Larsen & Toubro has secured India’s largest single-cluster AI infrastructure order to build a 10,000-GPU NVIDIA B300 AI Factory at its Chennai data centre campus for US-based Together AI. The contract, won by L&T’s AI infrastructure arm Vyoma.AI through its subsidiary LTN Compute, is classified as a mega order valued between ₹10,000 crore and ₹15,000 crore. It marks L&T’s formal entry into the AI Factory business and positions Chennai at the centre of India’s next generation compute expansion.
What Is an AI Factory?
An AI Factory is not just a data centre with servers. It is an integrated, end to end infrastructure stack designed to build, train, fine tune and run artificial intelligence models at scale. While a traditional data centre provides space, power and cooling for general computing, an AI Factory combines accelerated computing with high performance networking, ultra low latency interconnects, high throughput parallel storage and AI operations software in one unified system.
Think of it as a factory floor for intelligence. Raw data goes in, models are trained and optimised, and ready to use AI services come out for customers to deploy instantly. In this model, the Graphics Processing Unit (GPU), a specialised chip that can handle thousands of calculations in parallel, is the core machine tool. The NVIDIA B300, part of NVIDIA’s Blackwell Ultra generation, is among the most advanced GPUs available today for training large language models and for high speed inference, which means running those models to generate answers for users.
Inside the Mega Order: Value, Scale and Site Details
The announcement was made on 13 August 2026 from Mumbai. L&T said the order is classified as mega, a term L&T uses for contracts valued between ₹10,000 crore and ₹15,000 crore, or about $1.05 billion to $1.57 billion. The company has not disclosed the exact value within that band.
The scale is what makes the project stand out. The facility will host 10,000 NVIDIA B300 GPUs in a single cluster, which the companies describe as India’s largest single-cluster AI infrastructure to date. A single cluster means all 10,000 GPUs are interconnected as one unified system, which delivers much higher performance for large model training and inference than spreading the same number of GPUs across multiple smaller sites.
Vyoma.AI, LTN Compute and the Chennai Campus
The order was secured by Vyoma.AI, L&T’s sovereign, secure and integrated AI cloud and hyperscale data centre business, through its AI infrastructure subsidiary LTN Compute, formally LTN Compute Private Limited (LTCPL). Vyoma offers sovereign cloud platforms, GPU-as-a-Service (GPUaaS), hyperscale colocation and mission critical digital infrastructure for government, Banking, Financial Services and Insurance (BFSI), healthcare, manufacturing and other high compute industries.
The AI Factory will be hosted at Vyoma’s Chennai data centre campus at Kanchipuram, near Sriperumbudur in Tamil Nadu. This campus is planned as a gigawatt scale AI infrastructure site, with Phase 1 designed for 250 MW of IT load and power infrastructure readiness of 150 MVA. The campus was developed under a Memorandum of Understanding (MoU) signed with the Government of Tamil Nadu in November 2021 to build up to 90 MW in phases, with an initial ₹2,000 crore investment for a 12 MW facility scalable to 30 MW. L&T has more recently outlined a broader Gigawatt AI Infrastructure Mission, which includes a 30 MW GPU cluster on its 300 acre Chennai campus and a new 40 MW AI ready data centre in Mumbai. In early August 2026, L&T also transferred its data centre and cloud services business to Vyoma.AI Ltd for ₹1,400 crore to consolidate this vertical.
The integrated platform will combine hyperscale data centre infrastructure with accelerated computing, high performance networking, ultra low latency interconnects, parallel storage and AI operations, allowing customers to deploy and scale workloads through a single stack. L&T Chairman and Managing Director S N Subrahmanyan said the deployment marks a significant milestone in the company’s Gigawatt AI Infrastructure Mission and reinforces its commitment to making India a global hub for next generation AI infrastructure.
Understanding the NVIDIA B300 and Blackwell Ultra Platform
The NVIDIA B300 is a data centre accelerator of the Blackwell Ultra architecture, the successor to the Blackwell generation that included the B200. It was launched in September 2025 and began shipping through DGX B300 and HGX B300 server platforms in early 2026. While the B200 was built with 192 GB of HBM3e memory, the B300 increases that to 288 GB of HBM3e per GPU using 12 high HBM stacks, with memory bandwidth of about 8 TB per second.
HBM stands for High Bandwidth Memory, a stacked memory technology that sits close to the GPU chip to feed data at very high speed. More HBM means a single GPU can hold larger models and larger context windows without splitting the model across multiple chips. This is critical because modern large language models must store not only model weights but also a KV cache, which is the temporary memory that holds context during extended conversations and reasoning steps.
