Key Announcement: MSI Unveils $99,999 AI Workstation, Now Available for Pre-order
MSI has officially begun shipping its new AI workstation, the XpertStation WS300, to overseas markets as of August 28, 2026. Priced at $99,999 (approximately CNY 674,000 at current exchange rates), the unit is no prototype but actually available for purchase via Newegg.
Key specifications and availability facts:
- Release date: August 28, 2026 (today)
- Pricing: $99,999 (retail on Newegg)
- Core processor: NVIDIA GB300 Grace Blackwell Ultra desktop superchip
- Maximum memory: 748GB coherent memory
- Network interface: Dual ConnectX-8 SuperNIC, 400GbE Ethernet
- Scalability: Supports linking two units in series
- Storage architecture: PCIe Gen5/Gen6
- Model capability: Runs open models up to 1 trillion (1T) parameters
Technical Deep Dive: Desktop-Sized Supercomputing Power
The WS300 centers on NVIDIA’s GB300 Grace Blackwell Ultra chip—the first time this DGX-class architecture has been packaged for desktop deployment. It combines the Grace CPU and Blackwell GPU architectures via the Cobalt 700 platform to deliver unified memory access. The 748GB coherent memory configuration eliminates CPU-GPU data copying overhead, accelerating both inference and training workloads.
Network-wise, the system integrates dual Nvidia ConnectX-8 SuperNICs, each capable of 400 Gigabit Ethernet. This enables not only high-bandwidth single-machine operations but also CPU-scale interconnect for linking two WS300 units, forming a minimal 2-node cluster without requiring InfiniBand infrastructure.
Storage leverages a hybrid Gen5/Gen6 PCIe architecture (some lanes supporting up to 128 GT/s), paired with NVMe SSDs to sustain large model checksum loads and rapid parameter updates. An unexpected specification difference stands out: 748GB of memory dwarfs typical high-end workstations, which usually cap at 256–512GB. This brings workstation-class machines into the territory previously reserved for rack-mounted servers.
Target Use Cases and Ecosystem Support
Beyond standard AI model training, the WS300 integrates with NVIDIA NemoClaw, NVIDIA’s framework for deploying and governing AI agents. NemoClaw provides policy controls and sandboxing capabilities, enabling enterprises to run autonomous AI agents with controlled decision boundaries—ideal for secure use in customer service, code review, and compliant data analysis.
The product targets three primary user profiles:
- University AI labs and research institutes needing rapid iteration on large models but lacking budget for full DGX deployments
- Enterprise AI innovation teams requiring in-house 1T-parameter model validation before scaling
- Supercomputing centers supplementing distributed workloads with mobile edge nodes
Performance Comparison (Official Specifications Only)
| Feature | XpertStation WS300 | Standard High-End Workstation | NVIDIA DGX Station A100 |
|---|---|---|---|
| Core Chip | GB300 Grace Blackwell Ultra | RTX 6000 Ada / RTX 4090 | 8× A100 80GB |
| Max Memory | 748GB coherent | 256–512GB | 640GB |
| Network | Dual 400GbE (ConnectX-8) | 10/25GbE Ethernet | 4× 100GbE InfiniBand |
| Storage Bus | PCIe Gen5/Gen6 | PCIe Gen4 | PCIe Gen4 |
| Model Support | ≤1T parameters (single-node) | ≤10B (consumer) / ≤100B (pro) | ≤10B (single-node, requires distributed training for larger) |
| Price | $99,999 | $10,000–$40,000 | $200,000+ |
Who Should Buy Now—and Who Should Wait?
Buy Now If You:
- Plan to deploy 1T-parameter open models (e.g., Llama 3.1 405B, Qwen 3 preview) in Q4 and lack GPU cluster access
- Operate in highly regulated industries (finance, healthcare) requiring on-premises, air-gapped AI inference
- Value unified memory and ultra-low network latency enough to pay a premium over commodity hardware
Wait and Observe If You:
- Have budgets under $500,000 and only need routine LLM fine-tuning (<10B parameters)—current $10K–$40K workstations suffice
- Rely on rapidly evolving open-source toolchains—GB300 driver ecosystems remain nascent, with partial OSS support pending
- Are planning to scale beyond 10+ nodes—consider waiting for DGX SuperPod or cloud-optimized fleet solutions
Final Thoughts
MSI’s collaboration with NVIDIA signals a strategic shift: AI compute is being redefined at the desktop layer. By packing 748GB of coherent memory and 1T-parameter support into a single chassis, this workstation blurs the line between workstation and server—offering supercomputer-level capability without sacrificing mobility. It hints at a promising new paradigm for AI development: powerful, stand-alone clusters in a console form factor.