CHENNAI, 14 AUG 2026 — Larsen & Toubro has won an order worth ₹10,000–15,000 crore from the American AI cloud company Together AI. The contract is to build what L&T calls India's largest single-cluster AI infrastructure, with up to 10,000 NVIDIA B300 GPUs at a Chennai campus.
That is roughly US$1.2 to US$1.8 billion, and it is the second Together AI deployment to be announced in a week.
What was ordered
As one of India's largest engineering and construction groups, L&T is handling the build, not the operations. The customer is Together AI, the campus owner is Vyoma, and the final product will be an AI-native cloud.
One clarification, because two L&T announcements are being conflated
L&T also has a separate partnership with NVIDIA, announced at the India AI Summit under the IndiaAI Mission, for gigawatt-scale sovereign AI infrastructure — GPU clusters scaling to 30 MW in Chennai and a new 40 MW facility in Mumbai, about 70 MW in total.
Those are different numbers describing a different project, and coverage this week has mixed them. The Together AI order is at Vyoma's campus with a 250 MW first phase; the NVIDIA sovereign-infrastructure partnership is the 70 MW figure across two cities. Anyone reconciling the two will conclude one source is wrong. Neither is.
The thread worth pulling
Together AI is the name to watch here, because this is its second announced buildout in seven days.
Earlier this week IBM committed about US$240 million over multiple years for roughly 2,000 NVIDIA HGX B300 systems with Spectrum-X networking, deployed in the United States, explicitly to serve open models as an alternative to proprietary platforms. Now up to 10,000 B300s in Chennai.
Two purchases, two continents, one week, same silicon generation. A company whose pitch is serving open-weight models at scale is assembling capacity on both sides of the world at a pace that only makes sense if it has demand it can name.
This buying spree has implications beyond India. The open-model inference market has been described for two years as strategically important and commercially thin. Somebody is now buying like it is neither.
Where India actually stands
This lands into a market that has moved quickly and is not short of announcements.
| Project | Scale |
|---|---|
| Yotta, Noida | Over US$2bn for 20,736 liquid-cooled B300s, with 800Gbps InfiniBand and 40PB of storage. More than 10,000 of those GPUs committed to the IndiaAI Mission. |
| L&T for Together AI, Chennai | Up to 10,000 B300s. ₹10,000–15,000 crore. Phase 1 at 250 MW. |
| L&T with NVIDIA | Sovereign infrastructure under the IndiaAI Mission. 30 MW Chennai, 40 MW Mumbai. |
| IndiaAI Mission | State programme placing GPUs with startups and research institutions at subsidised rates. |
The projects serve different customers. Yotta's cluster has more than half its GPUs committed to a national mission — domestic startups, research institutions, sovereign model development. The L&T order is capacity built in India for an American company to sell globally.
Both are good for the Indian construction and power sector. Only one of them is sovereign AI in any meaningful sense, and the vocabulary in the press releases does not reliably distinguish them.
Check the electrical arithmetic
The two power figures in this announcement do not obviously fit together, and the discrepancy is worth examining.
The campus is described as having 150 MVA of power infrastructure readiness, supporting a Phase 1 designed for 250 MW.
Those are different units. MVA is apparent power — what the electrical infrastructure can carry. MW is real power, the part that does work. The ratio between them is the power factor, and at a realistic 0.9 for a well-corrected datacentre load, 150 MVA delivers roughly 135 MW.
Which is a little over half of 250 MW.
Most likely, the readiness figure describes what is energised now, while the 250 MW figure is the design target for the phase once further capacity is commissioned. That is an ordinary way for a large site to be built and an ordinary way for a press release to describe it.
But the two numbers are being quoted side by side as though one supports the other, and they do not. Nobody publishing them stated a power factor, and we could not establish whether the 150 MVA is the full Phase 1 provision or a first tranche. Anyone underwriting delivery on this site should ask which, because it is the difference between capacity that exists and capacity that is scheduled.
What ASEAN should take from it
The regional comparison is instructive and slightly uncomfortable.
Indonesia's Zankore announced a gigawatt target last week with a 200 MW first phase beginning construction in the first half of 2027. This Chennai order is a 250 MW first phase with a signed contract, a named customer, a booked order value and a builder that constructs refineries and metro systems for a living.
The difference is not ambition, but substance: one is a contract, the other an intention. When comparing regional AI infrastructure announcements, just ask whether anyone has actually bought anything. An order value booked by a listed contractor is a different beast from a memorandum with a target date.
The other lesson is about power, repeating a pattern we have covered four times this month already. The binding constraint here is not the 10,000 GPUs, but the electrical provision. The substations and interconnection took years to arrange and, as shown above, are the figures most worth interrogating. Read every capacity announcement for the electrical number first, then check whether it actually covers the capacity it is quoted next to.
What to be sceptical about
The order value is a range, ₹10,000 to ₹15,000 crore, which is a fifty per cent spread and unusual for a contract described as booked. Ranges of that width usually indicate phased scope with options attached rather than a single signed number.
"Up to 10,000 GPUs" carries the same qualifier. Up to is not the same as ten thousand, and B300 supply is contested globally — the constraint we described when Foxconn named CoWoS packaging as the 2027 ceiling applies to everyone buying this generation, including this campus.
No completion date has been published, which for a 250 MW first phase is the single most useful missing fact.
What to watch
Whether Together AI announces a third. Two buildouts in a week suggests one of two things, with very different implications: either a round of capital is being deployed, or the company has sold more inference than it can serve.
Whether Chennai's power holds. A 250 MW phase on a gigawatt-scale site is a serious draw on Tamil Nadu's grid, and India's data centre boom is increasingly a discussion about generation rather than construction.
And whether any of the sovereign framing survives contact with the customer list. Capacity built in India, owned by an Indian campus operator, filled with American demand is a real economic win and a weaker sovereignty story than the surrounding language implies.
View author profile → · Editorial policy
