India AI Data Centres: We Lost the Race, So We Rent the Land

Key takeaways

• HCLTech is putting Rs 14,257 crore into an AI data centre and technology hub in Odisha against 5,000 declared jobs. That is about Rs 2.85 crore of capital per job.
• PLI has disbursed Rs 35,354 crore for about 14.15 lakh jobs, roughly Rs 2.5 lakh each. A data centre costs about 114 times more capital per job.
• TCS has taken 200 acres at Anakapalli and 88 acres at Bopkhel in Pune for about Rs 641 crore, to host OpenAI capacity. Neither firm is building a frontier model.
• The IndiaAI Mission was approved with Rs 10,372 crore over five years. About Rs 401 crore had been released by 9 February 2026.
• India is not entering the AI race. It is renting land, power and cooling to the people already in it.

TCS buys land for OpenAI and the headline says India’s AI moment. HCLTech signs Odisha and the headline says sovereign AI. Google commits a gigawatt near Visakhapatnam and the headline says the world has chosen India.

Nobody on that list is building a model. They are building the building. It is the same decision that produced the headcount reckoning inside Indian IT, taken one layer down the stack.

The mechanism, step by step

Only step two touches employment, and only while the site is being built. It looks like this because rent is contractible and capability is not. For it to change, step four has to move: equity in the model, or guaranteed Indian compute, written into the lease.

The Rs 2.85 crore per job problem

Two honest caveats. PLI counts indirect jobs, and the 5,000 figure is mostly the Bhubaneswar delivery centre, not the data centre. A live data centre that size runs on a few hundred people. Both caveats make the real gap wider, not narrower. The full factory-side arithmetic is in the PLI reality check.

We are funding the building faster than the model

Rs 4,563 crore of that outlay was earmarked for compute. Rs 401 crore has moved in two years, split Rs 22 crore in FY25 and Rs 379 crore in FY26, with nothing yet against the FY27 estimate of Rs 1,000 crore. One private company committed thirty-five times the national two-year release in a single signing. Private capital moves at the speed of a rent cheque. Public capital for the model layer moves at the speed of a file.

We have run this film before

The 1990s built a genuinely large industry and never produced an Indian software product of global scale. This version rhymes with a heavier asset. Laptops get redeployed; thirty-year power contracts argue for continuity. Escaping upstream means owning silicon, and that fight is described in how the Nvidia moat is closing.

Keep the announcements honest

India added 258 MW in the first half of 2026 to reach about 1.8 GW live. The pipeline is 8.33 GW. That gap closes only if power connections and land clearances speed up, which is a state-government problem, not a demand problem. Note also where the money prefers to sit: buildings, not companies. The same preference runs through the why capital repriced Indian risk.

The counter-case, stated properly

• Sovereign compute. Guaranteed capacity for Indian researchers at a controlled price buys a seat the country cannot otherwise afford.
• Power economics. A gigawatt of committed baseload is the best anchor customer a grid can have, if the tariff is negotiated rather than given away.
• Model equity. HCLTech holding a stake alongside Sarvam is a different deal from charging Sarvam rent.
• Upstream manufacturing. Racks, cooling, transformers, eventually chips. A gigawatt of demand is a manufacturing brief if someone writes it.

All four are available. None happens automatically, and none of the agreements is public.

What I think happens next

• No Indian frontier model by 2030. Sarvam becomes excellent at Indian languages, which is a real business and a different claim.
• The 8.33 GW pipeline converts to nearer 3 GW live by 2030. Quiet re-phasings, not cancellations.
• Announced job counts keep bundling construction and delivery centres, because operating headcount alone is not a headline.
• The power tariff is the fight worth watching, not the ribbon-cutting.

What to do about it

For a business, compute gets cheaper and closer, and that is a real gift whether or not we own the model. If you are building, the useful move is to buy capacity in the Indian regions coming online instead of paying to move data across an ocean, and to keep the model layer above you swappable, because the thing being localised here is the rack, not the intelligence. If you are an engineer, this wave funds electrical, thermal, high-voltage and power-systems work far more than it funds another web framework, and in 2026 a data centre operations qualification is the cheaper bet. And when the next signing is announced, do the division yourself before you react: capital committed, divided by jobs promised. Rs 2.85 crore a job is not a scandal. It is simply not an employment programme, and knowing which one you are looking at is the whole skill.

Read next: the TCS layoffs case study for what is happening to the services layer these firms are pivoting away from, and what India got for Rs 35,354 crore of subsidy for the factory route, where a job costs 114 times less capital.

Sources

• HCLTech and Government of Odisha MoU, July 2026
• Ministry of Commerce and Industry, PLI progress to 31 March 2026
• MeitY, IndiaAI Mission reply to Parliament, February 2026
• Reported TCS land acquisitions, Anakapalli and Bopkhel
• Knight Frank India data centre report, 2026

Figures are as reported by the sources named above at the time of writing.