Why Together AI and L&T Are Building India's First 10,000-GPU Blackwell Unified Cluster

Engineering giant Larsen & Toubro has secured a massive contract worth between ₹10,000 crore and ₹15,000 crore (approximately $1.05 billion to $1.57 billion) to build India's first unified 10,000-GPU supercomputer for Together AI. Located at the Vyoma.AI data center campus in Chennai, the facility will house 10,000 NVIDIA B300 "Blackwell Ultra" GPUs. But the real story isn't the price tag—it is the choice of a unified single-cluster architecture over a traditional distributed network.

The Architecture of the Together AI Chennai Unified Cluster

To understand why Together AI is investing billions in Chennai, one must understand the difference between 10,000 networked GPUs and a unified 10,000-GPU cluster. In a standard distributed deployment, GPUs are housed in independent servers connected via standard ethernet or InfiniBand networks. While suitable for running inference or training smaller, parallelizable models, this approach falls apart when scaling frontier models.

The Together AI Chennai unified cluster avoids these bottlenecks by linking all 10,000 GPUs using NVIDIA’s high-bandwidth NVLink 5 fabric. NVLink 5 provides a staggering 1.8 terabytes per second of bidirectional bandwidth per GPU. Rather than functioning as a collection of individual nodes, the entire cluster acts as one massive, addressable machine.

Bypassing the Gradient Synchronization Bottleneck

During large-scale model training, systems must perform a mathematical operation called "all-reduce" to synchronize gradients across all participating nodes. In a disaggregated system, this synchronization is choked by network latency and bandwidth limits, leaving expensive GPUs sitting idle—a phenomenon known as the communication bottleneck.

With a unified single-cluster configuration, the NVLink network allows for direct memory-to-memory communication. The performance advantages of the NVIDIA B300 hardware make this scale possible:

• Memory capacity: Each B300 GPU features 288 GB of high-bandwidth memory (HBM3e).
• Memory bandwidth: The chips deliver 8 TB/s of local memory bandwidth.
• Server-level scale: An eight-GPU node provides 2.3 TB of fast GPU memory, capable of holding a 405-billion parameter model in 16-bit precision on a single server.

"A distributed deployment of 10,000 GPUs spread across independently networked servers and a unified single-cluster of 10,000 GPUs connected by a shared NVLink fabric are not the same machine. They are not even close to the same machine."

Ultrathink Analysis

Why Together AI Is Betting on Localized Sovereign Compute

For Together AI, which recently crossed $1.15 billion in bookings, the Chennai deployment represents a calculated bet on local, high-performance infrastructure. Instead of renting fragmented cloud compute from hyperscalers like AWS or Microsoft Azure, Together AI is building a localized physical hub for AI training in the Asia-Pacific region.

India is rapidly becoming a hotspot for AI development, but training foundation models requires massive local computing clusters to comply with data sovereignty preferences and reduce latency. By partnering with Larsen & Toubro—a conglomerate with deep infrastructure expertise—and Vyoma.AI, Together AI secures a premium physical footprint with the dedicated power and cooling required to run 10,000 Blackwell Ultra GPUs at peak capacity.

The Future of Enterprise AI Infrastructure

The L&T-Together AI deal signals a broader shift in how modern tech companies view AI infrastructure. The era of buying virtualized, general-purpose cloud instances for training is giving way to dedicated, physical AI supercomputers. For builders and researchers, this unified design means multi-node fine-tuning and massive pre-training runs can scale smoothly without hitting performance cliffs.

As startups and enterprises push past the limits of 100-billion-parameter models, the physical interconnect architecture of the data center will dictate the limits of AI capability. Software can only be as fast as the copper and silicon supporting it.