Sovereign AI compute, on home soil.
Portable, modular AI data centres — GPU-ready, liquid-cooled and running in about twenty-two weeks, not two years. Built so a nation’s most sensitive AI workloads stay inside its borders and under its own law.
- 22 weeks
- Order to operational
- 576
- NVIDIA GB300 GPUs
- 2.5 MW
- Total power capacity
- PUE 1.05
- Direct liquid cooling
- 8 → 1,152
- GPUs across the range
The physical layer of sovereign AI.
Compute, modular data centres, semiconductors and strategic infrastructure. Prometheon develops, integrates and partners across AI infrastructure.
GPU infrastructure, private and on-premises AI, sovereign compute.
From edge to multi-megawatt, for national and enterprise infrastructure.
A future strategic focus, through ecosystem partnerships.
Relationships capable of supporting modular sovereign-compute deployments.
- Infrastructure
- Intelligence
- Applications
- Products
- Customers
- Recurring revenue
A data centre that arrives.
Factory-built modules for compute, cooling, storage, power and network — integrated and tested before they ship, then joined on site into a single machine.
-
I
Compute
- 8 × NVL72 racks, 576 NVIDIA GB300 GPUs
- Rack-scale NVLink fabric
- About 1.2 MW of compute load
-
II
Cooling
- Direct liquid cooling to every GPU
- Closed glycol and water loop
- N+1 redundant pumps
-
III
Storage
- NVMe all-flash, petabyte scale
- GPU-direct data paths
- Fast checkpointing for long training runs
-
IV
Power & network
- 2.5 MW dedicated feed with UPS backup
- 800 Gb/s InfiniBand and Ethernet
- Clean-agent fire suppression
Capacity in steps. Add modules as demand grows, instead of committing to a whole building on day one.
Ready for the next generation. Designed around NVIDIA GB300 today, and Vera Rubin as it arrives.
Infrastructure, built in modules.
The same architecture at every size, so a ministry can start with one rack and grow to a national campus without starting again.
| System | GPUs | Built for |
|---|---|---|
| Managed rack | 8 | An enterprise or ministry starting out |
| Regional cell | 32 | Regional inference, or a four-node private AI cloud |
| Rack-scale module | 72 | One NVLink domain — also in a high-assurance sovereign build |
| AI factory | 576 | A nation’s sovereign AI compute |
| Campus block | 864 | A campus that grows block by block |
| Next generation | 1,152 | Reference architecture · design target |
Sovereign AI, on the desk.
Not every model needs a data centre. Personal and departmental systems keep sensitive work on hardware you hold.
-
I
AI laptops
- 48 to 128 GB of unified memory
- 14 to 16-inch, memory-rich
- A sovereign build for sensitive work
-
II
Desk systems
- 1 PFLOP FP4 on a single node
- 2 PFLOP as a linked pair
- Local models, no cloud
-
III
Team lab
- 8 PFLOP FP4 across eight nodes
- Each node scheduled independently
- A department’s own AI cluster
-
IV
Frontier station
- NVIDIA GB300 at the desk
- Hundreds of gigabytes of coherent memory
- Frontier-scale work without a server room
The silicon itself, sourced.
For clients building their own capacity, we source NVIDIA systems directly — from a single server to a full rack, and the next platform as it ships.
NVIDIA HGX B300
Eight-GPU Blackwell Ultra server node. By inquiry.
NVIDIA GB300 NVL72
Liquid-cooled rack: 72 GPUs in one NVLink domain. By inquiry, subject to allocation.
NVIDIA Vera Rubin
NVIDIA’s next rack-scale platform. By special inquiry.
Compute is a strategic asset, like energy.
-
I
Jurisdiction
Data, models and results stay inside the country, under its own data-protection law, with no dependence on a foreign hyperscaler.
-
II
Security
Air-gapped partitions for classified work, biometric entry, and encryption at rest and in transit, end to end.
-
III
Local capability
Local assembly, training and knowledge transfer, so the people who run the system are the nation’s own.
One system. A nation’s workloads.
Sovereign language models
Large models trained on a nation’s own languages and data, without relying on foreign platforms.
Defence & security
Signal intelligence and data fusion on air-gapped partitions, with zero outside exposure.
Healthcare
Population-scale genomics, AI radiology and faster drug discovery.
Climate & agriculture
High-resolution weather simulation, crop-yield prediction and resource planning.
Public services
Identity verification, compliance auditing and case analytics for accountable government.
Research & education
Supercomputing for universities and laboratories, and a home for national talent.
From order to live in twenty-two weeks.
A factory-built approach with engineering in parallel — against 18 to 24 months for a conventional data-centre build.
- Weeks 1–2
Order
Commercial terms, technical validation, approvals.
- Weeks 3–14
Build
Site preparation and module manufacturing, in parallel.
- Weeks 12–17
Ship
Logistics, customs and secure transport to site.
- Weeks 17–21
Install
Placement, interconnect, power, cooling and network.
- Week 22
Go live
Benchmarking, security hardening, handover and training.
The metal, shown.
The silicon that thinks and the hall that keeps it running, rendered live. Turn them with your hand.
The system, in numbers.
| Configuration | Factory-built modules for compute, cooling, storage, power and network, operated as one system |
|---|---|
| GPUs | 576 × NVIDIA GB300 (Blackwell Ultra) |
| Compute racks | 8 × NVL72 |
| Network fabric | InfiniBand and Ethernet, 800 Gb/s |
| Storage | NVMe all-flash, petabyte scale, GPU-direct |
| Power | 2.5 MW total · 3-phase 415 V · N+1 lithium-ion UPS |
| Cooling | Direct liquid cooling · PUE 1.05–1.08 |
| Security | Air-gapped partitions · biometric access · encryption at rest and in transit |
| Time to operational | About 22 weeks from order |
Indicative. Final configuration, timeline and pricing depend on site readiness, approvals and technical scope. Available to purchase or as a funded lease.