MSI XpertStation WS300 Review: Evaluating Grace Blackwell Deskside AI Performance and Sourcing Requirements

The MSI XpertStation WS300 represents a shift in local AI hardware, packaging enterprise-grade training and inference capabilities into a deskside form factor built around the NVIDIA DGX Station architecture. Powered by the GB300 Grace Blackwell Ultra desktop superchip, the system combines a 72-core Grace CPU with a Blackwell Ultra GPU, linked via a 900 GB/s NVLink-C2C interconnect. This architecture features a unified 748GB coherent memory pool—merging 252GB of HBM3e with 496GB of LPDDR5X system memory—allowing developers to run large-scale foundation models directly on-premises without cloud serialization latency.

From a component and supply-chain perspective, integrating high-density accelerators and large unified memory pools into a localized workstation creates significant engineering hurdles. The system relies on advanced liquid-cooling blocks spanning the processor, memory, and onboard NVIDIA ConnectX-8 SuperNICs to sustain its 1.6kW power envelope. Hardware engineering teams specifying similar high-performance local AI development nodes must account for dedicated high-wattage titanium power supplies, robust thermal dissipation strategies, and strict export compliance paperwork required for advanced accelerator architectures.

For enterprise deployment, the workstation also integrates dual 400GbE networking ports via the ConnectX-8 SuperNIC, enabling multiple units to cluster together for expanded local scaling. Hardware buyers and FAEs should evaluate how deskside supercomputing infrastructure impacts on-premises data privacy mandates, local agentic AI execution frameworks like NVIDIA OpenShell, and overall capital expenditure versus recurring cloud infrastructure fees.