· 10 min read

L&T Just Deployed India's Largest AI Cluster: 10,000 GPUs for One Customer. That's What Reserved Capacity Looks Like.

Stack Lessons

On August 13, 2026, Larsen & Toubro's AI infrastructure arm switched on 10,000 NVIDIA B300 GPUs in a single NVLink-connected cluster at its Chennai campus. That's India's largest single-cluster AI deployment to date, and every one of those 10,000 chips is spoken for by one customer: Together AI. The order, secured by L&T's subsidiary LTN Compute, is reportedly worth ₹10,000 to ₹15,000 crore. Big number, one buyer, and it's worth being precise about what that combination actually means for anyone who rents compute rather than builds it.

The numbers, verified against L&T's own release

I went straight to L&T's press release instead of trusting the aggregator digest that's been circulating, and the core facts hold up. Vyoma.AI, an L&T company, through its subsidiary LTN Compute, announced the deal on August 13, 2026, describing it as "India's largest single-cluster AI infrastructure": a 10,000-GPU NVIDIA B300 AI Factory hosted at Vyoma's Chennai data centre campus in Tamil Nadu. The GPUs sit on a shared NVLink fabric as one unified cluster, not a distributed pool stitched together across racks or facilities, which is the detail that makes this deployment notable rather than just large.

The site itself is bigger than this one order. Vyoma's Chennai campus is described as gigawatt-scale, with Phase 1 built for 250 MW and power infrastructure readiness of 150 MVA, meaning this 10,000-GPU cluster is the first tenant on a facility designed to keep growing. L&T Chairman and Managing Director S N Subrahmanyan called it a milestone in the company's "Gigawatt AI Infrastructure Mission." Together AI co-founder and CEO Vipul Ved Prakash put it more bluntly in the same release: "Making AI globally accessible is going to be the biggest infrastructure build-out in human history, and L&T understands that."

On the deal size, Reuters (via Business Standard) reported the order at ₹10,000 crore to ₹15,000 crore, converting that to $1.05 billion to $1.57 billion. Do the math on their own conversion and it implies roughly ₹95 to ₹96 per US dollar as of August 13, 2026, which lines up with where the rupee has been trading this year. I'm citing their number rather than doing an independent FX conversion because the rupee-crore figure itself, not the exact exchange rate, is the load-bearing fact here. Neither L&T's press release nor Together AI's own site had published the rupee figure independently as of this writing, so treat it as reported, not confirmed by either party directly.

Why "one customer, one cluster" matters more than the GPU count

10,000 GPUs sounds impressive until you remember that GPU counts alone are a vanity metric at this point. Yotta is deploying more than 20,000 Blackwell Ultra GPUs at its Greater Noida campus. What makes the L&T deal a different animal is that this entire cluster was built to spec for a single company's roadmap, under contract, before a single external customer could rent so much as an hour of it.

That's the distinction a lot of the coverage glossed over. A shared cloud region, the kind you'd hit through an API call to any GPU marketplace, pools capacity across thousands of customers and lets utilization even out across everyone's spiky demand. A dedicated cluster like this one is the opposite: it's Together AI's own infrastructure, sized and networked for Together AI's own training, fine-tuning, and inference roadmap, sitting inside a facility L&T built and now presumably operates or leases back to them. Nobody outside Together AI is getting direct access to slices of these specific 10,000 chips. What Together AI does with that capacity, and how much of it eventually shows up as public API throughput versus internal model development, is entirely their call.

What this actually means if you rent compute from Together AI

Here's the part that matters if you're a solo operator paying Together AI (or any inference provider) by the token or by the GPU-hour: deals exactly like this one are how that provider builds the capacity that eventually becomes the thing you rent. Together AI doesn't materialize inference capacity out of nowhere. It contracts for dedicated clusters like this Chennai buildout, brings them online, uses them to serve committed enterprise customers and to run its own model work first, and only later does spare or amortized capacity trickle down into more available, more competitively priced public API throughput.

That lag is real and it's not measured in weeks. Data center buildouts of this scale typically take multiple quarters between "cluster goes live" and "meaningfully changes the public pricing or availability picture" for an end customer buying by the hour or the token. If you're expecting this specific Chennai cluster to loosen up GPU availability or push Together AI's API pricing down by Q4, don't hold your breath. What it does tell you is something else useful: Together AI has enough committed revenue and enough investor confidence to underwrite a billion-dollar-plus infrastructure order, which is a decent signal that the company isn't going anywhere and is scaling its serving capacity aggressively. That matters if you're building a product with a hard dependency on their platform.

The broader India buildout, briefly

This isn't happening in isolation. Yotta is deploying more than 20,000 liquid-cooled Blackwell Ultra GPUs at its Greater Noida data center in a project reportedly north of $2 billion, alongside plans for tens of thousands more B300 and next-generation GPUs across its Noida and Navi Mumbai campuses, and it has a multi-year NVIDIA DGX Cloud partnership reportedly worth over $1 billion. Reliance Industries is separately building a gigawatt-plus-scale data center in Jamnagar, Gujarat, with reported cost estimates ranging from $20 billion to $30 billion and at least one confirmed tenant lease from Meta for 168 MW of that campus. India is trying to become a serious AI infrastructure hub, not just a market for AI products, and this L&T-Together AI deal is one data point in that push rather than the whole story. I'm not going to pretend to have fully verified every dollar figure in the Yotta and Reliance numbers; they're softened with "reportedly" on purpose because the reporting varies by outlet and I couldn't find primary press releases confirming every figure the way I could for L&T's own announcement.

What I'd actually do

If you're renting inference or fine-tuning capacity from Together AI or a comparable provider, treat announcements like this as a leading indicator of their balance sheet and roadmap, not a leading indicator of your own bill. Don't change your budget forecast based on this news. Do factor it into your vendor risk assessment: a provider willing to underwrite a $1B+ dedicated cluster is less likely to disappear or get acqui-hired out from under you in the next 18 months, which is worth something if you've built product infrastructure on top of their API. If you actually want cheaper or more available GPU capacity soon, look at spot markets and smaller regional providers riding the wave of last year's buildouts, not brand-new clusters that just went live.

The honest counter-take: I could be wrong about the lag. If Together AI is under real competitive pressure from Fireworks, Baseten, and the hyperscalers on inference pricing, they might push discounted capacity to market faster than the "multiple quarters" pattern I'm describing, specifically to defend market share while this cluster is fresh and underutilized. Watch their public pricing page over the next two quarters rather than taking my timeline as gospel.

Author

Lukas

Sources

More Stack articles →

Stay in the Loop

Get new posts delivered to your inbox. No spam, unsubscribe anytime.

Newsletter coming soon. Set PUBLIC_CONVERTKIT_FORM_ID in .env to activate.

Related Posts