Desktop MSI AI Supercomputer — GB10 Grace Blackwell, 128GB Unified Memory

A compact desktop AI supercomputer built on the GB10 Grace Blackwell Superchip, running large models up to 200B parameters locally.

$ 5 999
United States (USD)

Change currency

Prices follow the country you buy from. Pick a country to switch the currency.

Shipping calculated separately
Lead time: 2–4 weeks

This is a desktop-class AI supercomputer for teams that want to prototype, fine-tune and serve large language models on their own hardware instead of renting cloud GPU time. The GB10 Grace Blackwell Superchip pairs a 20-core Arm v9 CPU with a Blackwell GPU over an NVLink-C2C interconnect, so the CPU and GPU share one coherent pool of 128GB LPDDR5X unified memory at 273GB/s — no copying tensors back and forth across a PCIe bus.

It runs inference on models up to 200B parameters locally and fine-tunes up to 70B. Two units clustered over the built-in ConnectX-7 200Gbps RDMA NIC scale distributed training to 405B parameters, which lets a lab grow capacity one node at a time. The Blackwell GPU carries 6,144 CUDA cores and 5th-generation Tensor Cores rated at 1,000 AI TOPS (FP4), and the board ships with DGX OS (Ubuntu 22.04) and a pre-installed AI stack for PyTorch, TensorFlow and RAPIDS.

  • GB10 Grace Blackwell Superchip — 20-core Arm v9 CPU (10× Cortex-X925 + 10× Cortex-A725) plus Blackwell GPU over NVLink-C2C
  • 128GB LPDDR5X coherent unified memory, 256-bit interface, 273GB/s bandwidth
  • 1TB or 4TB self-encrypting NVMe M.2 SSD
  • ConnectX-7 Smart NIC (200Gbps RDMA), 10GbE RJ45, Wi-Fi 7, Bluetooth 5.3
  • 4× USB4 Type-C (40Gbps), HDMI 2.1a with 8K output
  • Secure Boot, ECC memory and Arm TrustZone confidential computing

At 150 × 150 × 50.5mm and 1.2kg with passive, silent cooling, it sits on a desk and draws about 170W in typical use. An optional kit makes it rack-mountable. Backed by a 3-year manufacturer warranty.

Specifications

Dimensions (L × W × H)
150 × 150 × 50.5 mm
Weight
1.2 kg
Power
170 W
Warranty
3-year manufacturer warranty
Chipset
GB10 Grace Blackwell Superchip
CPU
20-core Arm v9 (10× Cortex-X925 + 10× Cortex-A725)
GPU architecture
Blackwell
CUDA cores
6144
Tensor Cores
5th-generation, 1,000 AI TOPS FP4
RT Cores
4th-generation
Unified memory
128GB LPDDR5X
Memory bandwidth
273 GB/s
Memory interface
256-bit
Storage
1TB or 4TB self-encrypting NVMe M.2 SSD
Networking NIC
ConnectX-7 Smart NIC, 200Gbps RDMA
Ethernet
10GbE RJ45
Wireless
Wi-Fi 7 (802.11be), Bluetooth 5.3
USB
4× USB4 Type-C (40Gbps)
Video output
1× HDMI 2.1a (8K)
Peak power
224 W
Cooling
Passive airflow, silent
Form factor
FHFL dual-slot, rack-mountable with optional kit
Operating system
DGX OS (Ubuntu 22.04-based)
Local inference
Up to 200B parameters
Fine-tuning
Up to 70B parameters
Distributed training
Up to 405B parameters (dual-node)
Security
Secure Boot, ECC memory, Arm TrustZone
Screen size
19 in
Mouse & keyboard
Included

Frequently asked questions

What size AI models can this run locally?

The 128GB of coherent unified memory lets it run inference on models up to 200B parameters on a single unit, for example LLaMA 2 or DeepSeek-R1. It fine-tunes models up to 70B parameters. For larger work, two units clustered over the 200Gbps RDMA link scale distributed training to 405B parameters.

How does the unified memory differ from a normal GPU card?

The GB10 Grace Blackwell Superchip connects the Arm CPU and Blackwell GPU with an NVLink-C2C interconnect and gives them one coherent 128GB pool at 273GB/s. Both processors read the same memory without copying tensors across a PCIe bus, so you can load far larger models than a discrete GPU with separate VRAM would hold.

What operating system and AI frameworks come pre-installed?

It ships with DGX OS, an Ubuntu 22.04-based build, and a pre-installed AI Enterprise stack including cuDNN and TensorRT. PyTorch, TensorFlow and RAPIDS run out of the box, and you get access to the NGC container registry for additional images.

Can two units be clustered together?

Yes. Each unit carries a ConnectX-7 Smart NIC with 200Gbps RDMA for clustering. Linking two nodes lets you scale distributed training to 405B parameters, so you can grow capacity one node at a time rather than replacing the whole system.

How much power does it draw and how loud is it?

It draws about 170W in typical operation and up to 224W under heavy load, so a standard desk outlet is enough. The passive airflow design has no fans dedicated to noise, which keeps it silent on a desk or in an office.

What storage and security features are included?

Storage is a self-encrypting NVMe M.2 SSD in 1TB or 4TB — specify the capacity when ordering. Security covers Secure Boot, ECC memory protection and Arm TrustZone confidential computing, which suit teams handling sensitive data or regulated workloads.

Is shipping included in the price?

For in-stock items — anything you can add to the cart — shipping is included in the listed price. For made-to-order items, delivery is either free (marked on the product page) or calculated when you place your order, based on destination and volume.

How long does delivery take?

In-stock items typically arrive within one to three weeks. Made-to-order items take three to six weeks, since we work directly with manufacturers and need time for production and logistics before shipping. Your assigned manager confirms a more precise estimate once your order is placed. Please note we can't be responsible for delays caused by customs clearance in the destination country.

Are import duties and taxes included?

In most cases, yes — we aim to ship Delivered Duty Paid (DDP), which means duties and import taxes are handled so you don't face extra charges on delivery. In rare cases we may ship Delivered Duty Unpaid (DDU), where you would settle import charges directly with the carrier — but only when agreed with you in advance.

Do your products come with a warranty?

Yes. Most of our equipment comes with a one-to-two-year manufacturer's warranty covering factory defects and technical faults under normal use; the specific term is stated on each product page. If a defect appears, we arrange a replacement or refund in line with our Exchange & Return Policy, and our team guides you through the claim step by step.

Request a quote