Buy NVIDIA DGX Station GB300 | 9 Models for sale | buyzero

The NVIDIA DGX Station GB300 is a desktop AI supercomputer capable of running models of up to 1 trillion parameters locally, on 748 GB of coherent memory and the GB300 Grace Blackwell Ultra Desktop Superchip.

We sell nine DGX Station GB300 models from seven manufacturers, with transparent pricing from €93.277 (excluding VAT). 

NVIDIA DGX Station Specifications

We Sell the Full Range of GB300-Based Systems

NVIDIA does not sell the DGX Station GB300 directly. It is built by OEM partners under their own names, but they all use the same GB300 reference design. The superchip, the memory pool and the theoretical peak performance are identical across every model. Manufacturers ship models by chassis size, case acoustics, storage configuration, networking ports, rack mounting, the option to add a display GPU, price, and the service you get after the sale.

BuyZero carries the full range of GB300-based systems:

• ASUS ExpertCenter Pro ET900N G3

• Dell Pro Max with GB300

• Exxact Valence VWS-158270643, Valence VWS-117032084 and TensorEX TWS-188490844

• Gigabyte W775-V10-L01

• HP ZGX Fury G1n AI Station

• MSI XpertStation WS300 (WS300T60L)

• Supermicro Super AI Station ARS-511GD-NB-LCC

Deploy Large AI Models With Dual-Tier Unified Memory Architecture 

The DGX Station is the first desktop system to bring datacentre grade HBM3e memory to a deskside form factor, so any organization or individual can run large language models and other AI workloads locally.

Peak performance at 1,600 W

DGX Station GB300 uses HBM3e, the same memory generation as the datacentre Blackwell Ultra. High Bandwidth Memory moves data 10 to 20 times faster than DDR5 in a laptop. The speed comes from lane count rather than clock rate, and HBM3e runs 1024 data lanes where DDR5 runs 64. 

It pairs 252 GB of HBM3e at 7.1 TB/s with 496 GB of LPDDR5X at 396 GB/s, joined by an NVLink-C2C coherent bridge at 900 GB/s. 

This architecture means that software can use one 748 GB address space, so massive models can run without sharding or code changes. The primary tier alone is faster than an H200 at 4.8 TB/s yet can be run in any office or building with a standard plug socket.

A 100B model quantised to FP4 or FP8 runs entirely in the fast tier and generates tokens at full speed. Larger models spill into the LPDDR5X tier and slow for the overflow portion, but they still run, which is exactly what a 32 GB consumer GPU cannot do at any speed.

Unparalleled AI performance

The Station gives you 748 GB of coherent memory, 7.1 TB/s of HBM3e bandwidth, MIG partitioning for seven users, and server-grade remote management, which allows it to be a shared instrument for frontier models.

Critically, the GB300 has a memory bandwidth far beyond any consumer-grade device. For example, a Mac Studio runs around 819 GB/s vs the Station's 7.1 TB/s (>8X faster).

GB300 Benchmarks on big LLMs

Mixture-of-experts models (MOE) perform especially well in this environment. Keep active experts in HBM3e and inactive experts in LPDDR5X, with different quantisation per tier, and you can run models that would otherwise need a cluster at close to full speed on one machine that plugs into a standard socket. The DGX is an incredible piece of technology.

An AI Data Centre That Fits Under Your Desk 

Multi-Instance GPU slicing partitions the Blackwell Ultra hardware into up to 7 isolated instances, each with its own memory and compute, so seven users can work at once without one job degrading another.

Container work runs through the NVIDIA Container Toolkit, which integrates with Docker and Kubernetes. Remote management works the way it does on a rack server with a dedicated BMC independent of the host OS, IPMI 2.0 and Redfish for scripted sensor queries and reboots, DCGM for health checks and telemetry into Prometheus, and NVIDIA Fleet Command if you run several Stations across sites.

Software included

Every DGX station comes with DGX OS - an enterprise Ubuntu distribution with drivers pre-installed and kernel parameters tuned for Grace-Blackwell. 

The DGX range also comes with NVIDIA AI Blueprints, which are GPU-optimised reference architectures for common enterprise tasks. A Blueprint package can include the models, containerised NIM microservices, data pipelines, sample frontend code and Helm charts, so a workflow is reproducible from the start.

We also offer NVIDIA AI Enterprise subscriptions. 

Discounts Available

If you are buying a Station for educational use, you may be eligible for an NVIDIA Education (EDU) discount. Startups may be able to avail of the NVIDIA Inception programme, which also offers a discount. Contact us for more information.

Frequently Asked Questions

Which DGX Station model should I choose?

All nine have identical AI performance. They use the same NVIDIA DGX Station motherboard, an NVIDIA-approved cooling solution and a 1,600 W power supply, and NVIDIA controls this tightly. The difference between models is chassis size, acoustics, rack mounting options (Supermicro, HP), storage configuration, networking ports, price, and the service you expect after the sale.

Will my model run at full speed on the DGX station?

That depends on whether it fits in the 252 GB HBM3e tier. Models that fit run at 7.1 TB/s. Models that spill into the LPDDR5X tier serve those weights at 900 GB/s, so generation for the overflow portion is slower. However, quantisation can dramatically modify the results and allow even larger models (including Kimi K3) to run on the DGX Station. If you’d like more information on this, please contact us. 

Can several people use one Station at once?

Yes. MIG partitions the GPU into up to 7 hardware-isolated instances, each with its own memory and compute, so parallel workloads do not degrade each other.

How long does it take to receive a DGX station?

Delivery time varies by model. We can ship models with a lead time of as low as four weeks. 

Can the DGX station run in a normal office?

Yes. The 1,600 W system runs on a standard office socket and uses closed-loop liquid cooling. No server room, special power supply, or external cooling required.

Can I connect two DGX Stations together?

Yes. Each has a ConnectX-8 SuperNIC with two 400 Gb/s ports and RDMA support. Two Stations link directly, without a switch, into one logical system with around 1.5 TB of coherent memory and 40 PFLOPS FP4.

How is the DGX Station different from the DGX Spark?

The DGX Spark is a compact single-user device with 128 GB of unified memory, suited to individual developers and prototyping. The DGX Station GB300 has 748 GB of coherent memory and supports up to seven users through MIG.

How does the warranty work? 

Each DGX Station GB300 is an OEM product, so the hardware warranty runs through the manufacturer of the model you choose.

Do you offer free delivery? 

Delivery for the DGX AI Station is free within Europe, including Switzerland and the UK.

Global free shipping starts at 10 or more units.

Where should I buy the DGX station?

pi3g (buyzero) were one of the first resellers to offer the DGX Station in Europe, with customers in many European locations, for startups, established companies looking for powerful local AI compute, and for research institutes and universities.

Below is a photo of a DGX station we delivered to a customer.

Is there a smaller solution than the DGX Station? 

For smaller office requirements, NVIDIA provides the DGX Spark.