ASUS ExpertCenter Pro ET900N G3 – NVIDIA DGX Station GB300 AI Supercomputer, 748GB, 8TB

NVIDIA DGX Station GB300 AI Supercomputer

The ASUS ExpertCenter Pro ET900N G3 is a deskside NVIDIA DGX Station GB300 AI supercomputer built for enterprise AI development, LLM fine-tuning and local inference at data-centre scale.

The ASUS ExpertCenter Pro ET900N G3 brings the NVIDIA DGX Station GB300 architecture out of the data centre and onto your desk, pairing a 72-core NVIDIA Grace CPU Superchip with an NVIDIA Blackwell Ultra GPU across 748GB of coherent unified memory. It is built for enterprise AI teams, research labs and funded AI startups that need to fine-tune, inference and iterate on large language models locally, without renting cloud GPU capacity or queuing for data-centre time. Every NVIDIA DGX Station GB300 system shares the same Grace Blackwell Superchip architecture that powers NVIDIA’s largest AI data centres, scaled down into a single deskside unit that a small team can own, secure and control outright.

Grace Blackwell Ultra – Unified CPU-GPU Architecture

At the heart of the system is a single NVIDIA Grace Blackwell Ultra Superchip, a 72-core Grace CPU built on Arm Neoverse V2 cores, connected to a Blackwell Ultra GPU over NVIDIA’s NVLink-C2C interconnect. The CPU and GPU share a single coherent memory pool, 496GB of LPDDR5X on the CPU side and 252GB of HBM3e on the GPU side, so models and datasets that would normally need to be split, offloaded or quantised to fit in GPU memory can instead be addressed as one 748GB space. NVIDIA rates the system at up to 20 PFLOPS of FP4 AI compute, enough to fine-tune and run inference on frontier models of up to a trillion parameters on a single desk. In raw compute terms this sits comfortably above single high-end workstation GPUs and below full multi-GPU AI servers, bringing data-centre-class throughput to a single desk without server-room infrastructure. The GPU also supports NVIDIA Multi-Instance GPU (MIG), partitioning into up to seven isolated instances so several users or workloads can share the system securely instead of one project tying up the whole machine, and NVIDIA’s workload-optimised power-shifting automatically balances performance and efficiency as workloads change.

8TB NVMe Storage for Datasets and Checkpoints

The ET900N G3 ships with four PCIe 5.0 NVMe M.2 2280 SSDs totalling 8TB of raw capacity. NVIDIA’s recommended configuration mirrors two of those drives as a RAID 1 operating system volume for redundancy, with the remaining two allocated to training data, so usable space for datasets and checkpoints will be somewhat less than the full 8TB once that redundancy is accounted for. PCIe 5.0 throughput keeps data loading fast enough to avoid becoming the bottleneck ahead of a 7.1TB/s HBM3e GPU memory pool, which matters when every second of idle GPU time on a system this expensive has a real cost.

Networking – Dual 400GbE for Multi-System Scale-Out

Networking is built around two NVIDIA ConnectX-8 SuperNIC ports (QSFP112, 400Gb/s each, 800Gb/s combined), designed for connecting multiple DGX Station systems together or linking directly into existing data-centre AI infrastructure, plus a dedicated 10GbE Marvell port for fast data transfer and a separate 1GbE management LAN for out-of-band BMC access. Front-panel connectivity includes two USB 10Gbps Type-C ports, two USB 10Gbps Type-A ports and a USB 2.0 port for everyday peripherals, keyboards and external drives.

AI Software Stack – Ready to Train, Fine-Tune and Deploy

The ET900N G3 is supplied running Ubuntu Linux with the NVIDIA AI Enterprise software stack pre-installed, including CUDA-X libraries, RAPIDS for data science workflows and NVIDIA NIM microservices for deploying models as production-ready inference endpoints. It supports the frameworks AI teams already use, including PyTorch, TensorFlow and the NVIDIA NeMo framework, so existing training pipelines and notebooks can move onto the system with minimal rework.

Ideal for

• Enterprise AI and machine learning teams fine-tuning and deploying LLMs on-premise
• AI research labs and university departments running large-scale model experiments
• Funded AI startups that need dedicated compute without ongoing cloud GPU rental costs
• Data science teams working with sensitive or regulated data that cannot leave the building
• Robotics, automotive, aerospace and healthcare teams training and simulating physical AI models locally

For any organisation weighing up cloud GPU rental against owning hardware outright, the NVIDIA DGX Station GB300 architecture in the ASUS ExpertCenter Pro ET900N G3 offers a fixed, predictable cost, full control over where sensitive data is processed, and enough unified memory to work with frontier-scale models without the queueing, quota limits or ongoing spend of a shared cloud environment. Because the hardware is paid for once rather than billed per token or per hour, the more an organisation uses the system, the lower its effective cost per unit of AI work becomes relative to metered cloud API pricing, a calculation that increasingly favours ownership for teams running AI workloads every day rather than occasionally.

