NVIDIA DGX buying guide
NVIDIA DGX GB300 is rack-scale infrastructure for large AI training and inference deployments, not a consumer desktop computer. NVIDIA describes a GB300 NVL72 rack containing 72 Blackwell Ultra GPUs and 36 Grace CPUs, with 20TB of GPU memory and 37TB of total fast memory. If you are considering a deployment, the first questions concern rack power, liquid cooling, networking, data throughput and enterprise support.
Use this guide and the rack calculator to frame a procurement discussion, then obtain the actual design and pricing from NVIDIA or an authorized partner.
Quick answer
What do you need before buying DGX GB300?
You need a facility-ready design, not just the accelerator specification. DGX GB300 is a liquid-cooled rack-scale system with 72 Blackwell Ultra GPUs and 36 Grace CPUs in NVIDIA's published configuration. Confirm power distribution, liquid cooling, supported rack footprint, networking and storage with the deployment partner.
NVIDIA lists DGX GB300 as available for enterprise deployment; it is not a standard Amazon retail purchase. The marketplace listings below, where available, are individual supporting components only.
Enterprise procurement
Purchase complete systems through an enterprise partner
DGX GB300 racks and SuperPOD clusters require formal procurement, facility planning and vendor integration. The retail listings below, if any, are supporting hardware only.
Products and buying options
Supporting hardware to consider
Compare complete DGX computers separately from similar GB10 systems and optional supporting equipment. Confirm the configuration and current checkout price before ordering.
How to decide
Treat the entire rack as the unit of planning
A DGX GB300 purchase affects the site electrical plan, cooling loop, network fabrics, data storage and operational resilience. The published 20TB GPU memory figure is per complete rack configuration, not per card. Calculated rack power is a planning assumption until the integrator confirms actual equipment ratings and operating profile.
Use the tool to organize requirements, but have qualified specialists sign off on electrical and cooling infrastructure.
Decision table
DGX GB300 rack-scale published specifications
| Item | Current planning value | How to use it |
|---|---|---|
| Accelerators | 72 NVIDIA Blackwell Ultra GPUs | Use the exact rack generation when converting workload demand into racks. |
| CPUs | 36 NVIDIA Grace CPUs | Keep host-side services and CPU-memory needs in the rack plan. |
| GPU memory | 20 TB per DGX GB300 rack | Use for accelerator working-set planning, not as storage capacity. |
| Total fast memory | 37 TB per rack | Keep GPU memory and total fast memory as different figures. |
| Cooling | Rack-scale, liquid-cooled system | Validate facility water, CDU and OEM thermal requirements before deployment. |
Interactive planning tool
DGX GB300 Rack Planning Screen
Enter your planned workload to see an estimated memory or infrastructure requirement. Confirm results with the current manufacturer specifications and a real-world test before purchasing.
Before you buy
Four checks before you buy
Start with current documentation
Use the exact platform or OEM system guide as the source of truth for supported configurations and limits.
Keep assumptions visible
Every calculator input is an assumption until it is replaced by a measurement, vendor limit or facility design value.
Separate nameplate from application performance
Port speed, SSD peak rate, GPU memory and power ratings do not guarantee end-to-end workload results.
Escalate facility decisions
High-voltage distribution, rack electrical work, cooling design and liquid loops require qualified professionals and current codes.
Detailed buying guide
Understand the GB300 NVL72 configuration
NVIDIA's DGX GB300 NVL72 platform groups 72 Blackwell Ultra GPUs and 36 Grace CPUs in a high-density rack-scale AI system. The published memory capacity includes about 20TB of GPU memory and 37TB total fast memory. These are system-level figures, not specifications for a single workstation.
Verify the exact system generation and software release in an OEM quotation. Different NVIDIA platforms share terms such as GB300 and Blackwell, but their installed GPU count, network design and physical footprint can differ.
Plan liquid cooling first
NVIDIA describes DGX GB300 as a fully liquid-cooled rack architecture. That changes what a facility must provide: coolant distribution, flow, supply conditions, redundancy, maintenance access and trained operational staff all matter before hardware arrives.
Do not order a rack based on floor space and electrical nameplate alone. Request the latest mechanical and cooling specification from the vendor and have facilities personnel validate it against the site. A general-purpose room air conditioner is not a replacement for the required liquid cooling design.
