01
Sovereign data requirements
Capacity is placed in jurisdictions where clients need their data and workloads to remain, rather than routed offshore by default.
Infrastructure
The physical layer of AI.
AI GPU Data Center is developing dedicated GPU infrastructure as an integrated physical and operational system — hardware, fabric, power, cooling, security and operations planned together rather than assembled after the fact.
- Power — grid supply enters the unit
- Cooling — thermal capacity above the racks
- Network — interconnect and transit below the racks
- Unit 01 — 1MW planned first-phase capacity
- Expansion — phased pathway towards 5MW
Design principle
AI GPU Data Center's infrastructure model treats compute as an integrated physical and operational system. Hardware selection remains configurable according to workload requirements, with NVIDIA B300 systems forming the planned initial build-out.
Architecture layers
Nine layers, one operating envelope.
Layer 01
GPU systems
High-density accelerated compute forming the core of each deployment. NVIDIA B300 systems are the planned initial architecture; alternative configurations remain available according to client requirements.
Planned specification
Design intent, stated plainly.
The table below records current design intent for the first phase. It is not a product datasheet and contains no performance, uptime or availability claims.
Compute
- Initial GPU architecture
- NVIDIA B300 systems (planned initial build-out)
- Alternative configurations
- Selected according to client workload and commercial requirements
- Access model
- Dedicated bare-metal — no shared tenancy on client infrastructure
- Host compute and memory
- Sized to keep accelerators supplied for the target workload
Fabric and storage
- GPU interconnect
- High-speed low-latency fabric planned for distributed multi-node training
- Cluster networking
- Segmented client networking with controlled external connectivity
- Storage
- Throughput-oriented tiers designed around dataset size and checkpoint behaviour
Facility
- First-phase capacity
- 1MW planned
- Expansion pathway
- Phased towards 5MW, aligned with contracted demand
- Power
- Distribution, redundancy and headroom treated as a primary design input
- Cooling
- Thermal design matched to high-density racks and site conditions
Operations
- Monitoring
- Power, thermal, fabric and node telemetry for operations and lifecycle planning
- Security and access
- Physical, logical and administrative controls for data-sensitive workloads
- Managed operations
- Operation, maintenance and lifecycle activity under an agreed service framework
Unit 01
The first module, as ordered.
Equipment for the first modular deployment has been ordered. The figures below describe what is committed rather than what is intended.
- Status
- Equipment ordered — modular unit, racks, chillers and cooling
- Initial compute
- 8 × Supermicro NVIDIA HGX B300 servers — 64 B300 GPUs
- Pre-provisioned capacity
- 16 servers / 128 GPUs within the same module
- Cooling
- Rear-door heat exchangers matched to high-density racks
- Module cadence
- Further modules planned on an approximately quarterly import cycle
Locations and power
Sites chosen where power and data residency line up.
Deployment is modular, so each site is assessed on its own power position. Grid connection, generation, storage and redundancy arrangements are developed per site and published once contracted.
02
Wind and solar availability
Sites are shortlisted where renewable generation is available locally, so power strategy and site selection are decided together.
03
At least 1MW of green energy
A minimum of 1MW of green energy at the site is a selection criterion for each earmarked location.
Earmarked sites
Humansdorp
South Africa
Atlantis
South Africa
Harare
Zimbabwe
Integrated system
Every layer is a potential constraint.
- 01
Power
Supply, distribution, redundancy
- 02
Facility
Cooling, density, physical security
- 03
Compute
GPU, CPU, memory
- 04
Network
Interconnect, storage, transit
- 05
Workload
Training, inference, research
One integrated system — a constraint in any layer becomes a constraint on the workload
Phased capacity development
Built in increments, against contracted demand.
Phase 01
1MW planned initial deployment
Dedicated high-density infrastructure based on NVIDIA B300 systems, subject to final engineering and client requirements.
Expansion
Phased pathway towards 5MW
Additional capacity developed according to client uptake, power availability, technical requirements and contracted offtake.
Future Scale
Repeatable infrastructure growth
A modular operating model designed to support additional capacity and future infrastructure locations as demand develops.