The NVL72 rack
72 GPUs connected within a rack. The building block of the planned deployment.
Phase 01 / Planned architecture
Designed by WWT.
Built on NVIDIA.
A planned 32-rack NVIDIA GB300 NVL72 cluster, designed by World Wide Technology and to be deployed by WWT in cooperation with NVIDIA.
The compute foundation
Rack scale. Cluster ambition.
Keep the compute architecture and the complete service specification in view.
| Cluster design and deployment | World Wide Technology (WWT) has designed the cluster and will lead deployment. Meet our technology partners ↗ |
|---|---|
| Technology cooperation | The project is being developed in cooperation with NVIDIA, using its GB300 NVL72 platform. |
| GPU platform | Planned NVIDIA GB300 NVL72, based on Blackwell Ultra. |
| First-deployment compute scale | 32 compute racks × 72 GPUs = 2,304 GPUs. Supporting infrastructure is additional to the compute-rack count. |
| Campus development plan | The first deployment represents approximately 25% of the planned full-campus energy allocation. Further deployments are planned at Muxaro; their configurations will be defined separately. |
| Rack architecture | NVIDIA’s NVL72 platform combines 72 Blackwell Ultra GPUs and 36 Grace CPUs per rack, with rack-scale liquid cooling. |
| In-rack interconnect | NVLink within each NVL72 rack. The full 2,304-GPU cluster is not represented as one NVLink domain. |
| Inter-rack fabric | A separate scale-out networking layer. Final topology, bandwidth and selected components are subject to the approved design. |
| Storage and external connectivity | Final architecture, capacity and service specifications confirmed during diligence. |
| Server manufacturer and support | Final selection, support scope, acceptance and responsibility matrix confirmed in the proposal. |
Platform reference: NVIDIA GB300 NVL72 . Manufacturer specifications describe the platform; they are not measured SicurAI deployment results.
Three layers to agree
A system you can evaluate.
The cluster fabric
Inter-rack networking, storage and facilities that support the agreed workload.
The service boundary
Customer software, operator access, support and acceptance defined in the contract.
Beyond the compute rack
Every interface.
Considered together.
Power, cooling, connectivity, installation and software handover are part of the same deployment conversation.
The technical review defines the interfaces, responsibilities and acceptance criteria needed to turn a planned cluster into a customer service.
Request a technical discussionWorkload fit
Start with what
you want to run.
Large-model inference, training and fine-tuning have different requirements. Bring your workload, expected utilization, storage needs and connectivity constraints into the technical review.
Performance commitments should come from an agreed test method and acceptance criteria.
For model builders
Review model scale, distributed execution, checkpointing and storage throughput.
For infrastructure operators
Define customer software, orchestration, isolation, observability and the support handover.
For inference platforms
Discuss model mix, latency objectives, batch sizes and service boundaries.
Let’s review your technical requirements.
Tell us the capacity you need, your deployment window and what you want to build.