- 8x Blackwell Ultra, 2.1 TB HBM3e
- 144 PFLOPS FP4 sparse, 108 dense
- Turnkey 10U node for SuperPOD and BasePOD
Image: NVIDIA. Manufacturer render; units ship in manufacturer packaging.
NVIDIA, Rack-scale, new
NVIDIA GB300 NVL72
NVIDIA GB300 NVL72 is a fully liquid-cooled rack with 72 Blackwell Ultra GPUs and 36 Grace CPUs in a single NVLink domain, 37 TB of fast memory and 1,440 PFLOPS of FP4 compute with sparsity. Lineage Tech coordinates supply and delivery against a project offer.
Price on written offer
Send quantity and destination. We confirm supply, lead time and a fixed price in writing.
- Exact part number GB300 NVL72. No edition substitution without your written approval.
- Stated per offer: manufacturer, OEM or supplier backed.
- Lead time confirmed in your written offer. Agreed per order (Incoterms 2020).
- 30% deposit to secure the allocation, balance before dispatch. Bank transfer against pro forma invoice, or card payment link for approved orders. Card details are entered on Network International's secure payment page, never on our servers.
- Lead-time orders are secured with a 30% deposit, balance before dispatch. Written offers are valid for the period stated, usually 48 to 72 hours.
Offer terms
- Availability
- Source to order
- Unit price
- Quoted per order
- Minimum order
- 1 unit
- Ships from
- Confirmed in your written offer
- Dispatch
- Lead time confirmed in your written offer
- Incoterm
- Agreed per order (Incoterms 2020)
- Tax and duties
- Quoted excluding UAE VAT, duties and import taxes unless stated
- Payment
- 30% deposit to secure the allocation, balance before dispatch; Bank transfer against pro forma invoice, or card payment link for approved orders
- Condition
- New, factory sealed
Overview
GB300 NVL72 connects 72 Blackwell Ultra GPUs with fifth-generation NVLink at 130 TB/s, so the whole rack behaves like one very large GPU for training and reasoning inference. It carries 20 TB of HBM3e and 17 TB of LPDDR5X, with ConnectX-8 SuperNICs at 800 Gb/s per GPU for scale-out.
It is built by OEMs and ODMs to the NVIDIA reference design and needs facility liquid cooling and high-density power. We confirm facility readiness, delivery route and installation partners before an order is placed.
Typical uses
- Frontier model training
- Large-scale reasoning inference
- AI factory
Specifications
| GPUs | 72x NVIDIA Blackwell Ultra |
|---|---|
| CPUs | 36x NVIDIA Grace (2,592 Arm Neoverse V2 cores) |
| GPU memory | 20 TB HBM3e, up to 576 TB/s |
|---|---|
| CPU memory | 17 TB LPDDR5X, 14 TB/s |
| FP4 Tensor | 1,440 PFLOPS sparse, 1,080 PFLOPS dense |
|---|---|
| FP8 and FP6 Tensor | 720 PFLOPS (with sparsity) |
| FP16 and BF16 Tensor | 360 PFLOPS (with sparsity) |
| NVLink | 5th generation, 130 TB/s |
|---|---|
| Networking | ConnectX-8 SuperNIC, 800 Gb/s per GPU |
| GPU power | Configurable up to 1,400 W per GPU |
|---|---|
| Cooling | Fully liquid-cooled |
Source: manufacturer documentation, checked 16 October 2026. Tensor figures follow NVIDIA's stated basis; sparse figures are marked.
Delivery and warranty
Delivery. Lead time confirmed in your written offer. Agreed per order (Incoterms 2020). Shipments are insured to declared value and tracked. Delivery to every destination is subject to export and sanctions screening; see export compliance.
Warranty. Stated per offer: manufacturer, OEM or supplier backed. Keep serial numbers with your invoice; our warranty and RMA process explains how claims are handled.
Cancellation. Written offers state their validity and cancellation terms. See returns and cancellations.
Questions about the GB300 NVL72
What facility does a GB300 NVL72 need?
How quickly can you confirm availability?
Related guides
Alternatives
- 8x Blackwell Ultra, 2.1 TB HBM3e
- 144 PFLOPS FP4 sparse, 108 dense
- Configured by OEM: Supermicro, Dell, HPE and others
Often bought with
- 800 Gb/s XDR InfiniBand
- PCIe Gen6 x16
- Quantum-X800 and Spectrum-X fabrics


