L&T Launches Massive 10,000-GPU AI Factory in Chennai Using Nvidia B300 Hardware

Per CRN Asia, Larsen & Toubro just parked 10,000 Nvidia B300 GPUs into its Vyoma Chennai campus — the engineering giant's first move into the AI Factory market, and what L&T frames as India's largest single-cluster B300 deployment. The hardware lands with US AI cloud outfit Together AI through L&T's LTN Compute subsidiary, hitting training, fine-tuning, and inference workloads. I want to see whether the SLO math holds, or whether this is another floor-loaded press-release cluster.

What the box actually contains

The Chennai facility stacks 10K B300 GPUs with high-performance networking, parallel storage, and ultra-low-latency interconnects. Phase 1 of the campus lands at 250 MW with 150 MVA power infrastructure ready, and L&T is framing the site as gigawatt-scale with room to expand. The capacity number is the only spec I trust without a third-party audit. Everything else rides on which interconnect and which storage tier they actually shipped.

Together AI runs a multi-tenant training and inference cloud. A 10K-GPU single-tenant cluster means one team's checkpoint — and one team's OOM — can stall the fabric. I've watched a 2,000-GPU job crater a shared fabric because a single all-reduce exceeded timeout. At 10K scale that risk multiplies. Watch for any NCCL or network partition incidents in the first 90 days.

The reliability lens

AI Factories only pay off when utilization stays north of 70%. Below that, you're paying for power, cooling, and depreciation while the GPUs idle. The L&T pitch that customers can "deploy and scale AI workloads through a unified infrastructure stack" tells me nothing about tail latency or failure-domain isolation. A unified stack on top of 10K GPUs is great until a storage backend hiccups and 10,000 jobs stall simultaneously.

L&T chairman and managing director S.N. Subrahmanyan tied the build to a broader Gigawatt AI Infrastructure Mission. Together AI CEO Vipul Ved Prakash framed L&T's involvement as part of "the biggest infrastructure build-out in human history." Historical claims don't ship tokens. Show me MTTR, cross-rack bandwidth, and storage IOPS per GPU before I take the pitch.

Two outlets — Moomoo and Intellectia AI — filed separate reports this week on Nvidia's $500 billion compute financing play, per their headlines. Not part of the L&T deal, but worth a read if you're financing your own GPU cluster, since capex structures change when the chip vendor is bankrolling the buyer.

What I'm watching

Three signals will tell me whether 10K-B300 is real capacity or marketing: power utilization telemetry once Phase 1 hits steady state, public SLOs from Together AI on training throughput and inference latency, and MTBF on the fabric — a single switch failure at this scale can black-hole thousands of jobs. Until those numbers land, the announcement is a promise, not infrastructure. Deploy or discard applies to workloads, not press releases. Nobody should retrain a frontier model on the back of a launch event.