SA gets a 400-GPU AI cloud

South African companies can now rent computing power from a cluster of more than 400 of Nvidia’s latest B300 graphics processing units (GPUs).

The AI cloud has been launched through a partnership between Stratos Lab, Ecoblox and Digital Parks Africa, who describe it as “Africa’s most powerful AI cloud”. The system comprises more than 50 Gigabyte servers, each fitted with eight Nvidia B300 GPUs, at a Digital Parks Africa data centre in South Africa.

That places a massive pool of current-generation AI hardware within the country, and gives local developers, enterprises and research teams access to the machinery used for training large AI models and adapting existing models to specialist data.

Buying such systems outright is beyond the reach of most organisations. Apart from the cost and long hardware lead times, the servers demand specialist networking, huge electricity supplies and vast cooling systems. Gigabyte specifies power consumption of up to 109kW for a rack containing eight of its B300 servers, with direct liquid cooling built into the rack.

Stratos Lab will sell access to the cluster as cloud services. These range from bare-metal access, where a customer rents the physical server without a virtualisation layer, to GPU-as-a-service and managed platforms, on which developers can build and run applications. A planned Agent-as-a-Service offering is aimed at companies deploying AI agents that carry out tasks across business systems.

Nvidia’s reference specifications give each B300 GPU 288GB of high-bandwidth memory. A server containing eight of them has about 2.3TB, and a 50-server deployment would have roughly 115TB across the cluster. The GPUs inside each server are linked by NVLink, while very high-speed network connections allow work to be divided across multiple servers.

Model training is the obvious use, but the B300 generation has also been designed for inference: the work performed after training, when a model generates an answer, analyses an image or acts on a request.

“Deploying Africa’s most advanced AI Cloud powered by NVIDIA B300 GPUs represents a quantum leap for regional innovation,” says Chris Mostert, CEO of Stratos Lab. “By integrating our specialised GPUaaS orchestration and managed PaaS capabilities with world-class modular AI infrastructure, we are giving African enterprises and developers the raw, high-performance compute power needed to build world-class AI models locally.”

Data can remain in South Africa while it is processed, which can simplify governance for organisations handling personal, financial or commercially sensitive information. Users also avoid sending every request to a distant computing region, reducing network delay. Payment for a local service can remove some of the complications involved in buying cloud capacity offshore.

This is the basis of the partners’ use of “sovereign AI”: a country’s ability to develop and operate AI within its own jurisdiction, using infrastructure and data that remain under local control.

Africa has skilled AI researchers and a growing collection of local models, including work on African languages, agriculture, healthcare and climate, but access to large clusters remains scarce. The CSIR’s Centre for High Performance Computing supports scientific computing and has helped establish smaller university hubs across 11 African countries, but commercial cloud users often have to compete for scarce overseas GPU capacity and pay the data-transfer and currency costs that accompany it.

The new cloud could change the size of a project that an African team can attempt from within the region. For example, a bank could adapt a model to internal documents without moving the underlying data abroad, or a university could train language or scientific models on local datasets.

The three partners divide the job according to the parts that make an AI cluster usable. Ecoblox will integrate the servers, networking, modular infrastructure and management software. Stratos Lab will manage the computing resources and customer services. Digital Parks Africa will provide the data-centre space, power operations, connectivity and environmental controls.

Dense GPU clusters generate enormous heat and rely on tightly managed power and cooling. Menno Parsons, CEO of Digital Parks Africa, said: “Ultra-dense AI workloads demand an unyielding physical foundation, specialized power management, and pristine environmental controls.”

He described the Digital Parks Africa operation as “the carrier-neutral, highly resilient data centre facility required to sustain these 50+ NVIDIA B300 clusters, ensuring local data remains secure, accessible, and truly sovereign”.

* Arthur Goldstuck is CEO of World Wide Worx, editor-in-chief of Gadget.co.za, and author of The Hitchhiker’s Guide to AI – The African Edge.

The post SA gets a 400-GPU AI cloud appeared first on Gadget.