TL;DR

  • Real Estate vs. Compute Trade-Off: India offers substantial savings on land, shell construction, and labor, but imported compute hardware like GPUs receives no local discount and incurs additional import duties, logistics, and currency exposure.
  • Cooling and Power Penalties: Hot and humid climate conditions increase mechanical cooling demands and necessitate early liquid cooling adoption, while local grid instability requires heavier investments in power resilience and backup redundancy.
  • Strategic Capital Reallocation: Successful operators must reallocate budget away from physical buildings and toward modular design, advanced cooling, and power infrastructure to prevent underfunding critical operational systems.

# # #

Every AI infrastructure budget starts with the same global numbers: roughly $11.3 million per MW for a standard 2026 build, north of $20 million per MW once liquid cooling and high density enter the design, and $45-55 billion per gigawatt at hyperscale. Those figures are accurate, and applying them directly to an Indian project will misbudget it by 30-40%, in both directions. India doesn’t just cost less; it costs differently, because the drivers behind each line item behave very differently than they do in Northern Virginia or Frankfurt.

Getting this right matters more every quarter. India’s operational capacity reached about 1,123 MW of IT load by mid-2025, and roughly 387 MW was added in 2025 alone, more than double the prior year. The build wave is here. The benchmark it’s being budgeted against is the wrong one.

The India numbers, and where they diverge

To answer what it costs to build an AI data center for large-scale model training, across compute, power, cooling, and land, in India specifically, you have to adjust every line of the global benchmark rather than convert it.

Start with the headline. A Tier III colocation build in India costs roughly ₹50-65 crore per MW of IT load, about $6-8 million per MW. Compared with the global shell-and-core average of $11.3 million, that’s the 30-40% discount the market talks about. AI-ready high-density facilities (20-50+ kW per rack with liquid cooling) costs ₹70-95 crore per MW. A 50 MW hyperscale build including land lands around ₹3,000-4,500 crore ($360-540 million), with 100 MW campuses reaching ₹6,000-9,000 crore ($720 million-$1.1 billion).

But the discount is not uniform across the four cost components, and that’s the whole point.

Land and construction: India’s genuine advantage

This is where the 30-40% saving actually materialises. Land is significantly cheaper than in FLAP-D Europe or Silicon Valley, construction labour costs are a fraction of those in Western markets, and the shell-and-core scope has lower input costs across the board. For the physical building, India is one of the most cost-efficient large markets in the world. If your project were nothing but a shell, the global benchmark would overstate your cost by a third.

Compute: where the discount disappears

It isn’t anything but a shell. GPUs and accelerators account for roughly 60% of an AI facility’s capital, yet they receive none of India’s local advantage. Silicon is priced globally in dollars, and India then adds import duties, logistics, and rupee-dollar exposure on top. So the single largest cost bucket in an AI build either matches global pricing or exceeds it, and because it’s the biggest bucket, it drags the whole project’s effective discount well below the 30-40% the shell suggests. Budget compute at global-plus, not India-minus.

Cooling: an India penalty, not a saving

Here the direction reverses outright. India’s hot, humid climate is among the least favourable for data centre cooling. Temperate markets harvest hundreds of free-cooling hours a year from cool ambient air; across most of India, that opportunity is limited, so mechanical cooling runs harder, longer, and at higher energy cost. Water scarcity compounds the problem, limiting evaporative approaches and pushing designs toward liquid- and air-cooled chillers. Globally, mechanical scope has already climbed from about 22% to 33% of AI facility cost; in India’s climate, cooling carries even more weight. What global benchmarks treat as a moderate line item is, in India, one of the largest sources of both capital and operating cost.

Power: the wild card

Power cuts both ways. Electricity and land can be cheaper in the right state, and behind-the-meter and captive generation are increasingly viable. But grid reliability is the offsetting reality: variable supply and voltage instability demand heavier backup, larger UPS energy stores, more generator capacity, and higher redundancy than a stable Western grid requires. Grid-connection timelines and substation availability add schedule costs. Net, an India power train is often engineered for more resilience than the global benchmark assumes, and that resilience isn’t free.

The real insight: a different cost structure, not just a lower price

Put the four together, and the lesson is not “India is 30-40% cheaper.” It’s that India has a different cost structure. Relative to the gln obal template, an Indian AI build spends less on land, shell, and labour and more, proportionally, on imported compute, climate-driven cooling, and power resilience. A budget built by taking a global $/MW figure and shaving 30% off it will be wrong in a specific and dangerous way: it will over-fund the building and under-fund the cooling and power systems that actually determine whether the facility runs. The money doesn’t just shrink; it moves.

What this means for anyone building in India

You cannot copy-paste a global AI data centre design onto Indian soil and expect the economics to hold. Projects that come in on budget are re-engineered for local reality on three fronts. Liquid cooling stops being optional far sooner here, because the climate makes air cooling both costlier and less effective; the cooling investment that looks like a premium is actually the efficiency play. Power resilience has to be designed for the Indian grid from the first line, sized for a synchronous training peak on a supply that is less predictable than the benchmark assumes. And prefabricated, modular construction addresses both the speed the market demands and the shortage of specialised data centre and liquid-cooling skills by moving complexity into the factory.

Operators who treat India’s 30-40% divergence as a discount to pocket will misallocate their budgets. Those who treat it as a signal that the cost structure itself is different, and that the design must follow suit, will build facilities that are genuinely cheaper to own, not just cheaper to quote. In one of the fastest-growing data centre markets on earth, that distinction is the difference between a project that pencils and one that stalls. The global benchmark tells you what a data centre costs elsewhere. India requires you to do the harder work of costing it here.

# # #

About the Author

Sanjay Sahu is Director of Data Center Sales at Schneider Electric, Greater India, where he leads the company’s data center business across the region. With more than two decades in India’s data center and secure-power industry, spanning three-phase UPS, critical power, and end-to-end infrastructure, he advises IT/ITES, BFSI, and government organizations on designing data center solutions that are commercially and technically sound. He writes on data center power strategy and the evolving Indian market.