$99,900. Desktop form factor. Runs open-source LLMs with up to 1 trillion parameters — specs that, three years ago, would have required an entire server room.
According to an August 28 report by IT Home, the MSI XpertStation WS300 has started shipping overseas and is on sale at Newegg, powered by NVIDIA’s GB300 Grace Blackwell Ultra desktop superchip with up to 748GB of coherent memory. This is not a prototype — it’s an in-stock product sitting in regular retail channels.
1. What this machine is
The hard facts:
- Release date: August 28, 2026; priced at $99,999 (available on Newegg);
- Main chip: NVIDIA GB300 Grace Blackwell Ultra — the first time NVIDIA has brought DGX-class technology down to a desktop form factor;
- Memory: up to 748GB of coherent memory;
- Networking: dual ConnectX-8 SuperNICs, 400GbE per port;
- Expansion: two units can be linked into a 4-node mini cluster;
- Storage: hybrid PCIe Gen5 / Gen6 architecture;
- Model capacity: a single unit can run open-source LLMs with up to 1T (one trillion) parameters.
2. How can a desktop run trillion-parameter models
The original coverage only said “unified memory reduces copy overhead” — it skipped the most important layer.
1. 748GB is the threshold. Even with quantization, a trillion-parameter model demands memory measured in hundreds of GB — 748GB of coherent memory is precisely the prerequisite that makes the desktop viable. Mainstream consumer workstations typically ship with 64–512GB; the gap isn’t a notch, it’s crossing the line between “can it even hold the model” and “it can’t.”
2. The essence of unified memory is eliminating data movement. On the Cobalt 700 platform, the GB300 hangs the Grace CPU and the Blackwell GPU off the same pool of coherent memory — CPU and GPU share one physical memory space. In traditional architectures, once model weights exceed VRAM, you either shuttle them back and forth between system memory and VRAM, or you compress them with quantization. Unified memory makes that problem disappear in a desktop form factor — not by stacking more VRAM, but by using a shared memory pool to eliminate the movement overhead. That is the real mechanism that lets a desktop shoulder large models.
3. The sinking logic. NVIDIA has stuffed DGX technology into a desktop form factor; compute is sinking from the supercomputing center to a single deployable machine. Link two units into a 4-node cluster and you can assemble the skeleton of a small supercomputer right beside your desk. This points at where next-generation AI development platforms are heading: small but mighty single-machine clusters.
3. What to compare it against
| Item | XpertStation WS300 | Traditional high-end workstation | NVIDIA DGX Station A100 |
|---|---|---|---|
| Main chip | GB300 Grace Blackwell Ultra | RTX 6000 Ada / 4090 | 8× A100 80GB |
| Max memory | 748GB coherent memory | 256–512GB | 640GB |
| Networking | Dual 400GbE | 10/25GbE | 4× 100GbE IB |
| Model support | Up to 1T parameters on one box | ≤10B (consumer) / ≤100B (pro cards) | ≤10B on one box (bigger needs distributed) |
| Price | $99,999 | $10,000–$40,000 | $200,000+ |
The most informative contrast is in the last row: at less than half the price of a DGX, the WS300 runs larger models on a single machine — and it does so not with more expensive cards, but with a unified memory architecture.
4. Who should buy now, who should wait
- Buy now: teams planning to ship trillion-parameter open-source models in Q4 that lack an in-house GPU cluster; compliance-heavy scenarios like finance and healthcare with hard requirements for data locality and offline operation; research institutions willing to pay a premium for unified memory and ultra-fast networking.
- Wait: small and mid-sized businesses with budgets under ¥500K that only need routine fine-tuning of <10B models (a ¥50K–100K workstation already suffices); developers who are sensitive to rapid model version churn and depend on mature toolchains — the GB300 driver stack is still young, and parts of the open-source toolchain are not fully adapted yet; large teams building clusters beyond 10 nodes should hold out for DGX SuperPOD-class solutions.
5. Verdict
Compute is sinking from the supercomputing center down to a single machine at the desk — 748GB of unified memory turns “trillion-parameter models on a desktop” from a gimmick into something feasible. But the maturity of the driver stack and toolchain means that, right now, it’s closer to “early deployment by research institutions” than “standard kit for small and mid-sized teams.” The direction of the next-generation AI development platform has already poked its head out: small but mighty single-machine clusters.