The B300 also brings a generational jump in compute and connectivity. Key features include fifth generation Tensor Cores for AI calculations, a second generation Transformer Engine with native support for low precision formats such as FP4 and FP8 that reduce inference cost, and fifth generation NVLink with 1.8 TB per second bidirectional bandwidth. In rack scale systems such as the GB300 NVL72, up to 72 B300 GPUs and 36 Grace CPUs are connected in a single coherency domain, allowing trillion parameter models to be trained and served efficiently.
In practical terms, the B300 delivers about 15 petaFLOPS of dense FP4 compute per GPU, compared with about 9 petaFLOPS for the B200, and requires direct liquid cooling with a thermal design power of up to 1,400 watts per GPU. This density means data centres need specialised rack designs, power and cooling, which is exactly the hyperscale infrastructure L&T is building in Chennai.
| Feature | NVIDIA B200 | NVIDIA B300 |
|---|---|---|
| Architecture | Blackwell | Blackwell Ultra |
| Memory per GPU | 192 GB HBM3e | 288 GB HBM3e |
| Memory Bandwidth | About 8 TB per second | About 8 TB per second |
| Dense FP4 Compute | About 9 petaFLOPS | About 15 petaFLOPS |
| Interconnect | NVLink 5, 1.8 TB per second | NVLink 5, 1.8 TB per second |
| Power (TDP) | Up to 1,000 watts | Up to 1,400 watts |
PetaFLOPS here means a measure of how many AI calculations a chip can perform per second, and FP4 is a low precision number format that speeds up inference while maintaining quality for optimised models.
Who Are the Players: L&T, Vyoma.AI and Together AI
Larsen & Toubro (L&T) is India’s largest engineering and construction conglomerate, founded in 1938 in Bombay by two Danish engineers, Henning Holck-Larsen and Soren Kristian Toubro, and legally incorporated on 7 February 1946. Headquartered at L&T House, Ballard Estate, Mumbai, the group operates across engineering, construction, manufacturing, technology and financial services in more than 50 countries. It is listed on the National Stock Exchange (NSE: LT) and Bombay Stock Exchange (BSE: 500510) and is a constituent of the Sensex and Nifty 50. The conglomerate has steadily expanded into digital infrastructure through L&T Cloudfiniti, which was rebranded as Vyoma.AI, with dedicated subsidiaries such as LTN Compute to manage AI ready hyperscale data centres, sovereign cloud, AI Factory services, GPUaaS and managed AI platforms.
Vyoma.AI is positioned as L&T’s sovereign, secure and integrated AI cloud business. Sovereign cloud means that data, compute and operations remain within India’s jurisdiction, which helps meet requirements under India’s data protection and localisation norms, including for government and BFSI clients. Vyoma currently operates live facilities including a 2 MW site in Navi Mumbai and a 30 MW site in Chennai, with additional capacity under development, and targets building over 200 MW to 350 MW by 2030 under its long term plan.
Together AI (Together Computer Inc) is a San Francisco based AI Native Cloud platform founded in 2022 by Vipul Ved Prakash, Ce Zhang, Chris Re and Percy Liang, with Tri Dao also associated as a co-founder of its research lineage. The company provides accelerated infrastructure, open foundation models and developer services that let organisations train and deploy generative AI applications at scale. It operates on an open source model ecosystem, supporting deployments of models such as Llama, DeepSeek, Nemotron, MiniMax and Kimi, and serves more than a million developers. Backed by investors including NVIDIA, General Catalyst, Salesforce Ventures and Lux Capital, Together AI raised $305 million in Series B in February 2025 at a $3.3 billion valuation and $800 million in Series C in July 2026 at an $8.3 billion valuation led by Aramco Ventures. Co-founder and Chief Executive Officer Vipul Ved Prakash, who previously founded Topsy which was acquired by Apple in 2013, said the partnership with L&T brings scale, resilience and engineering excellence to India at a time when global AI access requires the largest infrastructure build out in history.
Why This Deal Matters for India
This contract is significant well beyond its headline value. It connects three national priorities, domestic high end compute capacity, sovereign data control and global competitiveness in AI services.
Alignment with the IndiaAI Mission and Sovereign Infrastructure
The Government of India’s IndiaAI Mission, backed by more than ₹10,000 crore, about $1 billion, in public funding, aims to democratise access to compute, build sovereign datasets, support indigenous foundation models and provide startup risk capital. Under its compute pillar, more than 38,000 GPUs have already been allocated, with additional capacity planned. India’s data centre capacity has grown more than four fold in six years to about 1,575 MW, and industry capacity is expected to reach about 1.8 GW by end of 2027, of which Mumbai and Chennai together account for about 64 percent of new additions.
