India’s AI Infrastructure Race Will Be Won at the Substation

Visakhapatnam’s proposed data center pipeline is already larger than India’s entire installed base. Google’s $15 billion AI hub includes a gigawatt-scale campus. Digital Connexion, the joint venture among Reliance Industries, Brookfield and Digital Realty, has outlined a separate 1-gigawatt project. Reliance has also floated a separate proposal for a reported 1.5-gigawatt project with captive renewable power and battery storage. The proposal has cleared a state investment committee.

India, by comparison, had about 1.5 gigawatts of data center capacity by the end of 2025, according to the Ministry of Electronics and Information Technology. The comparison is necessarily rough because planned capacity will arrive in phases, if it arrives at all. The useful distinction is between announced megawatts and working megawatts.

There is no shortage of capital commitments. On July 13, HCL Technologies Ltd joined the pipeline, planning to invest up to ₹3,500 crore, or about $365 million, in AI data centers, with capacity potentially scaling to 50 megawatts. But a 2026 study by the Council on Energy, Environment and Water (CEEW), a policy research institute, counted about $95 billion of commitments to India’s data center sector from 2019 through September 2025, and only about half could be tied to clearly identified recipient states and locations. 

India still has a powerful cost argument. Construction consultancy Turner & Townsend’s 2025-26 index puts Mumbai at $6.64 per watt, compared with $15.20 in Tokyo and $14.50 in Singapore. But a low construction cost does not secure a grid connection, shorten the delivery time for switchgear or settle who gets the water in a dry year. For companies planning AI workloads, those physical dependencies have moved from the facilities file to the business case.

Much Is Decided Before a Project Is Announced

Data center announcements can create the impression that a project begins with the investment decision. Operators describe the sequence differently. Land, power planning, connectivity and an approvals route usually come first. Customer demand may firm up while construction is underway.

“When AI-ready capacity is announced, the project has typically reached a stage where the key building blocks for execution are reasonably well defined,” said Vikram K., Chief Revenue Officer at Digital Connexion. He said the site and execution plan are generally established by then, while customer commitments may continue to evolve.

That does not make delivery routine. Power planning and grid connectivity can run beyond the construction schedule because they require utility coordination, feasibility studies, approvals and sometimes transmission upgrades. Specialized switchgear, generators and cooling equipment add another source of delay.

“The longest phase following an announcement is often the construction and commissioning process itself, which can take 18 to 24 months on average,” said Pratap Mane, President and Country Head for India at hyperscale data center operator Colt Data Centre Services. Testing is part of that timetable. Hyperscale customers pay for resilience, not merely a finished building.

AI has also changed what must be built inside it. Digital Connexion said average rack loads have risen from roughly 8-10 kilowatts to more than 150 kilowatts in some AI deployments. At that density, conventional air cooling may no longer be enough. Direct-to-chip liquid cooling, rear-door heat exchangers, and redesigned power distribution must be considered before construction, not added after racks arrive.

The skills mix follows the equipment. Operators now need more people who can work across electrical systems, mechanical cooling, accelerated computing, networks, automation, cybersecurity and resource efficiency. Digital Connexion said the strongest demand is for specialists who can connect physical infrastructure with AI platforms. A campus can have land and financing and still slip because the necessary utility, engineering, and commissioning work does not align.

Grid Access Has Become a Siting Decision

India’s data center capacity rose from about 375 megawatts in 2020 to roughly 1.5 gigawatts in 2025. Real estate consultancy Savills expects operational IT capacity across the main markets to reach about 4 gigawatts by 2030, while fellow consultancy Colliers estimates about 4.5 gigawatts.

The government’s forecast is larger. It expects electricity demand from data centers to reach 13.56 gigawatts by 2031-32. These figures measure different things. The market estimates refer to the IT capacity that facilities can host. The government figure refers to expected electrical demand. Combining them into a single growth series would overstate the forecast’s precision.

Whichever forecast lands closest, utilities still have to deliver far more power to a small number of zones, around the clock.

However, Ashish Arora, CEO of data center operator Nxtra by Airtel, said power generation itself was not the fundamental challenge. The constraint, he said, “lay in the availability, accessibility and delivery of power within specific data center zones.”

That distinction matters to a buyer. A state may have surplus generation while the proposed site lacks a ready substation, transmission capacity or a firm energization date. The 2026 CEEW study, based on consultations with 24 operators, cloud providers, renewable-energy developers, policymakers and specialists, found that grid connectivity remained a practical source of delay even where states advertised single-window clearances.

Policy has made capital easier to assemble. Data centers with an IT load above 5 megawatts have held infrastructure status since 2022. The 2026-27 Union Budget announced a tax holiday through 2047 for foreign companies providing cloud services to global customers through Indian data centers. Services sold to Indian customers must be routed through an Indian reseller. The SHANTI Act opened a route for private participation in nuclear generation, and the government presents small modular and micro reactors as future power sources for AI facilities.

None is a quick remedy for a project already in the queue. India’s nuclear mission aims to have at least five domestically designed small modular reactors operating by 2033. A campus due in 2028 still needs a credible grid and backup plan now.

