and conveniently ignores the fact that reasoning models consume a hundred times and video generators about 4,000 times that figure for a 5-second clip.

What is much harder to defend is the noise pollution caused by the incessant whine of fans that are needed to keep servers cool. Because of its steady, low frequency, it carries in a way that noise pollution caused by traffic does not. That said, a well-built data centre can keep the sound to between 45 and 65 decibels at its fence, which is quieter than a highway.

The other concerns are also less alarming than they appear to be at first glance. Data centres used about 1.5% of the world’s electricity in 2024 and produced 0.5% of its emissions from fuel combustion. This is about a fifth of the energy we consume for air-conditioning and the International Energy Agency expects them to account for less than a tenth of additional demand projected till 2030. The Council on Energy, Environment and Water puts Indian data centres at two or three hundredths of a per cent of our water demand.

Local Costs

None of this is to suggest that protestors in Thane are wrong to worry. The environmental cost of AI should not be calculated on a per-prompt basis, but must instead be evaluated in terms of where each data centre is built, what powers it, and who it is going to affect because they live next door to it. In Ireland, since hyperscalers were allowed to cluster around Dublin, 23% of its electricity now goes to data centres.

Two-thirds of the data centres built in the US since 2022 sit in places already short of water. By that reckoning, India is also building them in exactly the wrong places. Google’s $15 billion AI hub in Visakhapatnam is coming up in a district with the lowest groundwater in Andhra Pradesh, while Mumbai, which accounts for over half of our installed data centre capacity, had to ration water this summer.

The draft National Data Centre Policy that India’s ministry of electronics and information technology (IT) circulated last year ties its incentives to power-usage effectiveness (the ratio of the electricity a facility draws to what its servers actually use). However, it does not appear to have addressed issues about water, siting or noise. Since the cheapest route to power-usage effectiveness is evaporative cooling, we may find ourselves trading electricity efficiency for chronic water shortages.

The Other Half

Last year, I argued in this column that we should run the same playbook that built our IT industry to succeed at AI. But that is only half of what we need to do. The other half has to do with addressing the concerns of Balkum’s residents.

What needs to be done is not difficult. Wherever measures like this have been demanded, they have eventually been taken. When a Uruguayan court forced the government to disclose that Google’s Canelones data centre would draw 7.6 million litres of drinking water a day, the public reaction was enough to get the company to redesign it to run on air-cooled chillers. Microsoft’s Hyderabad data centre runs a closed loop that consumes almost no water for cooling. In Chandler, Arizona, the noise produced by a data centre can never exceed the ambient level of the baseline recorded prior to construction. Karnataka’s new policy requires data centres to report their water use every year, and Maharashtra has commissioned standards for siting, water and noise.

Our national policy must address these concerns head on. Every data centre should be obliged to report both its water-usage and power-usage effectiveness. Every new site should pass a test that weighs the water stress of the basin it draws from and the homes that depend on the same water. The noise it emits while operating at peak load should never exceed the baseline measured before construction. None of this will slow the industry down. These are all measures that have been adopted elsewhere and the technology exists to make it feasible.

Stories about the amount of water AI consumes will outlive our attempts to correct the narrative—because the myth is far more vivid than reality. But all that residents of Balkum really want is the assurance that the facility that’s being built would not impose an impossible burden on those who have to live with it. These are concerns that our data centre policy can easily address, and since we have not yet finalized it, it is not too late to include them.