DRACO GB200 NVL72 Rack-Scale AI System

Overview

The DRACO GB200 NVL72 Rack-Scale AI System is a turnkey, liquid-cooled rack-scale AI system built on NVIDIA’s GB200 NVL72 platform — 72 Grace-Blackwell GPUs joined in a single NVLink domain with 13,824 GB HBM3e of aggregate HBM3e. The whole rack behaves as one enormous accelerator, sized to train and serve frontier-scale models on-premises — in INR, on a GST invoice.

Engineered for national programmes, neoclouds and large enterprises building foundation-model capability in-house, it arrives racked, cabled, liquid-cooled and validated as a single SKU with one warranty and one support contract — RDP scopes the NVLink domain, InfiniBand spine, storage, power and cooling as one system so you don’t carry the integration risk.

Key highlights

• GB200 NVL72 · 13,824 GB HBM3e aggregate HBM3e — one unified NVLink domain of 72 Grace-Blackwell GPUs for frontier-scale training.
• Unified NVLink/NVSwitch domain — all GPUs in a rack act as a single accelerator; coherent Grace CPUs over NVLink-C2C; 400G+ InfiniBand spine for scale-out.
• 36× NVIDIA Grace (ARM) — Grace CPUs coherently attached to the Blackwell GPUs.
• 480 TB NVMe NVMe + parallel-FS ready — high-throughput data and checkpoint storage at frontier scale.
• Single-Rack (NVLink domain), liquid-cooled, turnkey — delivered racked, cabled, cooled and burned-in; one SKU, one warranty.
• On-prem data sovereignty — training data and weights stay in-house; DPDP-friendly, air-gappable.
• Make-in-India OEM — predictable INR pricing, GST tax invoice (HSN 8471), pan-India onsite support, GeM-procurable.
• Scale path — interconnect multiple systems into a multi-rack RDP AI SuperCluster.

AI workload fit (what it actually runs — honestly)

• Frontier training: pre-train and fine-tune the largest (multi-hundred-billion to trillion-parameter) models, tensor-/pipeline-/data-parallel across the NVLink domain.
• Large-scale inference: serve the very largest models with the whole rack as one memory pool, or many large models concurrently.
• RAG, multimodal & agentic at scale: the most demanding production AI platforms.
• Engineering note: the defining feature of an NVL system is the unified NVLink domain — 72 GPUs addressed as one accelerator with coherent Grace memory, which is precisely what trillion-parameter training needs; across racks, a high-bandwidth InfiniBand spine carries collectives. We validate real scaling for your workload rather than quoting a peak number.

AI workload positioning

This sits at the top of rack-scale train-and-serve: a complete AI factory in a rack. With 13,824 GB HBM3e of HBM3e in a unified NVLink domain, it is sized to sustain frontier-scale training and serving on-prem — the sovereign, owned alternative to renting the largest cloud systems continuously.

Industry use cases

• Government & sovereign AI — a national foundation-model factory on GeM-procurable infrastructure.
• Neocloud / AI providers — the flagship unit to build a competitive GPU cloud.
• BFSI & large enterprise — private frontier-model training under data-residency rules.
• Healthcare & life sciences — large-scale model and imaging research, data in-house.
• Research & national labs — an institutional frontier-AI system.
• Telecom & conglomerates — in-house foundation-model platforms.

Performance — and how to be sure

We don’t publish inflated peak numbers. The honest picture: 13,824 GB HBM3e of HBM3e across 72 Grace-Blackwell GPUs in a unified NVLink domain is sized for Trillion-scale training and serving. Want certainty? Request a free benchmark of your model and dataset — including a scaling test — on this exact configuration before you buy; we’ll send back real tokens/sec, scaling efficiency and timings.

Series & upgrade path

• DRACO (flagship rack-scale tier) — this.
• NVL ladder: GB200 NVL72 → GB300 NVL72 (Blackwell Ultra, more HBM/GPU) → multi-rack (2×+).
• When to step up: interconnect multiple NVL systems into an RDP AI SuperCluster — talk to an architect about the spine and facility.

On-prem vs cloud — the TCO case

At frontier scale, owning is a strategic decision: an always-on cloud system of this size dominates any AI budget, and on-prem removes egress fees and keeps sovereign data and weights in-house. RDP pricing is fixed in INR with a GST input-credit-eligible invoice — ask for a multi-year TCO and financing comparison.

Software & day-one readiness

Ships pre-integrated and validated: NVIDIA driver, CUDA, cuDNN, NCCL, the InfiniBand stack, Docker and NVIDIA Container Toolkit, with Slurm or Kubernetes, PyTorch and vLLM / Triton / TensorRT-LLM on Ubuntu LTS. Optional managed operations, scheduler and observability for the full system.

Power, cooling & rack integration

A single-rack (nvlink domain) liquid-cooled system with very high power density — RDP scopes facility power, CDU/manifold and water, the InfiniBand spine and floor layout as part of the design. (Exact PDU ratings, BTU, flow and facility figures confirmed on the build sheet.) Full out-of-band management across the system.

Deployment, warranty & support

• Made to order, integrated, racked, cabled, cooled and burned-in in India; realistic lead time confirmed at quote.
• Delivered as one system: the NVL rack(s), fabric switches, cabling, PDUs, cooling integration, and the pre-installed cluster software stack.
• Onsite warranty + AMC with pan-India coverage, system-level support and an RMA/escalation path (exact term & response window confirmed at quote).

Why RDP

14 years of Make-in-India infrastructure and 300,000+ devices shipped. Indian OEM, INR pricing, GST tax invoice (HSN 8471), pan-India onsite engineers, GeM availability, and DPDP / sovereign-AI-ready deployment.

Buy with confidence

This is a frontier-scale AI factory, made to order — talk to an RDP solution architect, size the NVL domain, spine and facility, get a multi-year TCO and financing plan, and benchmark your own model with a scaling test before you commit. Request a quote to begin.

Specifications