Every Navon server is designed to the job it runs, with each watt of the 20 kW envelope spent on compute, or each rack unit spent on capacity, sized to what the customer runs. The same engineering produces five turnkey reference configurations that cover the most common sovereign-compute use cases. Four are compute-led and one is storage-led, all shipping pre-integrated, cabled, witness-tested, and commissioned.
A single 20 kW rack pre-built for university, lab and research workloads: three CPU nodes for general compute, two 8-GPU training nodes, and a half-petabyte NAS, on a redundant 25 / 100 G fabric. Fourteen rack units left free for the team to grow into.
Five 20 kW racks built around 8-GPU training systems, each paired with a Blackwell-class PCIe accelerator node to fully utilise the 20 kW envelope. Engineered for large-model training and high-throughput inference, with a flat 400 G fabric across the cluster. Ships pre-integrated with vLLM, Ollama, and an OpenAI-compatible endpoint, the same serving stack as the sovereign cloud, on dedicated silicon. Two tiers: latest-generation or prior-generation HGX.
Five 20 kW racks built around dense PCIe inference nodes, tuned for serving traffic rather than training runs. Memory-bandwidth-rich silicon, packed as tightly as the 20 kW envelope allows, behind a vLLM-class serving stack with continuous batching and paged-KV-cache. Optimised for the metric that shows up on the bill: cost per million tokens out, not peak FP16 TFLOPS.
Five 20 kW racks built around a non-mainstream AI accelerator paired with general-purpose PCIe compute. For customers who want serious training and inference capacity without committing the entire stack to a single silicon supplier, giving sovereignty and supply-chain diversity by design.
Storage-led racks shaped to whatever the customer holds, from a few hundred terabytes for a research lab to multiple petabytes per rack for sovereign archives, scientific data lakes, and long-tail data residency. Mostly disk shelves, a slim controller, and the same dual-corded power and 25 / 100 G fabric the compute racks ride on. Configurations span all-flash for low-latency archive, hybrid NVMe + HDD, and pure HDD for true cold tiers.
We engineer infrastructure backwards from the application, not from the catalogue, to match each workload and its environment. That means optimal performance, cost-efficiency, and a clean fit with the modular Tier III data centre that hosts it.
Configurations are selected by end-use case, such as AI inference, HPC analytics, training, and storage. The most suitable hardware system is determined by the use case's performance requirements and cost constraints, not by what's on a SKU sheet.
Vertical and horizontal scaling paths to track software evolution, with margins of safety so customers never top off on compute. Cooling, IT-load and space envelopes are sized after the application is locked, not before.
Competitive CapEx operationalised through energy-aware configurations and PUE-tuned designs. A modular growth model, where capacity follows demand rather than the other way around, keeps every euro and every watt working.
Encrypted storage, tamper-evident racks, in-jurisdiction compute. High availability and extended warranties are included by default, never retrofitted at the contract stage.
Every dial sized to the workload, not the catalogue. Every rack lands in a Navon module.
Compute racks sized so steady-state load lands ≈95% of the 20 kW envelope. Storage racks sized to what the customer holds. CapEx on what gets used, not stranded.
Every node dual-corded across A and B PDUs, sized so single-feed failover carries the full ≈19 kW load within budget.
Cluster-wide Ethernet: 25/100 G for research and inference, full 400 G for training and open-accelerator clusters. Vendor-open switch stack, no network silo.
Crating, freight, on-site rack-and-stack, network and software bring-up. A specialist field team. Customer takes delivery of a working cluster, not a parts list.
Every configuration sized to the same 20 kW rack envelope as the 400 kW MDC. Compute and storage live side-by-side under the same facility, DCIM, and cooling budget.
Where the workload demands the latest GPU silicon, we ship it. Where you'd rather hedge supply-chain or sovereignty risk, we offer a credible non-mainstream accelerator path on the same rack, fabric, and operating model.
The research-rack reference design has three CPU nodes, two 8-GPU training nodes, a half-petabyte NAS, a dual leaf fabric, and dual-corded power. Click any unit to inspect its hardware, networking, and load.
Two GPU nodes for training and high-throughput inference. Three CPU nodes for general compute and orchestration. One NAS for shared training data and checkpoints. Two leaf switches and two PDUs in active/active redundancy. Hover any rack unit to highlight its connection.
—
| Device | PSU config | Feed A draw | Feed B draw | Total | Notes |
|---|---|---|---|---|---|
| GPU Node 1 | 2× 3 kW redundant | 3.25 kW | 3.25 kW | 6.50 kW | Active/active across A+B |
| GPU Node 2 | 2× 3 kW redundant | 3.25 kW | 3.25 kW | 6.50 kW | Active/active across A+B |
| CPU Node 1 | 2× 1.6 kW redundant | 0.75 kW | 0.75 kW | 1.50 kW | Active/active across A+B |
| CPU Node 2 | 2× 1.6 kW redundant | 0.75 kW | 0.75 kW | 1.50 kW | Active/active across A+B |
| CPU Node 3 | 2× 1.6 kW redundant | 0.75 kW | 0.75 kW | 1.50 kW | Active/active across A+B |
| NAS | 2× 1.6 kW redundant | 0.50 kW | 0.50 kW | 1.00 kW | Active/active across A+B |
| Asterfusion leaves | 2× PSU per switch | 0.28 kW | 0.27 kW | 0.55 kW | Primary on A, redundant on B |
| Feed totals | 9.53 kW | 9.52 kW | 19.05 kW | Failover load per feed: ≈ 19 kW (within 20 kW budget) |
Load figures are steady-state estimates derived from component TDPs (EPYC 9965 ≈ 500 W; RTX 5090 ≈ 575 W TGP) plus platform overhead. Peak training bursts can push GPU nodes ≈ 10-15% higher, so size PDUs and breakers for peak, not average. On single-feed failover the surviving feed carries the full ≈19 kW, so both PDUs are sized to 20 kW.