New York offers an early warning. On July 14, it became the first US state to impose a statewide moratorium on environmental permits for certain new data centers with a capacity of 50 megawatts or more while it develops tougher standards. “These hyperscale AI data centers consume enormous amounts of power, truly threatening to outpace our grid’s capacity, and they drive up costs for local ratepayers,” Governor Kathy Hochul said. New York is not a template for India, but it shows what happens when project approvals run ahead of an agreement over who pays for the grid.

Water Use Depends on the Cooling Design

Water figures for data centers are often repeated without the cooling technology attached to them. A peer-reviewed study in npj Clean Water found that a 1-megawatt facility using traditional evaporative cooling can consume about 25.5 million liters of water per year. That is not a universal rate. Climate, utilization, cooling design and the source of electricity all affect the footprint.

The distinction becomes more important at AI densities. Liquid cooling can remove heat more efficiently near the chip, but the full design still has to reject that heat somewhere. Some systems consume water through evaporation. Others recirculate coolant in a closed loop and use far less fresh water during normal operations.

“Water stewardship is also a key consideration,” Mane said. Colt DCS is evaluating closed-loop systems intended to minimize water consumption and wastewater. At Digital Connexion’s Chennai facility, Vikram said rainwater harvesting and on-site treatment support the reuse of water for cooling and reduce the call on fresh supplies where feasible.

India’s problem is not that every new campus will use the same amount. It is that many will be built in places where several large users compete for the same resource. India has about 18% of the world’s population and 4% of its freshwater resources. CEEW found that operators’ own assessments of future water risk ranged from low to medium, even though smaller facilities often rely on municipal supplies and the largest hubs sit in cities with heavy competing demand.

Visakhapatnam is testing one possible answer. The reported Reliance proposal sets aside land for a desalination plant. That could reduce dependence on freshwater, but it would bring higher costs, increased energy use, and greater marine impacts. These are engineering choices with public consequences, not sustainability language for an annual report.

The policy gap is visible. CEEW found that 15 states had dedicated data center policies or covered the sector through IT and industrial policies. Only five explicitly included sustainability provisions. Incentives are detailed. Water accounting, local carrying capacity and long-term reporting are not.

Data Sovereignty Is Narrower Than the Sales Pitch

Domestic hosting is often sold as if every Indian workload must stay in India. The legal position is more specific. The most substantive obligations under the Digital Personal Data Protection Act, including its cross-border transfer provision, are scheduled to take effect in May 2027. The law generally permits cross-border transfers while allowing the government to restrict specified destinations or classes of data.

Sector rules already go further. The Reserve Bank of India requires payment-system data to be stored in India. The Securities and Exchange Board of India requires regulated entities using cloud services to store and process their data and logs within India. Government contracts and customer security policies can add their own residency terms.

That means leaders should classify workloads before paying a blanket “sovereign” premium. Some data must remain onshore. Some customers will choose domestic processing for latency, security or procurement reasons. Other workloads may legally and economically sit elsewhere.

CoRover, which develops BharatGPT and AI assistants for government and enterprise users, has used subsidized compute through the IndiaAI Mission. “For government-focused AI deployments, hosting GPUs within India is important to meet data sovereignty, compliance, security, and public-sector policy requirements,” said founder and CEO Ankush Sabharwal. “However, the real need is access to reliable, cost-effective, and readily available compute capacity within India.”

Sabharwal’s experience also cuts through a common assumption. Domestic capacity matters only when teams can obtain the right accelerators, at the right time and with enough support to run them. The government said 38,231 GPUs had been onboarded through 14 providers by March 2026, at a subsidized average rate of ₹65 an hour. Sabharwal said CoRover’s own approval, assignment and provisioning process worked without major difficulty.

Four Questions to Ask Before You Commit

The physical limits do not make India a poor location for AI. They change the due diligence. Four questions should sit beside model performance and cloud pricing in any large deployment decision.

What will be energized, and when? Ask for the utility commitment, substation plan and staged energization schedule. Separate IT capacity from total facility draw. A promised 100-megawatt campus is of little use if only the first block has a firm connection date.

How will the site reject heat at the density you need? Specify rack density, the cooling method and expected power usage effectiveness (PUE) and water usage effectiveness (WUE) under local weather and realistic utilization. Identify the water source, reuse plan and fallback if supply is restricted. A low annual PUE can hide a poor answer to a peak-summer problem.

Which workloads actually require Indian residency? Map each workload to the applicable regulator, contract and risk policy. This avoids two mistakes: sending restricted data abroad and paying to localize data that does not need to be localized.

Can the campus expand without reopening every dependency? Check the land bank, future power route, fiber diversity, equipment lead times and specialist staffing. AI hardware changes faster than buildings. The site should allow a new cooling or power design without a prolonged shutdown or a second approvals battle.

Boards should also ask who bears the cost if the load does not arrive, if the substation must be expanded or if local water rules tighten. States face the same question from the other side. Tax incentives may win an announcement, but power tariffs, water safeguards and community benefits determine whether a project keeps public support.

India’s construction advantage can get a campus approved. It cannot energize a rack. The winners in the next stage will be the companies and states that can turn a promised megawatt into reliable compute on the date it was sold, without shifting the hidden bill onto the grid, the water system, or the surrounding community.