MSI has begun shipping the XpertStation WS300, a deskside AI workstation built on NVIDIA’s DGX Station architecture and powered by the new GB300 Grace Blackwell Ultra Desktop Superchip. The system is aimed at organisations that want to build, fine-tune and run large AI models on their own hardware instead of renting cloud capacity.
The Grace Blackwell Ultra design puts NVIDIA’s Grace CPU and Blackwell GPU on a single superchip with a shared, coherent memory pool, and the XpertStation WS300 carries that pool up to 748GB. That matters because the usual bottleneck in local AI work is not compute, it is moving model weights back and forth between separate CPU and GPU memory over PCIe. With enough memory attached directly to the superchip, a large foundation model can stay resident on the chip instead of being swapped in and out, which is also the reasoning behind NVIDIA’s separate work expanding custom NVHBM memory across its NVLink Fusion interconnect.
What’s inside the XpertStation WS300
Beyond the GB300 superchip and its memory pool, the WS300 can be linked to a second unit using NVIDIA’s ConnectX-8 SuperNICs, letting two workstations act as one clustered AI system rather than two independent boxes. MSI positions the machine for enterprise AI development, fine-tuning and multimodal inference: the kind of workload that is moving to local infrastructure as companies look to cut cloud dependence and keep model weights and training data in-house. It is not a consumer product, but it extends the rollout of Grace Blackwell hardware from the data centre onto a desk, using a DGX Station reference design that NVIDIA licenses out to system builders rather than selling itself.
Price and availability
The XpertStation WS300 has landed in overseas retail channels at $99,900, a figure reported by Guru3D. MSI is selling the system through enterprise channels, including ASI, D&H and Newegg, rather than as a general consumer purchase.
What that price buys, by the terabyte
At $99,900 for up to 748GB of memory, teqpost calculates that works out to roughly $133,556 per terabyte. That is a useful reference point for anyone weighing a single WS300 against renting equivalent GPU memory in the cloud or assembling a multi-GPU server from discrete parts: it is a steep number on its own, but it buys that memory capacity and bandwidth in one chassis, where matching it with networked, discrete GPUs would take considerably more rack space and cabling.
What to watch
The two-unit clustering option over ConnectX-8 SuperNICs is the detail worth following. If MSI publishes pricing and benchmarks for a clustered pair of WS300 units, that will show whether the system scales to genuinely larger models rather than just fitting more of one model into local memory. It is also worth watching whether other system builders bring their own GB300-based desk-side machines to market now that NVIDIA’s DGX Station architecture has a shipping example to point to.
