“Every time you ask LLM a question, you are consuming vast amounts of water.”
That claim has circulated widely, generating alarm among technologists and environmentalists alike about the environmental cost of artificial intelligence. Now, Sharad Agarwal, CEO of Sify Technologies, a company that has been at the centre of India's digital infrastructure story for over three decades, wants to put that myth to rest.
"The biggest myth is that every time they go and ask a question to AI, they are consuming water," he says, matter-of-factly. "No, it's not the case."
Cooling the Conversation
The confusion, Agarwal explains, stems from a genuine technical reality that no longer reflects how modern data centres operate.
There are broadly two ways to cool a data centre. The first—evaporative cooling—does consume significant quantities of water. In this method, potable water is brought in and allowed to evaporate to lower temperatures, much like a household cooler. According to a US Department of Energy report, this approach consumes 780–800 million gallons of water per MWh.
In a water-scarce country like India, that model was never sustainable.
The second method, now widely adopted in India, is the closed-loop water chiller system. Here, pipes are filled with water just once, requiring only 3,000–10,000 litres per megawatt as a one-time charge, and thereafter only minor leakage needs to be replaced, amounting to no more than 100–1,000 litres per megawatt annually. The water circulates endlessly within a sealed system, never leaving it.
"India has adapted the closed-loop water system," Agarwal tells AIM. "We are not dependent on water at all."
He adds a third layer to the picture: grey water usage. Sify and others are increasingly combining closed-loop chiller systems with treated or industrial grey water for any evaporative cooling needs, creating a hybrid approach that is both efficient and environmentally responsible. This also has the secondary effect of managing the ambient temperature around the data centre itself.
His views align with the AWS and Microsoft leadership, who have argued that their cooling systems use far less water than popularly believed. At Microsoft Build 2026, CEO Satya Nadella said new architectures are redefining data centre water consumption.
“The daily water usage over the course of an entire year is roughly equivalent to what a single restaurant would use," he said during his keynote.
For Agarwal, the industry consensus is now clear: the water-consumption narrative needs updating.
Why Every Company Ends Up in Mumbai
India's data centre geography tells a story of path dependency as much as planning. Mumbai remains the undisputed capital of the country's data centre market. "The first wave of data centre needs were in Mumbai and Chennai, and that created a stickiness," Agarwal explains.
The clinching factor came when global hyperscalers arrived. To set up operations in India, they needed three data centres located within 10–15 kilometres of each other to ensure redundancy and low latency across availability zones. At the time, Mumbai was the only city with the existing density of infrastructure to make that possible. The hyperscalers built there, and in doing so, effectively locked in the city's dominance.
For those looking beyond Mumbai, Agarwal points to the logic underpinning expansion into newer markets—including GIFT City and other tier-2 locations where Sify has been acquiring land. He is emphatic that this is not opportunistic land-banking driven by cheaper real estate.
"Real estate is not cheap anywhere in the country, especially in tier-2 cities," he notes. The rationale, instead, is demand-driven: four distinct categories of customers are pulling infrastructure into new geographies—hyperscalers, GPU neo-clouds, banking, and financial enterprises, and edge computing requirements from state governments now embedding AI into their operations.
The tier-3 vs tier-4 Question
One area where Agarwal pushes back against conventional wisdom is the assumption that India needs to aspire to tier-4 data centres—the gold standard in redundancy, where every component has a complete backup system running in parallel.
His reasoning is both technical and environmental. Tier-4 infrastructure requires N+N redundancy—effectively duplicating every system—which results in vast amounts of idle equipment consuming power and resources at all times. That redundancy, he argues, has a real sustainability cost that is often overlooked.
In practice, Sify and most global hyperscalers operate on an N+1 or N+2 basis, providing one or two redundant systems for every active one. Application-level redundancy, he points out, further compounds the protection already built into the infrastructure.
An enterprise deploying an ERP system will typically run it across two servers, with mirrored storage, a primary site, a near-disaster recovery site, and a far-disaster recovery site—each with its own mechanical, electrical, and cooling redundancies.
"The level of redundancies we create... are also wastage," he says. "We need to keep that in mind."
Only 2–3% of demand in India actually requires tier-4, he estimates, typically driven by specific regulatory requirements. For the rest, tier-3 is not a compromise; it is the appropriate solution.
To address latency, often cited as the argument for higher-tier infrastructure, Sify has taken a different approach. Sify, has established express fibre routes connecting 91 data centres across the country. This dedicated network, separate from carrier-grade public infrastructure, has allowed the company to deliver latency performance that it says competes with anything available in the region.
What Hyperscalers Cannot Replicate
As AWS expands its availability zones, Google commits to billion-dollar investments, and Reliance builds its own hyperscale infrastructure, the question of where an indigenous player like Sify fits becomes sharper. Agarwal's answer is confident: experience, credibility, and what he calls predictability.
"Fast time to market, that's what our customers get," he says. Building a data centre in India is not a purely technical exercise. It involves navigating more than 30 different regulatory authorities and approvals, coordinating across physical construction, power, water, and network infrastructure, and doing so within a system that rewards local knowledge and relationships built over decades.
"That Indian knowledge, the ability to navigate the Indian system, is important," he says. He claims Sify can take a data centre from land acquisition to production in under 24 months. That, he suggests, is something no global hyperscaler can match on Indian soil.
Looking ahead, he sees no shortage of demand. AI will continue to dominate data centre capacity in the near term, with inferencing taking priority over training in the immediate horizon. But as AI matures and enterprise adoption deepens, traditional workloads, banking, financial services, and sovereign cloud will continue to grow alongside it.
"For the next two to three years, we will always be playing catch-up to the demand."
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.