Liquid-cooled GPU chassis inside a Rhodium 45 container
The Platform

Sovereign by architecture. Industrial by manufacture.

The container is a unit of scale, not a unit of innovation. Every part is industrial-standard. The moat is in the network, the software, and the customer relationships.

Built forSOVEREIGN ENTITIES·FINANCIAL INSTITUTIONS· · ·DEFENCE & TELCOS
Fiber optic strands rising from a server rack into cloud
Fiber to Sovereign Cloud

Bytes routed only through the jurisdiction that owns them.

Reference Specification

One container. One megawatt.

IT capacity, design
1.0 MW envelope · 8 racks × 120 kW liquid-cooled
Stage 1 pilot config
256 H200-class GPUs (~360 kW utilized)
Year 5+ configuration
Up to ~500 GPUs · B200 / B300 · Full 1 MW
Cooling
Direct-to-chip liquid + rear-door heat exchangers
Power source
Behind-the-meter natural gas (primary) · Grid (secondary) · 48–72hr LNG buffer
Connectivity
Redundant fiber + private MPLS · Air-gap operating mode supported
Software platform
Federated control plane · Customer-controlled cryptography · Jurisdictional severability
Deployment time
12 weeks site-ready to revenue
Form factor
40-foot ISO-compatible modular data center
Sovereign Software Commitments

Where the real moat lives.

Orchestration that understands jurisdictional constraints. Billing that maps to sovereign procurement cycles. Audit interfaces that regulators can use directly.

Federated

Each container's control plane runs locally. Shared code, not shared runtime.

Customer-keyed

Customer-held HSM keys gate every workload. Even Rhodium 45 administrators cannot decrypt customer data.

Severable

Jurisdictional cutover is a configuration change, not a re-engineering project.

Workload Focus

Inference and fine-tuning. Not training.

Training is a winner-take-all market for hyperscalers with 100,000+ GPU clusters. Inference and fine-tuning are the addressable, growing, sovereign-bound markets — and they fit distributed architecture far better than centralized training does.

TrainingInferenceFine-Tuning
Cluster scale10k–100k+ GPUs1–8 GPUs / request8–256 GPUs
LatencyWeeks / monthsSub-secondHours / days
Data residencySometimes bindingOften legally requiredAlmost always required
Distributed fitPoorExcellentExcellent
Rhodium 45 mix0%50–80%20–40%
How an engagement runs

From enquiry to production in four defined steps.

01

Enquiry

Quote form or 20-min briefing. Reply within 2 business days under NDA.

02

Scoping

Workload, jurisdiction, power and timeline mapped. Indicative proposal within 10 days.

03

Contract

Executed under your jurisdiction, through your incumbent SI or framework where possible.

04

Deployment

Site-ready to production inference in approximately 12 weeks.

Design a deployment for your workload.