L&T’s AI Factory directly supports this mission. In February 2026, L&T, Yotta and E2E Networks were highlighted as partners working with NVIDIA to expand AI infrastructure under the mission. Yotta’s Shakti Cloud with more than 20,000 Blackwell Ultra GPUs, L&T’s gigawatt scale Chennai campus and E2E Networks’ HGX B200 cluster hosted at the L&T Vyoma Chennai site were presented at the AI Impact Summit in New Delhi. The Digital Personal Data Protection Act, 2023 and the sovereign cloud policy also push enterprises, especially in BFSI and government, to keep personal and sensitive data on India based infrastructure. A 10,000 GPU sovereign factory in Chennai gives India based training and inference that meets those residency needs while remaining connected to global AI supply chains.
Chennai as India’s AI and Cable Hub
The choice of Chennai and Kanchipuram is strategic. Chennai is already India’s second largest data centre market after Mumbai, with about 113 MW of colocation capacity and direct access to multiple submarine cable landing stations on the east coast. This makes it ideal for low latency connectivity to Southeast Asia and global networks. The state of Tamil Nadu has been active in AI infrastructure, including its MoU with Sarvam AI for a sovereign AI park in Chennai and its early support for L&T’s campus through assured power and infrastructure assistance.
For Together AI, the deal diversifies its geographic footprint. The company has to date relied largely on North American and European capacity through partners such as Rumble, Hypertec, 5C, CoreWeave and IBM Cloud, including a $240 million HGX B300 deployment with IBM Cloud. Adding a large India base gives it resilience, access to India’s engineering talent and a platform to serve global customers seeking cost effective, open source model inference outside the hyperscaler dominated US market.
For L&T, the move is a shift from one time engineering and construction revenue to recurring digital utility revenue. The company has outlined an ambition to invest about ₹25,000 crore over five years to build 350 MW of AI ready capacity by 2030, targeting $1 billion in annual revenue from this vertical. Building data centres in house, the company says, can reduce capital expenditure by about 15 percent due to its engineering and procurement strengths.
The Way Forward
The 10,000 GPU factory will power Together AI’s AI Native Cloud platform for large scale inference, fine tuning and training of open models. Together AI has stated that its bookings crossed $1.15 billion as utilisation of open source models accelerated, and its customer base includes startups building AI native applications. The Chennai cluster will allow those customers to deploy workloads through a unified stack without managing hardware, networking or storage separately.
For L&T, Phase 1 readiness of 250 MW and 150 MVA provides headroom to scale the site toward gigawatt capacity as demand grows. The company’s broader roadmap includes expanding Chennai, commissioning the new Mumbai site and offering GPUaaS and managed AI platforms to governments, enterprises, cloud providers and AI innovators. NVIDIA’s role remains central, as it supplies GPUs, CPUs, networking, accelerated storage platforms and the AI Enterprise software stack, alongside reference architectures to speed deployment.
Industry observers will watch three markers next. First, the timeline for commissioning and whether the cluster becomes available for production workloads in early 2027 as Blackwell Ultra supply ramps. Second, how pricing for India hosted B300 inference compares with global cloud rates, and whether sovereign advantages translate into adoption by Indian enterprises and government. Third, whether this mega factory catalyses further large cluster investments in Chennai, Hyderabad and Mumbai, helping India move from being a consumer of global AI compute to a builder and exporter of it. NVIDIA founder and Chief Executive Officer Jensen Huang has linked the L&T partnership to the wider vision of making India a global hub for digital infrastructure under the IndiaAI Mission.
Key Takeaways
- L&T’s Vyoma.AI through LTN Compute secured a mega order of ₹10,000 crore to ₹15,000 crore on 13 August 2026 from Together AI to build India’s largest single cluster AI Factory.
- The factory will deploy 10,000 NVIDIA B300 GPUs based on the Blackwell Ultra architecture at Vyoma’s Chennai (Kanchipuram) campus, a gigawatt scale site with Phase 1 of 250 MW and 150 MVA power readiness.
- Each NVIDIA B300 GPU provides 288 GB of HBM3e memory and about 15 petaFLOPS of dense FP4 compute, with NVLink 5 interconnect at 1.8 TB per second.
- Together AI, a San Francisco based AI Native Cloud founded in 2022 by Vipul Ved Prakash and team, will use the cluster for large scale inference, fine tuning and training of open source models.
- The project aligns with the IndiaAI Mission and India’s sovereign cloud push under the Digital Personal Data Protection Act, 2023, strengthening domestic compute while Chennai leverages its position as India’s second largest data centre hub.
