NVIDIA GB300 NVL72 Rack-Scale AI Factory

Overview

The NVIDIA GB300 NVL72 Rack-Scale AI Factory is a liquid-cooled, rack-scale AI infrastructure system built on the NVIDIA Blackwell Ultra platform. It unifies 72 Blackwell Ultra GPUs and 36 Grace CPUs into a single NVLink domain, so an entire rack operates as one accelerator for trillion-parameter training and test-time-scaling inference. Delivered, integrated and supported across India by RDP Technologies Limited — predictable INR pricing, GST invoice and pan-India onsite support.

Key highlights

• 72 NVIDIA Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs unified as a single NVLink domain — the rack behaves as one accelerator.
• 130 TB/s NVLink scale-up bandwidth across all 72 GPUs.
• 20 TB GPU memory, 37 TB fast memory and 17 TB LPDDR5X CPU memory for trillion-parameter models and long-context inference.
• 2,592 Arm Neoverse V2 cores (36 Grace CPUs) for data orchestration and CPU-side pipelines.
• Fully liquid-cooled rack-scale architecture for sustained high-density performance and energy efficiency.
• Engineered for AI factories: training, fine-tuning, high-throughput inference, test-time scaling and agentic / reasoning systems.
• Delivered as a validated reference-architecture rack — onsite integration, deployment and pan-India support by RDP.
• Make-in-India delivery; GST invoice; available on GeM for government and PSU procurement.

AI workload positioning

The GB300 NVL72 is built to be an AI factory, not a single server. With 20 TB of GPU memory, a 130 TB/s NVLink scale-up fabric and 54 TB of total fast and CPU memory across 72 Blackwell Ultra GPUs and 36 Grace CPUs, one rack can train and serve trillion-parameter models, run test-time-scaling (reasoning) inference and host large agentic systems that would otherwise require a room of disaggregated servers. Because the whole rack is one NVLink domain, parallelism scales up rather than out, removing network bottlenecks at frontier scale.

Built for AI factories — who it is for

• Enterprise AI labs and frontier-model teams
• Cloud, neocloud and hosting providers building AI capacity
• Government and sovereign AI programmes
• Research organisations and national HPC / AI centres
• Large IT and infrastructure teams standing up AI factories
• System integrators and hyperscale AI operations

Why it matters

• Rack-scale integration — delivered as a validated reference-architecture rack, not a parts list.
• Scale-up fabric — a single NVLink domain so parallelism scales up, not out, removing network bottlenecks.
• Very high memory capacity — 20 TB GPU + 37 TB fast + 17 TB LPDDR5X for the largest models and longest contexts.
• Liquid cooling — sustained high-density performance with better energy efficiency and thermal headroom.
• Enterprise deployment readiness — onsite integration, support and financing options for production AI factories.

Buy with confidence

• Make-in-India OEM — designed, integrated and supported in India; 14 years and 100,000+ devices shipped.
• Pan-India onsite support with reference-architecture integration and deployment.
• Transparent INR pricing — GST invoice, eligible for input-tax credit; financing and as-a-service options available.
• Sovereign-ready — India data residency (DPDP), MeitY-recognised, ISO 27001 / SOC 2 deployment paths.
• Government procurement — available on GeM for PSU and government AI programmes.
• Talk to a solution architect — a no-cost reference-architecture consultation to design and size your AI factory.

Price on application. Configured and quoted per deployment (power, cooling, fabric and financing). Talk to a solution architect to design your AI factory.

Specifications