Entrepreneur News Network

Nominate Now for the Global Women Power List 2026

Entrepreneurnewsnetwork Logo
Linkedin Envelope
Menu
Menu
Menu
India data center

India Doesn’t Have a Data Center Problem. It Has a Jurisdiction Opportunity.

by

“India has 20% of the world’s internet users. It holds 1.5% of global data center capacity. The chips inside cost more than the building and the land combined. Even the largest tech companies lease more than they own, because owning sinks returns. Countries think differently. They need jurisdiction, not hardware. So what happens to a country that has the users but not the machines?”

— Nikhil Kamath, LinkedIn

That closing question is worth taking seriously, because the honest answer is more reassuring than the framing suggests. Roughly one in five people who go online anywhere in the world do so from India. The country’s active internet base crossed 886 million in 2024 and was projected to pass 900 million in 2025, according to the IAMAI-Kantar Internet in India report — against a global internet population of roughly 5.5 to 6 billion people tracked by DataReportal’s Digital 2026 estimates. That puts India’s share of the world’s internet users somewhere in the 15-20% range depending on methodology, and separate industry analysis suggests India generates close to a fifth of the world’s data traffic outright.

Now set that against how much of the world’s data center capacity actually sits inside India’s borders: somewhere between 3% and 4% today, according to 2026 estimates from CBRE, JLL and KPMG’s India Data Centre Opportunity report — up from a smaller base a few years ago, and projected to reach only about 5% of global capacity even by 2030. Some earlier industry estimates, using narrower definitions that count only fully operational (as opposed to under-construction) capacity, have put India’s live share closer to 1.5-2%. Whichever number you use, the gap is the same story: a country that produces close to a fifth of the world’s internet activity hosts a low single-digit share of the physical infrastructure that runs it.

The instinctive reaction is that India — or any country in a similar position, from Indonesia to Nigeria to Brazil — is “behind” and needs to close the gap by building and owning more steel, concrete and racks. That instinct is largely wrong, and the reason why says a lot about how the economics of computing have quietly inverted over the last three years.

The building is now the cheap part

For most of the data center industry’s history, the physical facility was the dominant cost. Land, power infrastructure, cooling and shell construction were the capital-intensive core; the servers inside were a recurring, relatively modest line item that got refreshed every few years.

AI has flipped that ratio. A detailed cost breakdown of AI training infrastructure — tracing a dollar of spend from land to silicon — estimates that building a 100-megawatt data center facility costs roughly $0.9 billion to $1.5 billion for construction, electrical systems and mechanical/cooling infrastructure. Within that construction budget, land and the building shell together account for only about a quarter of the spend; electrical systems eat up roughly half, and cooling most of the rest. But populate that same 100 MW facility with 100,000 Blackwell-class GPUs at Nvidia’s prevailing $25,000-$40,000 per-unit pricing, and the hardware bill alone runs to $2.5 billion to $4 billion — two to three times the cost of the entire building that houses it, and roughly seven to ten times the cost of the land and shell specifically.

Nvidia’s own bill-of-materials data makes clear why the chips are so expensive: on a single B200 GPU, high-bandwidth memory accounts for nearly half the manufacturing cost, with advanced CoWoS packaging and logic silicon splitting most of the rest — a roughly $6,400 manufacturing cost that Nvidia sells for $30,000-$40,000, reflecting the pricing power of a company that controls an estimated 92% of the data center GPU market. Scaled up, a single Nvidia GB200 NVL72 rack now runs $2 million to $3 million on its own — often $3.9 million once networking and storage are included — a price tag that would have bought a small standalone building a decade ago.

Put simply: in the AI era, the chips are the expensive, scarce, fast-depreciating asset. The building is comparatively cheap, generic, and — critically — nearly identical whether it sits in Virginia, Frankfurt, Mumbai or Johannesburg.

Why even the richest tech companies would rather rent the shell

If the building is the cheap, commoditized part of the stack, it follows that owning it outright is not where the returns are — and the world’s largest technology companies are behaving exactly as if they believe that.

Data center lease commitments among U.S. tech companies passed $850 billion in the first quarter of 2026 alone, a 63% jump from a year earlier, with Microsoft and Meta each committing close to $50 billion in additional leases in a single quarter. Even Amazon, which is often held up as the most vertically integrated of the hyperscalers, owns roughly 24 million square feet of data center space and leases a comparable amount — a deliberately hybrid model rather than a pure ownership strategy.

