South African startup Stratos Lab has begun implementing what it and its partners call the most powerful cloud service on the continent. This service is built on the latest Nvidia B300 'Blackwell Ultra' GPUs and is located in a data center in Centurion, north of Johannesburg. The entire system is planned to be ready within three months.
The project was unveiled on Wednesday at the Activate AI conference in Cape Town. In this initiative, Stratos Lab acts as the cloud operator, while Dubai-based company Ecoblox and neutral operator Digital Parks Africa (DPA) serve as the initial host. The servers are supplied by Taiwan's Giga Computing, which is part of Gigabyte.
The first phase includes 50 Giga Computing servers equipped with 400 Nvidia B300 GPUs (eight per server). Partners estimate the aggregate throughput of these systems at 7.2 exaflops. Ecoblox's Intercross CEO, Doug Makishima, stated that the order was placed about a month ago, and the machines are undergoing final checks for compliance with Taiwanese export requirements. He noted that these GPUs are under strict export control by the governments of the US and Taiwan, and expects them to be shipped by air within two to three weeks.
Stratos Lab CEO, Chris Mostert, said his team, together with Ecoblox and DPA, will complete the layout, cooling systems, and power supply at DPA's Samrand campus this week, after which the cluster will be installed in phases over five to six weeks. Makishima added that Stratos Lab will subsequently build its own sites using modular, container-based data centers from Ecoblox.
Mostert estimated the initial investment at $25 million USD, which, according to Makishima, only covers the GPU compute power. Additionally, an extra, yet unconfirmed, contribution of $25 million from a Mauritian bank under co-financing terms is being discussed. A second, unfinished order for 50 servers with AMD 'equivalent performance' GPUs would 'nearly double the capacity.'
Verification of Power Claims
Partners are cautious about the claim that their system is the 'most powerful.' Makishima acknowledged that competitor Cassava Technologies, who presented at the same conference, has more chips, but argued that those chips are older generation Hopper H100 and H200, whereas Stratos Lab uses newer Blackwell B300 chips. He emphasized: 'We are building on the B300, Nvidia's newest Blackwell architecture, so if measured in exaflops, our cluster is the most powerful currently coming online.' (Exaflop, or exaflop, is a measure of computer processing speed, meaning the computer can perform one quintillion (one with 18 zeros) calculations per second).
B300s began shipping in January 2026, and Nvidia has already introduced its successor, Vera Rubin, which is expected in the cloud in the second half of this year. Cassava received 3000 GPUs in South Africa, with plans for 12,000 across the continent.
The 7.2 exaflop headline is based on low-precision AI measurement, not double-precision calculations used for ranking world supercomputers. This figure represents a total of 144 petaflops on each of the 50 servers, totaling 7200 petaflops, or 7.2 exaflops, since one exaflop equals 1000 petaflops.
Comparing 'analogously' is more complex to verify. Recalculated in FP8, a higher precision measure, this figure decreases by about half to 3.6 exaflops, whereas Cassava's 3000 Hopper GPUs, at approximately four petaflops FP8 each, yield around 12 exaflops, more than three times greater. Although B300s are significantly more efficient per chip, rack, and watt, Makishima may have been referring to a narrower metric, but in raw computational power, the smaller cluster does not lead. Companies stated that their CTOs will follow up with training-related data; until then, the comparison remains unresolved.
Power, Not Floor Space
The cluster is flexible for GPU-as-a-service, bare-metal, and 'agent-as-a-service' workloads; the choice between them depends on the interconnects determining whether the system is suitable for training or inference. Makishima noted that early clients 'were largely leaning towards inference,' although the configuration can be changed for training.
Each rack consumes up to 65 kW, significantly exceeding standard computing capacities, and is sealed with a rear door heat exchanger connected to the building's cooling loop. Makishima explained: 'This means they can stand side by side. Thus, the only limitation is how much energy the data center can provide, not the floor space.'
Stratos Lab has secured an agreement with one anchor client, who remains undisclosed, and others are 'in the queue.' Mostert stated that the launch intentionally precedes demand. 'I would prefer to present the narrative that we have deployed more than we have clients, and that we are happy to find new customers.'
Sovereignty and Switches
Both leaders emphasized data sovereignty—a current political issue as the government considers South Africa's position in the global AI race. Makishima asserted that local ownership reduces dependence on foreign clouds such as Amazon, Google, and CoreWeave. Mostert stated that the South African entity can 'lock this down' so that the equipment 'stays in the country and never leaves it,' which is a strategic national resource.
Makishima reported that, 'to their knowledge,' there were no hidden loopholes for remotely disabling the GPUs after they left the factory. However, US legislators insist on requiring tracking of exported AI chips amid Washington's tightening control over where its processors end up.