L&T Vyoma — KancheepuramLarsen & Toubro Vyoma (LTN Compute) · Sriperumbudur, Kancheepuram district · 30 MW claimed

L&T is the largest engineering contractor in India and it is now an operator. Its data centre business, renamed Larsen & Toubro Vyoma, has a campus at Kancheepuram that has been published at 90 MW, then 30 MW, then 24 MW, then 18 MW, then 12 MW — and separately described as a 250 MW Phase 1 on 150 MVA of live power. In August 2026 it took a ₹10,000 to ₹15,000 crore order from the American AI cloud company Together AI to host 10,000 Nvidia B300 GPUs. Five weeks earlier, the Tamil Nadu government announced a ₹15,000 crore data centre expansion by L&T at the same campus. Both were reported as news. They are almost certainly one number.

1At a glance

2Seven capacities for one campus

3The 12 plus 6 does not equal 30

In May 2026 the company reported that its data centre business had commissioned 12 MW at Kanchipuram, that a further 6 MW was at an advanced stage of commissioning, and that this brought total capacity available to 30 MW. Twelve plus six is eighteen.

Three explanations are consistent with the public record and Gridvik cannot choose between them:

Gridvik uses 18 MW and states the conflict rather than resolving it. A vendor planning against 30 MW at this site should confirm energisation directly.

4Two ₹15,000 crore stories

This is the most consequential thing on the page, and it is a reporting observation rather than an accusation. Two announcements, five weeks apart, both carrying the number ₹15,000 crore, both about the same campus.

It is unlikely to be a coincidence. The most economical reading is that the Together AI order is the anchor contract that funds the ₹15,000 crore expansion the state announced in June, and that the two were reported as separate pieces of news because they were announced by different parties in different cities.

Why this matters commercially: if both numbers are real and they are the same money, the Indian AI infrastructure order book is being overcounted by ₹15,000 crore (about $1.7 billion) in most summaries of it. If they are genuinely two separate commitments, that is a materially larger and much more interesting fact. No public document reconciles them, and Gridvik will not assert which it is.

5What 10,000 B300 GPUs need

This can be checked arithmetically, and it validates the reported numbers in a way that most capacity claims in this market cannot.

The arithmetic closes. 23 MW of AI factory plus 6 MW of colocation is 29 MW, which is the 30 MW L&T claims, and 10,000 B300s at DGX densities need about 18 MW of IT load, which leaves about a quarter for networking and storage. Every number here is internally consistent, and that is unusual enough in this market to be worth saying out loud.

It also explains the "12 plus 6 equals 30" puzzle. The 30 MW is a campus capacity figure covering both the AI conversion and the residual colocation, not the sum of two commissioned phases.

6The contractor becomes the operator

L&T builds data centres. It has been the engineering, procurement and construction contractor on hyperscale projects for years. In 2026 it turned the same capability inward, and did it in three moves inside eight months.

The incentive to watch is competitive as well as commercial. L&T is a contractor to operators including some of the companies in this registry. Every move it makes as an operator is a move its clients will notice — and its pricing as a contractor now competes with its own hosting business.

7Cost per megawatt does not fit

One number on this page does not agree with the rest of the market, and it is worth flagging because it changes equipment budgets.

Two explanations fit. Either the ₹2,000 crore figure covers more than the first 12 MW — a campus, land, a substation sized for far more, or later phases bundled into one number — or L&T built a genuinely more expensive specification, which is plausible for an AI-ready campus with high floor loading.

Gridvik records the discrepancy rather than picking one. The practical consequence is real either way: anyone using ₹167 crore per MW as a benchmark for Indian build costs will over-budget by a factor of three.

8Timeline

• 24 November 2021
  L&T signs a memorandum with the Tamil Nadu government to build a data centre at Kanchipuram, described as up to 90 MW in phases over five years.
• December 2024
  A ₹2,000 crore, 12 MW facility at Sriperumbudur nears completion. Testing under way. Plan stated as 30 MW in the near term and 90 MW over five years.
• May 2026
  Data centre business rebranded Larsen & Toubro Vyoma. 12 MW commissioned, 6 MW at an advanced stage of commissioning, total capacity claimed at 30 MW.
• 4 June 2026
  Tamil Nadu signs a ₹18,600 crore three-project memorandum with L&T, including a ₹15,000 crore hyperscale and edge AI data centre expansion in Kancheepuram. The first large industrial memorandum of the new state government.
• 11 August 2026
  L&T agrees to transfer its data centre and cloud services business to Vyoma.AI for ₹1,400 crore.
• 13 August 2026
  LTN Compute wins a mega order from Together AI: 10,000 Nvidia B300 GPUs in a single cluster at the Chennai campus, classified at ₹10,000 to ₹15,000 crore. Phase 1 design of 250 MW with 150 MVA of power ready is disclosed for the first time.
• 22 October 2026
  18 MW commissioned or in commissioning. Around 24 MW being converted to AI workloads. The AI factory is not live

9Procurement

Rows with a lock are the paid layer.

10Assumptions and method

11Limits and open questions

12Sources

This record was compiled entirely from publicly available sources, cross-checked against one another.