Where companies do build and “own,” the ownership is increasingly a legal fiction designed to keep debt off the parent company’s balance sheet rather than a straightforward capital commitment. Meta’s roughly $30 billion Hyperion campus in Louisiana was financed through a joint-venture special purpose vehicle with Blue Owl Capital, in which Blue Owl-managed funds hold 80% of the equity and Meta retains just 20% — while Meta locks in long-term capacity through an operating lease and offtake commitments. Structures like this let the hyperscaler secure the compute it needs without the facility’s construction debt showing up on its own books, at the cost of a minority equity stake and higher effective financing costs. Industry analysts increasingly frame this as the emerging default architecture for AI infrastructure buildouts, not an exception.

None of this means hyperscalers are indifferent to ownership altogether — some forecasts suggest hyperscalers will still own roughly two-thirds of overall data center capacity by 2031, and Meta in particular is pursuing a dual-track strategy of owning AI-specific campuses while continuing to lease general cloud capacity. But that ownership is concentrated precisely where it is strategically differentiating — custom AI campuses tied to proprietary model training — while the generic, replaceable shell keeps getting financed off-balance-sheet or leased outright. The chips and the software running on them are the asset worth owning. The building is a cost to be minimized.

What a country actually needs is legal reach, not concrete

This is where the calculus for a country like India differs sharply from the calculus for a hyperscaler — and where the “build more capacity” framing misses the point. A company optimizes for return on capital; a government optimizes for something closer to control: the ability to compel a platform operating on its citizens’ data to answer to its courts, its regulators and its laws, regardless of whose balance sheet the servers sit on.

That is precisely the model countries have started converging on, and it doesn’t require owning racks. Microsoft’s EU Data Boundary commitment, for instance, guarantees that AI data processed by its services — including prompts, embeddings and inference logs from Microsoft 365 Copilot — stays within the EU/EFTA region and is processed in-country by default, extending well beyond simple data residency into the AI pipeline itself. France’s Mistral has built its sovereign cloud explicitly around the promise that no data touching it is subject to the U.S. CLOUD Act, so that banks, hospitals and government agencies bound by GDPR have a jurisdictionally clean option. The UAE and France went further still in 2026, structuring a bilateral AI cooperation framework around what officials describe as “virtual data embassies” — infrastructure that is physically hosted in one jurisdiction but operates under agreed legal protections for the other. Saudi Arabia’s HUMAIN initiative and the UAE’s Stargate buildout, worth roughly $100 billion and targeting a gigawatt of capacity respectively, follow a similar template that analysts have summarized as a recurring playbook across the EU, Canada, the UAE and Saudi Arabia: a national fund, a cloud or telecom partner, a GPU allocation, and — the actual point of the exercise — a data sovereignty clause.

India, in fact, is already running a live experiment in exactly this model, even if it isn’t usually described that way. The IndiaAI Mission has empaneled more than 38,000 GPUs — including Nvidia H100s and H200s — for subsidized access by Indian startups, researchers and institutions, with a target of 100,000 GPUs by the end of 2026. Critically, the government does not own a single one of those chips. Private partners — Jio, Tata, Yotta and CtrlS among them — own and operate the hardware; the state’s role is to empanel providers, subsidize the per-hour access rate to roughly ₹65-116 depending on the subsidy tier, and set the terms under which that compute can be used. The Digital Personal Data Protection Act layers a second mechanism on top, giving the government the authority to designate categories of sensitive data that must be processed domestically regardless of who owns the infrastructure processing it. India isn’t trying to own the means of computation. It’s asserting jurisdiction over how computation touching Indian data and Indian users has to behave — which is a fundamentally cheaper, faster and more durable form of leverage than trying to out-build hyperscalers at their own capital-intensive game.

The better scorecard

None of this means India’s physical data center buildout doesn’t matter — latency, energy security, and the ability to attract compute-hungry AI workloads all benefit from more capacity within the country’s borders, which is why KPMG and others project India’s share climbing toward 5% of global capacity by 2030 on the back of roughly $30 billion in announced investment. But treating gigawatts owned as the primary measure of digital power is fighting the last war. The GPU dollar shows where the real, scarce, appreciating value in this stack actually sits — increasingly in Taiwan’s advanced packaging lines, South Korea’s memory fabs and Nvidia’s design studios, not in anyone’s data center shell — while the shell itself is being leased, off-balance-sheeted and commoditized by the very companies that build it.

The right question for India, or any large digital economy without a chip industry of its own, isn’t “how do we own more of the hardware.” It’s “how much legal reach do we have over what happens to our users’ data and compute, no matter which company’s balance sheet the servers sit on.” By that measure — not gigawatts, but jurisdiction — the country with 20% of the world’s internet users has considerably more leverage than a 1.5-4% capacity share suggests.

Sources

Leave a Comment