India’s data centre industry has spent the past decade solving one problem: how to build enough capacity to accommodate the country’s accelerating digital economy. Artificial intelligence is changing the nature of that challenge.
The constraint is no longer simply the availability of space, connectivity or even headline megawatts. AI workloads are increasing rack densities to levels that put power distribution, cooling and the economics of infrastructure under a different kind of pressure.
For Ankit Saraiya, Director & CEO, Techno Digital, one of the industry’s biggest challenges is also one it has created for itself. “The challenge is neither, in my opinion. The challenge is self-inflicted,” he says.
His argument centres on geography. Data centres are increasingly concentrated in Tier 1 cities, even as they become among the most resource-intensive forms of digital infrastructure.
Saraiya offers a simple illustration. If a city has 4 GW of transmission or distribution capability and 1 GW of data centres is added, one industry would consume a quarter of that capacity. The competition is no longer only between businesses. It is also for resources required by the wider community.
“The challenge is not that we don’t have the generation capability or we don’t have the transmission or distribution capability,” he says. “But till the time you locate resource-guzzling industries within prime locations, it’s always going to be a challenge.”
For an industry that has historically preferred proximity to connectivity and customers, the argument points towards a more distributed geography. Large campuses located closer to generation assets and the grid could eventually become a more logical proposition than continuing to concentrate them within resource-constrained urban centres.
From power availability to power economics
Saraiya’s perspective is shaped by Techno Electric’s nearly four-decade history in the power sector. The company has worked across power generation, transmission and EPC, before Techno Digital entered data centres in 2020.
The first data centre project began in Chennai in 2021 and has since been commissioned with 5.6 MW of IT load, although the facility has the potential to support almost 40 MW. The company also entered into a revenue-share partnership with RailTel for edge data centres across 102 cities, of which facilities in Gurgaon and Mumbai have been delivered. A separate RailTel partnership is developing a Noida facility that will begin with 10 MW and has the potential to expand to 40–50 MW. The first 5 MW is expected by March 2027, followed by the remaining 5 MW over the subsequent 24 months. A Kolkata facility is at the foundation stage and is targeted for commissioning by FY29.
But the more consequential change is happening inside these facilities. For years, PUE has been one of the industry’s primary measures of efficiency. AI is introducing another metric: how much useful computation can be generated from every unit of power consumed.
“It starts from the fact that how much power I’m consuming and how many tokens I’m generating,” Saraiya points out.
The economics are compelling. He estimates that power accounts for 40–60% of an operator’s cost. If the same amount of electricity can produce more AI output, efficiency becomes directly linked to revenue rather than being merely an infrastructure metric.
The change becomes stark at the rack level. A facility designed around 8–10 kW racks faces a very different engineering problem when AI workloads push requirements towards 150 kW, 200 kW or even 250 kW. Moving from an 8 kW rack to a 200 kW rack represents a 25-fold increase in power consumption per rack. “It is the game of power now,” he says.
The implication is that real estate itself becomes less important relative to the amount of electrical capacity that can be delivered to it. Saraiya cites an example from Chennai where a facility originally designed around 8–10 kW racks and roughly 24 MW of capacity could potentially deliver around 40 MW from the same real estate when configured for high-density racks. “The real resource is power. Who can get me 30–40 MW today? And at what price?” he asks.
Efficiency is moving beyond automation
AI will not necessarily replace the automation already embedded in modern data centre operations. Instead, Saraiya sees it as another layer of intelligence over increasingly automated infrastructure.
The same evolution has already taken place in the power sector. Substations that once required personnel on site can increasingly be monitored and operated remotely through automation and software.
Data centres are following a similar trajectory. Techno Digital is already using software with built-in AI functionality for operational efficiency, performance improvement and maintenance predictability. Saraiya says that at its Chennai facility, consumption was initially almost twice the eventual level. Through engineering and operational practices, automation and AI, consumption has been reduced by almost 30%.
The significance is less about AI being deployed as a fashionable technology and more about what happens when intelligence is applied to a facility where energy is one of its largest operating costs.
The power advantage is also a supply-chain advantage
Techno Digital’s power-sector heritage also affects how it approaches data centre development. Saraiya identifies three advantages. The first is understanding grid reliability before placing mission-critical infrastructure on it. The second is understanding how different sources of electricity can be accessed, including renewable and conventional power. The third is the company’s EPC capability.
He says that in June, 96% of the power consumed was sourced from renewables and that the figure rose to 97% in July, with the electricity coming from wind turbines rather than being offset through certificates.
The EPC capability has another consequence. Because of established engineering teams, labour access and supply-chain relationships, Saraiya says Techno Digital can deliver a data centre 30–40% faster than many operators that remain dependent on general contractors, MEP contractors and external supply chains.
The claim is particularly relevant in a market where customers increasingly want capacity available in months rather than years. The company’s access to transformer manufacturing capacity and established OEM relationships, he argues, can shorten the development cycle.
Cooling is becoming a function of rack density
The same power-density equation is reshaping cooling. Saraiya does not see air cooling, direct liquid cooling and immersion cooling as competing technologies where one will simply replace the others. Their relevance depends on the amount of heat generated by the workload.
“The denser the racks become… the more concentrated cooling is required,” he says.
At 6–8 kW rack densities, air cooling can remain practical. As density rises, direct liquid cooling becomes increasingly relevant because it can remove heat closer to the chip.
The sustainability calculation is also more nuanced than measuring water consumed inside the facility. Saraiya cites an illustrative PUE of 1.6 for an air-cooled facility compared with 1.25 for a direct liquid-cooled facility. At 1.25 PUE, every kilowatt of IT load requires 25% additional power, compared with 60% additional power at 1.6 PUE.
His argument is that the additional electricity consumed by an air-cooled facility also carries an upstream resource cost, because power generation itself requires cooling and water.
The sustainability question, therefore, cannot be confined to the data centre’s physical boundary.
AI-ready means power-ready
For existing facilities, becoming AI-ready may not require rebuilding the entire structure. But the electrical and thermal architecture has to accommodate dramatically higher densities.
“If you had designed a data centre for an 8–10 kW rack, can you accommodate a 150-kW rack?” Saraiya asks.
The answer depends on whether the power distribution system can carry the increased current and whether the cooling architecture can handle the resulting heat. Higher-density workloads can require changes to busways, power distribution and cooling technology, including direct liquid cooling.
Modern facilities are likely to have an easier transition than older data centres because they can be designed with greater flexibility from the outset.
The case for power beyond the grid
The longer-term search for power could eventually take the industry towards technologies such as small modular reactors. Saraiya sees SMRs as promising because they could enable generation and consumption behind the meter. But he does not regard them as an immediate solution. Both the technology and regulatory environment still have to mature.
He expects aspects of SMR technology to become commercialised over the next four to six years, while noting that data centre economics would require substantial scale. “Unless you have a gigawatt, 500 MW-plus operational data centre with you, until that point of time these SMRs may not make sense,” he adds.
For a company rooted in the power sector, he says the opportunity could also lie in the grid infrastructure required around future SMR projects.
The direction is clear even if the timeline is not. AI is turning the data centre into an increasingly energy-intensive industrial asset, forcing operators to reconsider where facilities should be located, how power should be sourced and how much computing value can be extracted from each kilowatt.
The next phase of India’s data centre race, therefore, may not be won by whoever builds the largest facility.
It may be won by whoever solves the harder equation between power, density, cooling and useful compute.