Stratos Lab Brings Nvidia Blackwell B300 GPUs to Centurion Data Centre

Hardware Specifications and Lead Times

The initial deployment consists of 50 servers supplied by Taiwan’s Giga Computing, a Gigabyte company. Each server houses eight Nvidia B300 GPUs, representing the latest Blackwell Ultra architecture that began shipping in January 2026. For local enterprise buyers, this hardware represents a massive leap in available processing capability, bringing cutting-edge silicon directly to the South African market.

Getting these highly restricted components into South Africa requires navigating strict international regulations. According to TechCentral, the hardware is currently clearing final export-compliance checks with both Taiwanese and United States authorities. Dubai-based integrator Ecoblox expects the servers to arrive via air freight within two to three weeks, bypassing traditional shipping delays.

Once available, the physical installation will move rapidly at the carrier-neutral Digital Parks Africa campus in Samrand. Stratos Lab and its partners plan to finalise the layout, power, and cooling this week, before installing the cluster in stages over five to six weeks. This rapid deployment schedule ensures commercial availability before year-end.

The financial scale of this local infrastructure is substantial. The initial compute hardware alone represents a US$25-million investment by the partners. Stratos Lab is also discussing an additional US$25-million matching facility with a Mauritius-based bank. This highlights the massive capital expenditure required to compete in the modern artificial intelligence hosting market.

Power Density and Cooling Realities

Deploying Blackwell Ultra chips introduces extreme power and cooling demands that traditional local facilities cannot easily support. Each rack in this new Centurion cluster will draw up to 65kW. This figure sits far above standard enterprise compute loads, requiring highly specialised data centre engineering to prevent catastrophic overheating and hardware failure.

To manage this immense thermal output, the setup relies on advanced liquid cooling techniques rather than traditional air cooling. The racks are completely sealed and utilise a dedicated rear-door heat exchanger. This exchanger ties directly into the building’s primary cooling loop, efficiently removing heat before it can affect the surrounding infrastructure.

This specific thermal management strategy allows for incredibly dense physical layouts on the data centre floor. Because the heat is contained and extracted at the rack level, the server cabinets can stand shoulder-to-shoulder. Consequently, the primary deployment constraint shifts entirely from physical floor space to the facility's total power provisioning capabilities.

Stratos Lab plans to eventually expand its footprint beyond the initial Samrand campus deployment. Future phases will utilise modular, container-based data centres designed by Ecoblox. This containerised approach allows operators to deploy high-density compute exactly where adequate power grids exist, effectively bypassing the physical limitations of legacy brick-and-mortar hosting facilities.

Navigating the Compute Performance Metrics

Evaluating the true power of this new cluster requires looking closely at the underlying precision metrics. The partners claim a combined throughput of 7.2 exaflops. However, this headline figure relies on FP4, a low-precision measure where each of the 50 servers produces 144 petaflops of raw performance.

When measured using the higher-precision FP8 standard, the Stratos Lab cluster's output halves to approximately 3.6 exaflops. This distinction is vital for technical buyers matching specific workloads to hardware capabilities. While FP4 is excellent for certain rapid calculations, complex training models often require the higher precision that halves the headline speed.

The local market is already seeing fierce competition in raw compute volume. Rival firm Cassava Technologies has already landed 3,000 older Hopper-generation GPUs in South Africa. Operating at roughly four petaflops of FP8 each, Cassava’s larger footprint delivers around 12 exaflops, significantly outpacing the newer Blackwell cluster in sheer processing volume.

Despite having fewer chips, the B300 architecture offers superior capabilities per chip, per rack, and per watt. Stratos Lab notes that early customer interest leans heavily toward inference workloads, though the interconnect allows reconfiguration for training. A planned second order of 50 AMD-based servers will eventually double the total capacity.

Data Sovereignty and Enterprise Security

A major selling point for this Centurion deployment is strict data sovereignty. Processing sensitive corporate or government data locally reduces reliance on foreign cloud providers like Amazon, Google, and CoreWeave. South African institutions can effectively ring-fence this hardware, ensuring their proprietary training data never leaves the country.

Security at the hardware level remains a critical consideration for enterprise buyers evaluating cloud partners. Ecoblox notes that, to their knowledge, these GPUs contain no factory backdoors that would allow remote deactivation. This provides crucial peace of mind for local operators investing heavily in what constitutes a strategic national resource.

However, international oversight of these powerful processors remains incredibly tight. United States lawmakers are actively pushing for location-tracking requirements on exported artificial intelligence chips. Washington continues to tighten controls over where its advanced processors end up, making local ownership of immediately available hardware a significant strategic advantage for South African businesses.

Demand for this sovereign capacity is already materialising in the local market. Stratos Lab has secured an undisclosed anchor customer, with others actively lining up for access. For South African IT leaders, securing a slice of this local GPU capacity early may be crucial as regional artificial intelligence adoption rapidly accelerates.