Supplied and supported in the UK by Punch Technology, the ASUS ExpertCenter Pro ET900N G3 is covered by ASUS’s 3-year direct warranty with a 7-day dead-on-arrival replacement period, backed by Punch’s UK-based pre-sales advice and technical support and over 30 years of experience specifying professional computing hardware.

Full Specification

What is the ASUS ExpertCenter Pro ET900N G3 best for?

The ASUS ExpertCenter Pro ET900N G3 is primarily an on-premise AI development and inference platform for organisations that need data-centre-class AI compute without sending data to the cloud. It is built to fine-tune and run inference on large language models, including models scaled up to around a trillion parameters, entirely on local hardware.

In practice this covers workloads such as fine-tuning open-weight LLMs (Llama, Qwen, Mistral and similar model families) on proprietary data, running retrieval-augmented generation and agentic AI pipelines with NVIDIA NIM, training and simulating robotics and autonomous-system models under NVIDIA’s Physical AI workflows, and supporting data science teams running RAPIDS-accelerated analytics on datasets too large or sensitive to move off-site.

Where does the ASUS ExpertCenter Pro ET900N G3 sit in the range?

The ET900N G3 is a category of its own on the Punch Technology range, sitting well above our most capable creator and AI-focused gaming PCs, which use consumer GPUs such as the RX 7900 XTX for local Stable Diffusion and content-creation workloads. Where those systems suit individual creators experimenting with AI tools, the ET900N G3 is built for teams that need dedicated, data-centre-class compute they own outright. For organisations that only need occasional AI experimentation rather than dedicated training and inference capacity, our 24GB VRAM creator PCs remain a significantly lower-cost starting point.

Frequently Asked Questions

What AI frameworks and models does the ASUS ExpertCenter Pro ET900N G3 support?

The system ships with Ubuntu Linux and the NVIDIA AI Enterprise software stack pre-installed, including CUDA-X libraries, RAPIDS and NVIDIA NIM microservices, alongside NVIDIA NemoClaw for secure, policy-controlled agentic AI workflows. It supports PyTorch, TensorFlow, the NVIDIA NeMo framework and vLLM for inference, and can run open-weight models including Llama, Qwen, Mistral, DeepSeek and NVIDIA’s own Nemotron models locally, covering workloads from LLM training and inference through to computer vision, generative AI content creation and scientific computing, with enough unified memory to work with models up to roughly a trillion parameters.

How much memory does the ET900N G3 have, and why does that matter for AI work?

It has 748GB of coherent unified memory: 496GB of LPDDR5X attached to the Grace CPU and 252GB of HBM3e attached to the Blackwell Ultra GPU, connected over NVIDIA’s high-bandwidth NVLink-C2C interconnect. Because the CPU and GPU share one memory pool rather than two separate ones, large models and datasets do not need to be split, quantised or offloaded just to fit in GPU memory, which is usually the limiting factor on smaller AI workstations.

Why buy an on-premise DGX Station instead of renting cloud AI compute?

Cloud GPU instances are billed continuously and are shared, queued or capacity-limited at busy times, and sending proprietary or regulated data off-site is not acceptable for every organisation. Owning an NVIDIA DGX Station GB300 system outright gives a fixed one-off cost, dedicated always-available capacity, and full control over where data is processed and stored, which matters for organisations in regulated sectors or working with confidential model training data.

What power and space does the ASUS ExpertCenter Pro ET900N G3 need in my office?

The system uses a single 1600W Titanium-rated power supply and should be run on a dedicated circuit appropriate for that load, in a well-ventilated space. ASUS specifies an operating range of 10C to 35C ambient, and recommends keeping ambient below 30C during sustained heavy workloads. It weighs 27kg and measures 584mm x 232mm x 565mm as a deskside tower, so it needs floor or desk space similar to a large tower workstation rather than a standard office PC.

Who is this PC best suited for?

The ASUS ExpertCenter Pro ET900N G3 is best suited to enterprise AI and machine learning teams, research labs and well-funded AI startups that need dedicated, on-premise AI compute for LLM fine-tuning, inference and data science work, including domain-specific teams in automotive, aerospace, healthcare and medical research, and smart infrastructure. It is not intended for general office use, gaming, or as a first step into AI, those buyers are better served by Punch Technology’s business PCs or creator-focused gaming PCs.

What support and warranty comes with the ASUS ExpertCenter Pro ET900N G3?

The system is covered by ASUS’s 3-year on-site warranty, with onsite service available after remote diagnosis to help minimise downtime on critical AI workloads (availability varies by region), plus a 7-day dead-on-arrival replacement period. A closed-loop AIO liquid cooling system with a data centre-grade thermal design gives the ET900N G3 enough headroom to run full-load AI training and inference 24/7 without thermal throttling, so it is built for sustained work rather than short bursts. Punch Technology also provides UK-based pre-sales advice and technical support to help specify and deploy the system correctly, contact Punch directly to confirm current lead times and delivery arrangements before ordering.