DRACO 16× GB300 NVL72 AI SuperCluster

Overview

The DRACO 16× GB300 NVL72 AI SuperCluster is a turnkey, liquid-cooled GB300 NVL72 SuperPOD — 1152 Grace-Blackwell Ultra GPUs across 16 unified NVLink-domain racks, joined by a non-blocking spine-leaf InfiniBand fabric, delivering ~332 TB HBM3e of aggregate GPU memory. It arrives as a complete, validated AI factory — power, cooling, fabric, storage and software — ready to train frontier models on-premises, in INR, on a GST invoice.

Engineered for sovereign-AI programmes, neoclouds and national-scale enterprises, it is delivered as a single engagement: RDP designs the reference architecture, integrates and burns it in, and hands over one validated SuperPOD with one warranty and one support contract — removing multi-vendor integration risk at frontier scale.

Key highlights

• 16× GB300 NVL72 · ~332 TB HBM3e aggregate — 1152 Grace-Blackwell Ultra GPUs for the largest foundation-model training.
• Unified NVLink domains + non-blocking InfiniBand spine — each NVL72 rack is one 72-GPU accelerator; the racks scale over a full-bisection fabric.
• 576× NVIDIA Grace (ARM) (coherent) + Grace LPDDR5X coherent memory — Grace CPUs coherently attached to the Blackwell Ultra GPUs.
• 16 PB parallel NVMe parallel filesystem — high-throughput training data and checkpoint storage at SuperPOD scale.
• 16-Rack SuperPOD, liquid-cooled, turnkey — delivered racked, cabled, cooled, validated; 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 — extend the fabric to multi-pod data halls.

AI workload fit (what it actually runs — honestly)

• Frontier training: distributed pre-training of the largest foundation models across 1152 GPUs with 3D parallelism.
• Large-scale fine-tuning & serving: fine-tune and serve many large models in parallel, or shard the very largest across NVLink domains.
• RAG, multimodal & agentic platforms: national- or organisation-wide production AI on the SuperPOD’s storage and fabric.
• Engineering note: a GB300 SuperPOD combines unified NVLink domains (72 Blackwell Ultra GPUs each, 288 GB per GPU) over a non-blocking InfiniBand spine — within a rack the GPUs act as one accelerator, across racks the spine carries collectives. This is the architecture frontier-model training actually uses; we validate real scaling for your workload, not a peak number.

AI workload positioning

This sits at the top of the data-hall / foundation-model stage: a complete frontier AI supercomputer. With ~332 TB HBM3e of GPU memory on a non-blocking fabric, it is sized to sustain the largest training runs and national-scale serving on-prem — the sovereign, owned alternative to a hyperscale cloud region.

Industry use cases

• Government & sovereign AI — a national foundation-model SuperPOD on GeM-procurable infrastructure.
• Neocloud / AI providers — a premium GPU-cloud region built and validated end-to-end.
• BFSI & conglomerates — private frontier-model training under data-residency rules.
• Healthcare & life sciences — national-scale research and imaging programmes, data in-house.
• Research & national labs — a frontier AI supercomputer.
• Telecom & public sector — sovereign foundation-model platforms.

Performance — and how to be sure

We don’t publish inflated peak numbers. The honest picture: 1152 Grace-Blackwell Ultra GPUs (~332 TB HBM3e) on a non-blocking fabric are sized for the largest foundation-model training and national-scale serving. Want certainty? Request a free scaling benchmark of your model and dataset on a representative configuration before you commit; we’ll send back real tokens/sec, scaling efficiency and projected time-to-train.

Series & upgrade path

• DRACO (flagship supercluster tier) — this.
• Scale ladder: 8-rack → 16-rack → multi-rack data hall GB300 SuperPODs.
• When to step up: extend to a multi-pod data hall — talk to an architect about the fabric, power and facility.

On-prem vs cloud — the TCO case

At SuperPOD scale, owning is a sovereign and economic decision: an always-on hyperscale cloud region dominates any AI budget over a multi-year horizon, while on-prem removes egress 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 model.

Software & day-one readiness

Ships pre-integrated and validated: NVIDIA driver, CUDA, cuDNN, NCCL, the InfiniBand stack, Slurm and/or Kubernetes, container registry, PyTorch and vLLM / Triton / TensorRT-LLM on Ubuntu LTS, with monitoring and a scheduler. Optional managed cluster operations and an MLOps platform.

Power, cooling & rack integration

A 16-rack superpod liquid-cooled SuperPOD with very high power density — RDP scopes facility power, CDU/manifold and water, the InfiniBand spine, floor layout and redundancy in the reference-architecture design. (Exact power, BTU, flow and facility figures confirmed in the design package.) Full out-of-band management.

Deployment, warranty & support

• Made to order, integrated, racked, cabled, cooled and burned-in in India; project timeline confirmed at quote.
• Delivered as one system: NVL72 racks, leaf/spine switches, cabling, PDUs, cooling, parallel storage, and the full cluster software stack.
• Onsite warranty + AMC with pan-India coverage, cluster-level SLAs and an RMA/escalation path (exact terms 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 turnkey frontier AI SuperPOD, made to order — talk to an RDP solution architect, co-design the reference architecture, fabric and data hall, get a multi-year TCO and financing plan, and run a scaling benchmark before you commit. Request a quote to begin.

Specifications