MSI · NVIDIA GB10 · LOCAL AI WORKSTATION
A compact AI development platform for your desk
Build and test compatible AI applications locally with the MSI EdgeXpert-34SAU. The NVIDIA GB10 Grace Blackwell Superchip, 128GB unified memory and 4TB Gen5 NVMe SSD bring substantial model capacity into a compact desktop system, backed by a 3-year Australian on-site warranty.
The configuration you are buying
EdgeXpert-34SAU · 128GB LPDDR5x · 4TB Gen5 NVMe · NVIDIA DGX OS. This page describes the Australian configuration. Other regional bundles, memory options and promotional illustrations may differ. The price is for one system.
What the hardware does for your work
GB10 Grace Blackwell
A 20-core Arm CPU and NVIDIA Blackwell GPU work together in a platform designed for AI development and inference.
128GB unified memory
A coherent memory pool is shared by the CPU and GPU, giving compatible large-model workloads room to operate.
4TB Gen5 NVMe storage
Store model files, datasets, containers and development environments locally. Available space is lower after formatting and software installation.
NVIDIA DGX OS
An Ubuntu Linux-based environment for the NVIDIA AI software ecosystem. Check Linux/Arm support for your chosen tools.
10GbE networking
The RJ45 Ethernet interface supports fast links to a compatible network. End-to-end speed depends on the switch, cable and other devices.
ConnectX-7 expansion
High-speed networking supports planned multi-system AI deployments. Interconnect equipment and configuration are separate.
Compact desktop design
The 151 × 151 × 52 mm chassis fits a development workspace while retaining dedicated cooling.
Australian on-site warranty
The supplied Australian model has a 36-month on-site warranty. Ask us about the applicable coverage and service terms.
Official MSI networking illustration. Additional EdgeXpert systems, QSFP cables and other network equipment shown are sold separately.Where it can fit in your business
Local language-model experiments
Evaluate supported LLMs, compare prompts and test retrieval-augmented generation against your own datasets. Choose model size, quantisation and context length around the available memory.
Development and data science
Build reproducible workflows with compatible notebooks, frameworks and containers. MSI identifies tools such as PyTorch, Jupyter and Ollama in the platform ecosystem; support varies by release and architecture.
Education and research
Provide a local platform for hands-on inference, model evaluation and supported fine-tuning experiments. Training capacity depends on model size, method and precision.
AI vision and edge development
Prototype compatible computer-vision and edge-AI applications before a wider deployment. This system is not itself a preconfigured CCTV recorder or a ready-made security analytics licence.
Understanding the AI performance figures
The platform is rated at up to 1 petaFLOP of FP4 sparse AI tensor performance. That is a specific low-precision AI measurement, not an all-purpose speed rating. Model architecture, quantisation, batch size, context length, framework and software versions affect usable performance.
MSI describes a single 128GB system as capable of running models up to 200 billion parameters under supported conditions. A model that fits in memory may still have a different response speed from a smaller model. Share your intended model and workflow with us so the configuration can be checked against the software requirements.
Linux/Arm software compatibility
EdgeXpert combines an Arm processor with NVIDIA DGX OS. Verify application binaries, containers, Python packages, drivers and third-party extensions for that platform. A package available for a Windows or x86 Linux workstation does not automatically have an equivalent Arm build.
The NVIDIA AI software stack supports local development workflows, but application licences, cloud subscriptions and commercial support contracts should be assessed separately. This listing does not include custom AI model deployment, application development or an enterprise software subscription.
Official MSI thermal-design illustration. Keep the chassis ventilation clear and follow the manufacturer’s installation guidance.Plan your workspace and setup
-
Confirm the workload: identify your model, framework, precision, context size and Linux/Arm dependencies.
-
Prepare connectivity: select a suitable network connection and peripherals. Four built-in USB Type-C ports require an adapter for USB-A-only accessories.
-
Allow ventilation: place the system on a stable surface with clear airflow and follow the manufacturer’s power and environmental instructions.
-
Complete system setup: follow MSI and NVIDIA documentation, install approved updates and configure user access and backups.
-
Validate with your data: measure the actual workload before committing to a larger rollout or multi-system design.
Scaling to a second system
ConnectX-7 networking provides a route to larger multi-system experiments. Two units do not become a larger workstation simply by connecting a cable: your topology, interconnect and software must support the intended distributed workload. A second unit, QSFP cable and any switch or deployment services are additional purchases. Request a configuration check or multi-system quote.
