MSI Details XpertStation WS300 DGX Station Workstation
On March 17, 2026, MSI detailed the XpertStation WS300T60L, a tower workstation based on NVIDIA's DGX Station architecture and GB300 Grace Blackwell Ultra platform. MSI's product page lists a 72-core Grace CPU, one Blackwell Ultra GPU, 748 GB of coherent memory, dual 400GbE ports, and Ubuntu 24.04 LTS with NVIDIA AI developer tools. The system targets AI development, data science, inference, and personal cloud deployments.
MSI has published detailed specifications for the XpertStation WS300T60L, a tower-form-factor AI workstation based on NVIDIA's DGX Station architecture and the GB300 Grace Blackwell Ultra platform. MSI's product page lists a 72-core Arm Neoverse V2 Grace CPU, a single NVIDIA Blackwell Ultra GPU, 748 GB of coherent memory, and two 400GbE QSFP112 ports supplied by an integrated NVIDIA ConnectX-8 SuperNIC.
MSI announced the XpertStation WS300 in March 2026 as a deskside AI system for large language models, generative AI, and data-science workflows. The manufacturer's current barebones listing identifies intended applications as AI development, data science, AI inference, and personal cloud deployments.
Compute, memory, and networking
According to MSI, the system combines GPU HBM3e and Grace CPU LPDDR5X memory in a coherent memory domain, intended to enable CPU-GPU data sharing. The published configuration includes:
- •A 72-core NVIDIA Grace CPU Superchip using Arm Neoverse V2 cores
- •One NVIDIA Blackwell Ultra GPU
- •748 GB of coherent memory
- •Two PCIe 5.0 x4 M.2 slots populated with a 2 TB RAID1 array connected to the CPU
- •Two open PCIe 6.0 x4 M.2 slots connected to ConnectX-8
- •Three PCIe 5.0 expansion slots
- •Dual 400GbE networking, or up to 800 Gb/s aggregate bandwidth
MSI's March announcement states that the system includes high-speed PCIe Gen5 and Gen6 NVMe storage and support for the NVIDIA AI software stack. The product listing specifies Ubuntu 24.04 LTS with NVIDIA AI Developer Tools pre-installed. It also lists a dedicated 1GbE management port, ASPEED AST2600 baseboard management, IPMI 2.0, DMTF Redfish, hardware root of trust, and TPM 2.0 support.
Deskside GB300 deployment
ServeTheHome reported from NVIDIA GTC 2026 that MSI displayed the WS300 in a 1.6 kW power envelope, with liquid cooling for the Grace CPU, Blackwell Ultra GPU, ConnectX-8 controller, and optical cages. The publication described the displayed unit as pre-production and noted that labels on its 400GbE ports did not yet match the final QSFP112 specification.
ServeTheHome also reported a 252 GB GPU-memory specification and 496 GB of Grace LPDDR5X memory, which together match MSI's advertised 748 GB total coherent-memory capacity. MSI's product page does not separately state the GPU and CPU memory allocations.
For ML infrastructure teams, a system with coherent CPU-GPU memory, local NVMe expansion, and 400GbE connectivity can reduce data movement constraints in workflows that combine preprocessing, fine-tuning, and distributed inference. Comparable deskside systems, however, still require teams to account for power delivery, cooling, rack or office networking, and software compatibility before treating them as substitutes for conventional data-center nodes.
Key Points
- 1MSI's WS300 combines a 72-core Grace CPU and Blackwell Ultra GPU in a deskside DGX Station-derived system for AI workloads.
- 2The published 748 GB coherent-memory figure combines CPU and GPU memory, a configuration relevant to memory-bound training and inference pipelines.
- 3Dual 400GbE and PCIe Gen6 NVMe expansion make the workstation relevant to distributed AI environments, subject to local power and networking requirements.
Scoring Rationale
The WS300 is a notable high-end AI workstation configuration, particularly for practitioners evaluating deskside GB300 compute with high-bandwidth networking. Its impact is narrower than a new accelerator or model release.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