Estimate power and facility overhead
Dense Blackwell Ultra racks require a deliberate power-distribution design. The calculator includes a user-entered rack kilowatt figure and PUE for preliminary facility load planning; that default is not a guaranteed measured draw for the hardware shipped to you.
Use vendor-provided peak, sustained and fault-condition requirements to choose feeds, redundancy and monitoring. High-voltage or rack electrical installation must be handled by qualified professionals according to applicable codes.
Networking and compute fabrics
Large AI jobs depend on fast communication between accelerators and across racks. NVIDIA specifies ConnectX-8 and BlueField networking in DGX GB300 platform materials, alongside the NVLink switch system inside the rack. Those interfaces serve different traffic and management roles.
Plan separate compute, storage and management requirements with the integrator. Validate topology and transceiver compatibility rather than choosing generic high-speed switches or cables solely from port-speed labels.
Storage for model data and checkpoints
AI training and enterprise inference can move large datasets and checkpoints. Storage must deliver throughput under realistic concurrency and failure conditions, not only enough capacity in terabytes. The correct design depends on whether data is streamed, cached locally or staged from shared storage.
The NVMe products shown below are possible infrastructure components, not certified drop-in parts for a DGX GB300 rack. Confirm supported SSDs, storage appliances, namespace design and network paths in the integrator's bill of materials.
Software, management and support
NVIDIA lists Mission Control and NVIDIA AI Enterprise among DGX GB300 platform software, with enterprise support offerings. Those capabilities require the correct licensing, security design and operational skills. Treat software operations as part of the platform purchase.
Clarify who owns orchestration, monitoring, patching, incident response, backup and component replacement. A fast rack without a reliable operational model can become less useful than a smaller, well-managed deployment.
How to obtain meaningful pricing
A request for 'DGX GB300 price' should result in a scoped system quote, not an isolated GPU price multiplied by 72. Ask for the rack configuration, fabric, storage, liquid-cooling integration, commissioning, software, support, delivery and ongoing power costs.
NVIDIA's product site directs enterprises to sales and partners. Amazon search results can help compare supporting commercial switches or power hardware where appropriate, but they are not quotations for a complete NVL72 rack.
When a smaller system makes more sense
If the main need is local model development, a DGX Station or DGX Spark may be far more practical than a data-center rack. A cloud GB300 instance can also be useful to validate a workload before committing to on-premises infrastructure.
Move to DGX GB300 only when model scale, utilization, data governance and cost justify the facility and operating commitment. Include a measured proof-of-concept and partner-led design review in the purchase process.
Methodology and official references
Specifications for the DGX GB300 rack are based on NVIDIA's current DGX GB300 product page and datasheet.
Power, PUE, storage and network figures entered in the calculator are planning assumptions rather than certification of an installation.
Complete rack procurement requires a partner quotation and site engineering review; retail component listings are deliberately separate from enterprise system purchasing.
As an Amazon Associate, Cloudzat may earn from qualifying purchases. Amazon listings are optional buying choices, not proof of compatibility. Check the exact model, firmware, software, interfaces, electrical requirements and warranty with the manufacturer before purchasing.
Frequently asked questions
How many GPUs are in DGX GB300 NVL72?
NVIDIA specifies 72 Blackwell Ultra GPUs and 36 Grace CPUs in the published rack configuration.
How much GPU memory does DGX GB300 provide?
NVIDIA lists 20TB GPU memory and 37TB total fast memory for the complete rack configuration.
Is DGX GB300 a desktop AI workstation?
No. It is rack-scale enterprise AI infrastructure, distinct from deskside DGX Station.
Can I buy a complete DGX GB300 on Amazon?
The appropriate route is an NVIDIA or authorized enterprise partner quote. Marketplace supporting parts are not complete GB300 systems.
Does DGX GB300 require liquid cooling?
NVIDIA describes the rack architecture as liquid-cooled. Confirm the exact cooling and CDU requirements with the integrator.
Can I use the calculator figure for electrical installation?
No. It is an early estimate; use current vendor engineering specifications and qualified facility professionals.
What software comes with DGX GB300?
NVIDIA lists Mission Control, NVIDIA AI Enterprise and DGX OS/Ubuntu support; confirm entitlement and versions in the quote.
Should my team buy DGX Station instead?
Possibly, if the main need is large-model development at a deskside location rather than rack-scale training or serving.