Professional HardwareMore Performance

New · Teklif ile tedarik

NVIDIA DGX GB300

Brand: NVIDIA SKU: AI-NVIDIA-DGX-GB300

The NVIDIA DGX GB300 is a liquid-cooled rack-scale AI system that combines 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single NVLink domain. It is designed for reasoning models, agentic AI and very large-scale inference. It is not sold as a single product but planned and quoted on a project basis together with the data centre infrastructure.

  • Garanti ve servis koşulları teklif aşamasında üretici ve tedarik kanalına göre netleşir.
  • Delivery and installation terms are finalised at the quotation stage
  • Your card details are not stored on our site

The NVIDIA DGX GB300 is a liquid-cooled rack-scale AI system that combines 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single NVLink domain. It is designed for reasoning models, agentic AI and very large-scale inference. It is not sold as a single product but planned and quoted on a project basis together with the data centre infrastructure.

NVIDIA DGX GB300 overview

The DGX GB300 is a DGX-branded rack-scale system based on NVIDIA's GB300 NVL72 architecture. The rack forms an NVLink domain in which 72 GPUs behave like one giant GPU.

Hardware architecture

Each rack contains 72 NVIDIA Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs; it is a fully liquid-cooled, rack-scale architecture. According to NVIDIA's rack data, GPU memory increases from 13.4 TB in the GB200 generation to approximately 20 TB. Exact performance, power and network figures are determined by NVIDIA's current specification and the project design.

Use cases in AI workloads

Training and inference of trillion-parameter models, reasoning models, agentic AI services and AI factories.

Business advantages

With the NVIDIA DGX software stack and Mission Control, rack-level management, installation and monitoring are standardised.

Scalability and infrastructure requirements

It requires liquid cooling infrastructure (CDU), high power-density racks and a data centre design. Third-party sources put rack power in the range of approximately 120–142 kW; NVIDIA's official rack power figure could not be verified in this work. It scales across many racks as a DGX SuperPOD.

Who is the NVIDIA DGX GB300 for?

Suitable for AI cloud providers, organisations developing large-scale models, national research centres and operators of high-volume inference services.

Key features

  • 72 × NVIDIA Blackwell Ultra GPU, 36 × NVIDIA Grace CPU
  • Single NVLink domain (NVL72)
  • Fully liquid-cooled rack-scale architecture
  • Increased GPU memory compared with the GB200 generation (~20 TB per rack)
  • Multi-rack scaling with DGX SuperPOD
  • Project-based design and installation

Technical Specifications

MarkaNVIDIA
ModelDGX GB300 (NVL72 rack-scale system)
Product typeRack-scale liquid-cooled AI system
MimariNVIDIA GB300 NVL72 (Grace Blackwell Ultra)
GPU72 × NVIDIA Blackwell Ultra
CPU36 × NVIDIA Grace
GPU memory (rack)~20 TB (NVIDIA raf verisi; GB200: 13,4 TB)
CoolingLiquid cooling
Form factorRack-scale system
Rack powerCould not be verified with the manufacturer in this work; determined in the project design
SoftwareNVIDIA DGX software stack, NVIDIA Mission Control

Procurement and corporate quotation process

The NVIDIA DGX GB300 is supplied on a project basis: first your data centre's power, cooling and network infrastructure are assessed together, then a corporate quotation covering the number of racks, installation and support scope is prepared. Price and delivery schedule are finalised in the quotation. Stating your intended use and delivery country in the quote form speeds up the process.

Similar Products