MSI XpertStation WS300 is a desktop workstation built on the NVIDIA DGX Station architecture, priced at around $99,900-100,000 (Newegg retail, US); it visually lacks a conventional graphics card because the GPU is fused with the CPU into a single module. Headlines like "a computer costs as much as an apartment and ships without a graphics card" spread across tech media in 2026 for a reason.
Key takeaways:
- Price is around $99,900-100,000 (Newegg retail, US); platform is NVIDIA GB300 Grace Blackwell Ultra.
- There's no discrete graphics card: CPU and GPU are fused into one module.
- Unified CPU+GPU memory totals 748 GB (252 GB HBM3e + 496 GB LPDDR5x).
- Performance is 20 petaflops FP4; ConnectX-8 SuperNIC networking reaches 800 Gbps for clustering.
- Base configuration ships with just 2 TB of SSD storage; expansion drives are sold separately.
Why a $100,000 computer "has no graphics card"
The irony comes from one detail: the XpertStation WS300 has no discrete graphics card, meaning no separate board that plugs into a PCIe slot with fans sticking out of the case. The graphics accelerator is very much present, and it's one of the most powerful on the market - it's just not packaged as a separate board.
The reason is the NVIDIA GB300 Grace Blackwell Ultra architecture. It's a "superchip" where the CPU and GPU are fused at the module level instead of connected through an expansion bus. The 72-core ARM-based NVIDIA Grace processor works alongside the DGX B300 accelerator built on Blackwell Ultra with 288 GB of HBM3e memory. There's no graphics card in the sense a gaming PC has one, but there is a solution more powerful than most retail cards, just assembled differently.
748 GB of shared memory - and what it actually means
The station's main engineering breakthrough isn't chip power, it's memory design. CPU and GPU share a single coherent pool here: 252 GB of HBM3e (ultra-fast memory located right next to the GPU) plus 496 GB of LPDDR5x, the processor's memory. That's 748 GB total, and both processors see it as one address space. There's no need to copy data between "GPU memory" and "system memory" the way regular PCs do.
A gaming PC gets nothing from this. But for the tasks the station was built for - training and running large language models - it's decisive. AI models with billions of parameters run into a hard limit: the memory available to the GPU. If a model doesn't fit into GPU memory, it has to be split into parts or data has to move back and forth over a slow bus, losing speed. That problem disappears here, and a model that once needed several server-grade GPUs fits entirely into one station's memory.
20 petaflops on a developer's desk
A single Blackwell Ultra accelerator delivers 20 petaflops of FP4 compute (a low-precision data format widely used for inference in modern neural networks). Just a few years ago, numbers like that belonged to server racks in data centers, not a tower sitting under a developer's desk.
The idea behind XpertStation WS300 is to move onto a local machine work that used to require access to a cloud data center or cluster. Training and fine-tuning large language models, running inference for generative AI systems - all of that usually turns into a bill for rented cloud GPU capacity. Heavy data science computations fall into the same category. Now that work can be done locally, without sending data outside the company. For companies dealing with sensitive information, or simply tired of cloud bills, the argument needs no further explanation.
What's inside: MSI XpertStation WS300 specifications
| Parameter | Value |
|---|---|
| Platform | NVIDIA GB300 Grace Blackwell Ultra (NVIDIA DGX Station architecture) |
| Processor | NVIDIA Grace, 72-core ARM |
| Graphics accelerator | NVIDIA DGX B300 (Blackwell Ultra), 288 GB HBM3e |
| Unified memory | 748 GB (252 GB HBM3e + 496 GB LPDDR5x), shared coherent CPU+GPU pool |
| Performance | 20 petaflops FP4 (single Blackwell Ultra accelerator) |
| Storage | 2 TB SSD standard, 4 M.2 slots for expansion |
| Networking | NVIDIA ConnectX-8 SuperNIC, up to 800 Gbps, 400GbE ports |
| Cooling | Liquid |
| Power supply | 1600 W |
| Form factor | Deskside tower |
| Price | Around $99,900-100,000 (Newegg retail, US) |
2 TB of storage for that money - and why it's not a nitpick
Against all the other numbers, one spec line looks downright modest: the base configuration comes with just 2 TB of SSD. Datasets used by teams training models often don't fit into that space, so more storage has to be bought separately. Expansion is possible via 4 M.2 slots, but the drives cost extra on top of the $100,000 price tag.
800 Gbps networking: stations combine into a cluster
A single station is powerful on its own, but the XpertStation WS300 also has networking ambitions. The NVIDIA ConnectX-8 SuperNIC adapter delivers up to 800 Gbps through 400GbE ports. Several machines can be combined into a small cluster and exchange datasets and model parameters at speeds regular networking hardware can't match. A lab or a small AI team can build its own mini data center out of several WS300 units and skip cloud providers entirely.
A computer priced like an apartment: where the headline irony comes from
The roughly $100,000 price tag is comparable to the cost of an apartment in many European cities, and that contrast - "a supercomputer priced like real estate" - is exactly what spread across headlines alongside the joke about the missing graphics card. Looking at the market for server-grade AI GPU solutions, though, the price makes sense: comparable computing power in the form of server-grade NVIDIA cards for data centers costs about the same or more. Here, everything is packed into a single deskside tower with liquid cooling that can sit in an office next to a desk.
The station was shown at Computex/GTC 2026 and is already available through MSI partners: ASI, D&H, and Newegg. It's not meant for gamers or regular users, and it's not within their budget. Demand comes from research labs and AI startups that need their own local compute without depending on the cloud. What's notable here is different: computing density that recently required a server room now fits into a case the size of a desktop tower, even if it costs as much as real estate.
How much does the MSI XpertStation WS300 cost?
US retail price (Newegg) is around $99,900-100,000 for the base configuration with 2 TB of SSD storage; expansion drives are sold separately.
Why doesn't the MSI XpertStation WS300 have a graphics card?
It does have a graphics accelerator, it's just not packaged as a separate PCIe board: the NVIDIA GB300 Grace Blackwell Ultra architecture fuses the Grace processor and the DGX B300 accelerator into a single module with shared memory.
What is the 748 GB of unified memory for?
That much shared CPU and GPU memory lets a single machine fully host large language models that previously required several server-grade GPUs, without splitting the model into parts.
Can several stations be combined into a cluster?
Yes: the NVIDIA ConnectX-8 SuperNIC adapter, with up to 800 Gbps of bandwidth and 400GbE ports, lets several WS300 units be connected into a small local data center.
Who is the MSI XpertStation WS300 for?
The station is built for research labs and AI startups that need local compute power for training and running inference without relying on cloud GPU capacity; it's not intended for regular users or gamers.