SA gets a R798m Nvidia B300 AI cloud build in Centurion

South Africa is about to get a serious chunk of high-end AI compute on home soil.

Johannesburg-based Stratos Lab, infrastructure integrator ECOBLOX and Digital Parks Africa are deploying more than 50 Nvidia B300 HGX servers at Digital Parks Africa's Samrand campus in Centurion. The planned cluster will contain more than 400 Blackwell Ultra GPUs and the partners put the first two years of the programme at about $50 million, roughly R798 million at the exchange rate used when the project was announced.

The hardware matters because these are not ordinary data-centre accelerators. Nvidia's B300 sits at the high end of the company's shipping Blackwell generation, aimed at large AI training and inference workloads. Stratos Lab's own GPU-cloud page lists B300 capacity as "coming soon", with expected availability in the third quarter of 2026.

For South African companies that currently rent large amounts of GPU compute offshore, the bigger story is where the machines will live. Hosting the cluster locally gives businesses, researchers and developers another option for keeping sensitive data and workloads inside South Africa while cutting the network distance to the compute they use.

More than 400 GPUs in the first build

The announced configuration is more than 50 Nvidia B300 HGX servers supplied by Giga Computing, a Gigabyte Group company. With eight GPUs per server, the deployment comes to more than 400 GPUs.

The partners quote 7.2 exaflops of performance for the cluster. That figure needs context: AI infrastructure performance varies dramatically depending on the numerical precision and workload being measured, so headline exaflop numbers are not clean apples-to-apples measures between different GPU estates. What is less ambiguous is the scale and generation of the hardware: hundreds of current-generation B300 accelerators are a substantial local deployment.

MyBroadband reported that Stratos Lab CEO Chris Mostert put the programme target at $50 million over its first two years. The spend covers the 50-plus B300 HGX servers as well as the high-density infrastructure required to run them at Digital Parks Africa.

TechCentral reported that the stack is expected to be operational within about three months of the 26 August announcement. The project was unveiled at the Activate AI conference in Cape Town.

Stratos runs the cloud, DPA hosts it

The three main partners have distinct jobs.

Stratos Lab will operate the cloud service and GPU cluster. Its planned products include bare-metal compute, GPU-as-a-Service and managed platform services for model training, inference and other AI workloads. The South African company says it already works across cloud infrastructure, AI integration, hardware supply and colocation, with operations spanning African and Middle Eastern markets.

ECOBLOX is handling systems integration and infrastructure deployment, including cluster design, procurement, installation and optimisation.

Digital Parks Africa supplies the data-centre environment at its Samrand campus in Centurion, including the power and facility layer needed for unusually dense AI racks.

That last part is increasingly important. The constraint on modern AI infrastructure is no longer just whether a company can buy GPUs. Power delivery, cooling, networking and the ability to install dense clusters quickly have become part of the competitive advantage.

South Africa is becoming an AI-compute market

The Centurion deployment lands as local demand for AI infrastructure accelerates. Banks, telecoms operators, software companies and large enterprises are moving from small generative-AI pilots into production systems, while model builders need much larger pools of compute for training and inference.

South Africa also has a useful geographic position for a regional AI cloud: mature data-centre infrastructure, multiple international subsea cable routes and a sizeable enterprise technology market. Local compute does not automatically make AI cheaper, but it gives customers another choice besides sending every demanding workload to hyperscale regions outside the country.

The sovereignty argument may be just as important. Organisations handling regulated, commercially sensitive or personal information increasingly care about where data is stored and processed. A large GPU cloud hosted in Gauteng makes it possible to run more of those workloads domestically.

There is still execution risk. The full cluster is not yet live, and claims about being Africa's "most powerful" AI cloud depend heavily on which performance metric is used and what competing installations are counted. Cassava Technologies, for example, already operates a much larger estate of older-generation Nvidia H100 and H200 GPUs across the continent.

What Stratos Lab and its partners are doing differently is bringing a concentrated pool of Nvidia's newer B300 hardware into a South African facility. If the planned deployment arrives on schedule, Centurion will have one of the continent's most significant pools of current-generation commercial AI compute, and South African businesses will no longer have to look exclusively offshore for this class of hardware.