Prime Cell P72
One rack, one NVLink domain.
A complete rack-scale system in a twenty-foot module: seventy-two GPUs that work as one, with the coolant distribution, fabric, storage and power to run them. GB300 NVL72 today, Vera Rubin NVL72 in the Vera Rubin edition.

Styling concept. Configuration not finalised.
StatusGB300 NVL72 shipping, allocation constrained. Vera Rubin edition: production ramp, allocation and certification.
Questions first? Ask the team about Prime Cell P72
A closer look.
The whole rack-scale cell on one front line, beside its cooling package.
Styling concepts from the HYPERCUBE Product Family Design Collection, Edition 3. Configuration not finalised; renders do not depict a supplied configuration.
Two editions.
Blackwell Ultra today and Vera Rubin next, in the same module and the same layout.
| Specification | GB300 edition | Vera Rubin edition |
|---|---|---|
| Rack | GB300 NVL72 | Vera Rubin NVL72 |
| GPUs | 72 Blackwell Ultra B300 | 72 Rubin |
| CPUs | 36 Grace: 2,592 Arm Neoverse V2 cores | 36 Vera: 3,168 Arm cores |
| FP4 inference | 1.44 EF sparse | 3.6 EF NVFP4 |
| FP4 dense or training | 1.08 EF dense | 2.5 EF NVFP4 training |
| FP8 | 720 PF | About 1.26 EF: 72 × 17.5 PF per GPU |
| GPU memory | 20 TB HBM3e at 576 TB/s | 20.7 TB HBM4 at 1.6 PB/s |
| CPU memory | 17 TB LPDDR5X at 14 TB/s | 54 TB LPDDR5X |
| NVLink | NVLink 5: 130 TB/s; 1.8 TB/s per GPU | NVLink 6: 260 TB/s; 3.6 TB/s per GPU |
| Scale-out | 800 Gb/s per GPU, ConnectX-8 | 1.6 Tb/s per GPU, ConnectX-9 |
| Rack power | 135 kW TDP; 155 kW peak | About 190 to 230 kW, third-party figure |
| Facility | 170 to 220 kW envelope | About 220 to 280 kW, planning |
| Cooling | Liquid: CDU and dry cooler | 100 per cent liquid at 45 °C inlet |
| Storage | 0.5 to 1.0 PB usable NVMe | 0.5 to 1.0 PB usable NVMe |
| Status | Shipping; allocation constrained | Production ramp; allocation and certification |
Vera Rubin rack power is a third-party figure; NVIDIA has not published an official rack power. Vera Rubin FP8 is derived from NVIDIA's per-GPU figure. The facility range for Vera Rubin is a first-principles planning range, not an approved duty. Reference figures from the HYPERCUBE Product Family Design Collection, Edition 3, and NVIDIA published platform data. They are budgetary and indicative, not a supplier quotation. Final configuration, pricing, availability and delivery are confirmed by quotation for a named country and end user. Images are styling concepts.
What goes in.
One heat-rejection package beside the module. Vera Rubin's 45 °C water allows dry cooling in most climates.
Modules and packages
- 1 Twenty-foot module
- 6.06 × 2.44 × 2.90 m
- 1 Heat-rejection package
- Dry cooler, 4.0 m long, beside the module
What the site provides
- Low-voltage supply at 415 or 480 VAC, sized per edition
- Fibre uplinks to the customer network
- No water in operation: a closed liquid loop to a dry cooler
- Concrete pad for the module and the package, crane access
| Item | Qty | What it is |
|---|---|---|
| Rack-scale systemNVL72 | 1 | 18 compute trays of four GPUs and two CPUs, 9 NVLink switch trays and power shelves; GB300 NVL72 or Vera Rubin NVL72 |
| GPUs and CPUs | 72 + 36 | One NVLink domain |
| Fabric rackF1 | 1 | Scale-out, management and out-of-band network |
| Storage rackS1 | 1 | 0.5 to 1.0 PB usable NVMe |
| Power cabinet and buswayU1 | 1 | Distribution and protection to the rack power shelves |
| Coolant distribution unit | 1 | Liquid-to-liquid exchanger and pumps, sized for the edition ordered |
| Controls panel | 1 | Building management, access control and monitoring |
| Clean-agent fire suppression | 2 | Cylinders with detection, beside the entrance |
| Heat-rejection packageT1 | 1 | Dry cooler outside the module |
Options.
Chosen at quotation and fixed in the bill of materials for the edition ordered. Prices are given by quotation.
Compute
- GB300 edition: one GB300 NVL72
- Vera Rubin edition: one Vera Rubin NVL72 in the same MGX rack position
- Plant pre-sized for the Vera Rubin transition
Storage
- 0.5 or 1.0 PB usable NVMe
- Immutable backup tier
Network
- InfiniBand or Spectrum-X scale-out
- 800 Gb/s per GPU today, 1.6 Tb/s per GPU with Vera Rubin
- Out-of-band management network
Power and autonomy
- UPS ride-through to the rack power shelves
- 415 or 480 VAC; 800 VDC option for Vera Rubin
- Single or dual utility feed
Cooling and climate
- Dry cooler sized per edition and design ambient
- Hot-climate package
- Package in louvre fascia or sleeved finish
Security
- Access control and CCTV
- Customer-held encryption keys
- Telemetry under customer control
Site and services
- Factory, site and workload acceptance
- Remote monitoring
- In-country support and spares
Finish
- Pearl white sleeve, the reference finish
- Graphite sleeve with a white logo
- Titanium-grey sleeve for coastal and industrial sites
Best fit.
Where Prime Cell P72 earns its place.
- National champions and large state enterprises
- Telecom AI capacity and enterprise AI services
- Large research institutes and simulation
- Rack-scale reasoning, training and high-concurrency inference
Price Prime Cell P72 for your site.
Tell us the quantity, the delivery country and the timing. We reply with a written quotation: price, lead time and delivery terms.