L&T Lands $1.57 Billion AI Data-Centre Deal with Together AI Powered by NVIDIA Silicon

L&T secures a landmark $1.57B contract with Together AI to engineer hyperscale AI data-centre infrastructure across India, powered by next-gen NVIDIA chips.

Indian engineering and infrastructure giant Larsen & Toubro (L&T) has secured a monumental $1.57 billion multi-year contract from Silicon Valley-based cloud compute leader Together AI to design, construct, and manage hyperscale artificial intelligence data-centre campuses across India. The infrastructure deployment will be powered by enterprise clusters of NVIDIA Blackwell and Hopper AI processors, establishing one of South Asia's largest dedicated sovereign AI compute hubs.

The milestone transaction underscores India’s rapid emergence as a global destination for high-density compute infrastructure. As global foundation model developers face intense power constraints and server real estate shortages in North America and Western Europe, India’s expanding green grid capacity and engineering prowess are attracting multi-billion-dollar enterprise commitments.

Engineering Architecture: Powering Next-Gen GPU Densities

Under the terms of the $1.57 billion agreement, L&T’s data centre and digital infrastructure division will deliver modular hyperscale facilities across key tech nodes including Chennai, Navi Mumbai, and the NCR corridor. The facilities are architected specifically to handle the unprecedented thermal and electrical demands of modern dense AI clusters.

Standard legacy data centres are typically engineered for power densities between 6 kW and 12 kW per rack. In contrast, the Together AI facilities engineered by L&T will support up to 100 kW per rack, incorporating state-of-the-art liquid-to-chip cooling loops and closed-circuit evaporative systems to maintain a Power Usage Effectiveness (PUE) target of 1.18.

This partnership represents a structural inflection point for India’s digital engineering sector,

said S. N. Subrahmanyan, Chairman and Managing Director of Larsen & Toubro. "Deploying gigawatt-scale AI infrastructure requires deep domain expertise spanning power substation construction, microgrid integration, seismic mechanical isolation, and ultra-high-efficiency thermal design. In collaboration with Together AI and NVIDIA, we are building the computing bedrock for the next generation of global intelligence."

The strategic deployment complements national infrastructure initiatives outlined in our analysis of India's sovereign AI compute bottleneck, while integrating with regional developments such as HCLTech and Sarvam AI's Odisha compute hub.

Technical Specifications: L&T Hyperscale AI Data-Centre Blueprint

The operational specifications and electrical distribution architecture for the Together AI buildout highlight the immense scale of the project:

| Infrastructure Metric | Baseline Enterprise Facility | L&T – Together AI Hyperscale Hub | Architectural Innovation | | :--- | :--- | :--- | :--- | | Total Target Compute Capacity | 30 MW – 50 MW | 250 MW Scalable Phase 1 | Modular grid substations with 100% redundant feeds | | Rack Power Density | 8 kW – 14 kW / rack | 85 kW – 100 kW / rack | Direct-to-chip liquid cooling manifolds | | Target PUE Rating | 1.45 – 1.60 | 1.18 Annualized | Closed-loop adiabatic chiller heat dissipation | | Silicon Architecture | General Purpose x86 / Mixed | NVIDIA Blackwell & H100/H200 | Ultra-dense NVLink-connected multi-node clusters | | Renewable Energy Integration | 20% – 35% | 65% Dedicated Solar/Wind PPAs | Onsite battery energy storage systems (BESS) | | Network Fabric | 100G Ethernet | 800G InfiniBand Quantum-X800 | Low-latency non-blocking leaf-spine topology |

#

Overcoming Thermal and Energy Challenges

The primary engineering bottleneck in AI cluster deployment is thermal management. High-performance accelerators running continuous distributed training workloads generate immense localized heat. L&T is implementing direct-to-chip cold plates circulating deionized dielectric fluid, eliminating air-cooling inefficiencies and reducing server parasitic fan draw by up to 32%.

Furthermore, to mitigate grid transmission volatility, the data centres will be directly coupled with dedicated renewable energy purchase agreements (PPAs) and grid-scale lithium iron phosphate (LFP) battery storage backups. This guarantees five-nines (99.999%) operational uptime even during peak regional demand periods.

#

Economic Impact on the Indian Deeptech Ecosystem

The presence of domestic low-latency NVIDIA compute capacity will drastically reduce inference and fine-tuning costs for Indian startups and enterprise labs. Rather than routing sensitive enterprise datasets to overseas servers in Singapore or Virginia, Indian developers can train domain-specific models locally under strict data residency protocols.

Industry leaders note that large-scale infrastructure investments will also spur downstream venture creation. As detailed in our breakdown of experienced founders driving larger rounds and India's AI startup growth engine, readily accessible compute serves as the catalytic multiplier for indigenous enterprise applications and sovereign AI foundation models.