Collection: NVIDIA DGX Station AI Supercomputer

NVIDIA DGX Station is a desktop AI supercomputer built on the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip — a Grace 72-core Arm Neoverse V2 CPU paired with a Blackwell Ultra GPU, connected via NVLink-C2C at 900 GB/s. It delivers up to 20 petaFLOPS of AI compute (FP4) with 748 GB of total coherent memory (252 GB HBM3e + 496 GB LPDDR5X), and is built to train, fine-tune, and serve AI models up to 1 trillion parameters locally. Designed for serious on-prem AI: enterprise data science, research labs, agentic AI, and multi-user inference.

DGX Station fits AI research labs, enterprise data science and ML teams, university HPC programmes, AI-product engineering teams, and businesses needing on-prem inference at scale. The common need: train and run very large models locally (LLMs, VLMs, agentic AI, physical AI) without paying cloud GPU per-hour costs and without sending sensitive data to public cloud.

NVIDIA positions DGX Station for models up to 1 trillion parameters — the full coherent 748 GB memory pool (252 GB HBM3e GPU + 496 GB LPDDR5X CPU, unified via NVLink-C2C) gives it the headroom to fit, fine-tune, and serve frontier-scale models that don't fit on workstation-class GPUs.

Realistic ranges: 7B–70B (excellent throughput, low latency); 120B–200B (strong); 405B (production-feasible); 600B–1T (frontier — the kind of work that previously required multi-node DGX H100/B200 clusters).

LLM training and fine-tuning up to 1T-parameter scale — pre-training, SFT, RLHF, DPO, LoRA / QLoRA at production sizes. Multi-user inference via NVIDIA Multi-Instance GPU (MIG) up to 7 isolated partitions — one Station serves a whole team or department. Agentic AI with NVIDIA NemoClaw for secure tool-calling and autonomous workflows. Vision-language models, multimodal, physical-AI training for robotics/autonomous systems. Massive-dataset data science via RAPIDS. Real-world: private GPT-class fine-tuning on enterprise data, multi-team inference, AI Foundry-style internal LLM deployments.

Yes — DGX Station is the first NVIDIA desktop platform built for serious on-prem training. Pre-training small-to-mid foundation models, fine-tuning any model up to 1T parameters (SFT, RLHF, DPO, LoRA, QLoRA), continued pre-training on domain data, RAG indexing and embedding work. For training scale that exceeds a single Station, the platform is built to compose into clusters with QSFP112 networking at 400 Gb/s per port; EMARQUE AI can size multi-node setups.

Yes. NVIDIA Multi-Instance GPU (MIG) partitions the Blackwell Ultra GPU into up to 7 fully-isolated GPU instances — each with its own dedicated memory, compute, and L2 cache. This means one DGX Station can serve up to seven simultaneous users, models, or workloads without resource contention. Ideal for shared internal AI environments, research teams, or multi-tenant inference services.

Same family, very different scale. DGX Spark: GB10 Grace Blackwell, 128 GB unified memory, ~1 PFLOPS FP4, models up to ~200B (practical ~120B), 240W — the personal AI development unit. DGX Station: GB300 Grace Blackwell Ultra, 748 GB coherent memory, 20 PFLOPS FP4 (153 PFLOPS sparse peak), models up to 1 trillion parameters, MIG up to 7 users, 1,600W.

Spark = local AI development, prototyping, fine-tuning small-to-mid LLMs. Station = serious on-prem AI infrastructure, multi-user inference, frontier-scale training. EMARQUE AI helps you size the right tier (email business@emarque.co).

RTX PRO 6000 Blackwell Workstation — choose when you need mixed AI + creator workloads (rendering, simulation, visualization, AI). Workstation flexibility on a familiar form factor. 96 GB GDDR7 VRAM.

DGX Station — choose when AI is the primary workload, models exceed what fits in 96 GB VRAM, multiple users need to share one box (MIG up to 7), or you're doing serious fine-tuning / training. 748 GB coherent memory. 20 PFLOPS FP4. Built specifically for the AI stack.

Multi-GPU AI servers (DGX B200 etc.) — choose when one Station isn't enough: multi-Station clusters via 400 Gb/s networking, full data-center AI deployments, training that scales beyond a single GB300. EMARQUE AI scopes and integrates these.

DGX Station ships with NVIDIA DGX OS (Ubuntu base + NVIDIA AI Developer Tools, qualified and locked for AI workloads) and the full NVIDIA AI Enterprise stack — CUDA, container runtimes, NIM microservices, AI Workbench, and pre-built framework images (PyTorch, vLLM, RAPIDS, NeMo). Plus NVIDIA NemoClaw for secure agentic AI workflows. You can run frontier models out of the box without spending weeks on stack setup.

DGX Station ships with the NVIDIA ConnectX-8 SuperNIC at up to 800 Gb/s Ethernet, plus 2× QSFP112 ports at 400 Gb/s each and 1× RJ45 10 GbE. Designed to compose into multi-Station clusters for training that exceeds a single box, and to integrate cleanly into existing data-center fabrics. EMARQUE AI handles ConnectX-8 / QSFP112 deployment, NCCL/MPI tuning, and rack-level networking as solutions on request.

DGX Station is officially launching in 2026 via NVIDIA's OEM partners (MSI, ASUS, Gigabyte, Dell, HP, Supermicro). EMARQUE is among the first solution providers in Malaysia taking pre-orders. Pre-ordering with EMARQUE locks your allocation when partner units arrive, gives you direct access to the EMARQUE AI team for deployment scoping (MIG configuration, ConnectX-8 networking, private LLM stack, team training), and confirms final pricing once partner pricing is locked. Click ENQUIRE NOW on any variant page or email business@emarque.co.

Yes. EMARQUE has helped 300+ Malaysian clients integrate AI and high-performance computing — enterprises, SMEs, research labs, universities, government agencies, and creative studios. Deployments span private LLM stacks, on-prem RAG, document AI, computer vision, custom AI workstations, and multi-GPU server integration. Common thread: local data control, predictable performance, Malaysia-based support. Doing this since 2016.

EMARQUE is a DGX Station solution provider and systems integrator in Malaysia — not just a reseller. The EMARQUE AI team handles sizing and configuration; ConnectX-8 / QSFP112 networking deployment; MIG partitioning for multi-user serving; NemoClaw integration for agentic AI; private LLM / VLM / RAG / agentic-AI stack deployment; team training; and Malaysia-based post-sale support — all available as EMARQUE AI Solutions on request. Scope, timing, and charges quoted on enquiry. Email business@emarque.co.

On-prem AI is the right choice when privacy (data sovereignty, PDPA compliance, sensitive enterprise records), predictable cost (vs. unbounded cloud GPU bills as model use scales), latency (no cloud round-trip for real-time inference and agentic loops), or customisation (full control of models, weights, fine-tunes, and the full stack) matters more than cloud convenience. EMARQUE designs DGX Station deployments for these realities — for many Malaysian businesses, on-prem with EMARQUE is the better long-term fit.

Yes. EMARQUE AI is our dedicated AI compute division — full NVIDIA AI hardware stack and on-prem deployment for Malaysian businesses, research labs, and enterprises: NVIDIA Personal AI (DGX Spark, 2× Spark stacking), AI workstations (RTX PRO 6000 Blackwell and others), DGX Station GB300, multi-GPU AI servers, plus full integration support. Visit EMARQUE.AI or contact business@emarque.co.