Technical specifications
| Australian manufacturer model |
EdgeXpert-34SAU |
| Platform |
NVIDIA DGX Spark / NVIDIA GB10 Grace Blackwell |
| CPU |
20-core Arm: 10 Cortex-X925 + 10 Cortex-A725 |
| GPU architecture |
NVIDIA Blackwell |
| AI tensor rating |
Up to 1 petaFLOP; FP4 sparse operations; workload-dependent |
| Memory |
128GB LPDDR5x coherent unified system memory |
| Memory interface |
256-bit |
| Memory bandwidth |
273GB/s |
| Storage supplied |
4TB NVMe M.2 SSD; PCIe Gen5 Australian configuration |
| Operating system |
NVIDIA DGX OS, based on Ubuntu Linux; Arm architecture |
| USB connections |
4 × USB Type-C |
| RJ45 network |
1 × 10 Gigabit Ethernet |
| High-speed NIC |
NVIDIA ConnectX-7; up to 200GbE platform interface |
| Wireless |
Wi-Fi 7 and Bluetooth; performance depends on compatible network and region |
| Display connection |
HDMI 2.1a |
| Audio output |
HDMI multichannel audio |
| Video engines |
1 × NVENC and 1 × NVDEC |
| Chassis dimensions |
151 × 151 × 52 mm |
| System weight |
Approximately 1.2kg; shipping weight is higher |
| Power |
External system power supply; follow the supplied adapter requirements |
| Warranty |
36-month Australian on-site warranty |
| Barcode |
4711377403856 |
What is included—and what to add
Included: one EdgeXpert-34SAU system in the configuration above and its system power supply.
Optional: display, keyboard, pointing device, adapters, network cabling, a second system and compatible QSFP interconnect equipment. Manufacturer demonstrations can show multiple systems or additional equipment.
For a simple wired pointing device, the Logitech M90 can be connected through a suitable USB-C to USB-A adapter. Confirm your complete input-device arrangement and application requirements before ordering.
Official MSI video
Official MSI platform demonstration. Regional configurations, software and optional equipment shown may differ from the EdgeXpert-34SAU supplied here.
Official documentation and buying help
Frequently asked questions
What makes this different from a normal PC?
EdgeXpert is built for compatible local AI workloads. Its NVIDIA GB10 combines an Arm CPU with a Blackwell GPU and a shared 128GB memory pool. It runs NVIDIA DGX OS, based on Ubuntu Linux. It is intended for developers and technical teams rather than as a drop-in Windows office or gaming PC.
What does 128GB unified memory mean?
The CPU and GPU share one coherent LPDDR5x memory pool. This helps large models use memory without the same separate CPU-RAM and graphics-VRAM split found in many PCs. Operating-system use, model weights, runtime overhead and context cache all consume that pool.
Can it run a 200-billion-parameter model?
MSI describes support for models up to 200B parameters on a 128GB system. This is a workload-dependent capacity claim, not a guarantee for every model, precision or context length. Quantisation, framework support and available working memory determine whether a particular model fits and how fast it runs.
Is the advertised petaflop performance available for every task?
No. The platform rating is up to 1 petaFLOP for FP4 sparse AI tensor operations. It does not describe general CPU speed, FP32 compute or guaranteed tokens per second. Real performance depends on the model, precision, software and cooling conditions.
Does it run Windows software?
It ships with NVIDIA DGX OS on an Arm processor. Check for a supported Linux/Arm build or compatible container of your application. Standard Windows x86 applications and drivers are not directly interchangeable with this platform.
Does local AI automatically keep everything private?
Local inference can keep model processing and data on your own network, but privacy still depends on the applications, update services, external APIs, permissions and network policies you configure. Hardware alone does not certify a deployment as compliant or eliminate all data-transfer risks.
Can I connect more than one system?
The ConnectX-7 interface supports high-speed multi-system designs. A second system, compatible QSFP interconnect cable, and suitable software configuration are separate. Ask us to confirm the topology and software requirements before ordering a cluster.
What is included at this price?
One Australian EdgeXpert-34SAU system with 128GB unified memory, a 4TB Gen5 NVMe SSD, NVIDIA DGX OS and the system power supply. A monitor, keyboard, mouse, QSFP cable and paid software or deployment services are separate.
What do I need for the first setup?
Plan a suitable workspace with ventilation, mains power, network access, and a supported client or local display/input arrangement. Review the NVIDIA setup documentation and your software requirements before deployment. The built-in USB ports are Type-C; USB-A peripherals need an appropriate adapter.
What warranty applies?
This Australian supplier configuration is listed with a 36-month on-site warranty. Retain your invoice and contact CCTV Guru for warranty assistance and the applicable service terms. On-site warranty is not a promise of a specific response time or included AI software development.
