advertisement
Sovereign AI: Africa Is Buying The Chips, Not The Rules
If you missed Africa Tech Festival at the Cape Town International Convention Centre this week, you missed two different ideas of sovereign AI being discussed in rooms next to each other.
The first idea comes with a price list. It is counted in GPUs, megawatts and years to full build-out. On Day Two it had its own session, called “From AI Ambition to AI Factories.”
The second idea was in the next room. A regulator from the Independent Communications Authority of South Africa (ICASA) took questions on AI governance and customer data. The Africa leads of OpenAI and Meta discussed how to govern the AI economy. This idea has no price list. Instead, it comes down to one question. When a machine makes a decision about a person, can that person get a human to look at it again?
advertisement
It is only the first idea that is being built at scale.
What, Precisely, Is Being Built?
We now know the investment numbers are large enough to take seriously. Cassava Technologies has committed up to $720 million to put 12,000 NVIDIA GPUs into Africa. The company says it has bought all of them, and they are arriving in about 28 containerised shipments. In the first phase, 3,000 go to South Africa. The Cape Town facility is already running, and a 20-megawatt site is planned inside an existing data centre in Johannesburg. The other 9,000 go to Kenya, Nigeria, Morocco and Egypt over the next three to four years. “Our aspiration is to deploy between 12,000 and 13,000 GPUs across the continent,” said Ziaad Suleman, CEO of Cassava’s South Africa and Botswana operations.
In Egypt, Vodafone Business and Cassava launched what has been called the country’s first sovereign AI factory. A data centre venture with Elsewedy Electric is targeting 20 megawatts within three years and 200 megawatts in the long run. Egypt’s Ministry of Communications puts the first round of foreign investment at about $200 million, rising towards $1 billion at full capacity. It expects 200 direct jobs and around 2,000 indirect ones.
advertisement
Compare that with where the continent started. Strive Masiyiwa, who controls Cassava, has said he was told at the beginning that there were only 80 NVIDIA GPUs in the whole of Africa. Going from 80 to 12,000 in a few years is real change.
South Africa’s Deputy President Paul Mashatile opened the festival on Tuesday and made the political case. “We will compete by solving our own challenges, harnessing our own talent, building our own capabilities,” he said. His priorities ran from data centres and fibre to sovereign cloud and responsible AI governance.
Who Actually Owns What
Once you look at who owns the compute, the word “sovereign” means even less. NVIDIA and Google are shareholders in Cassava. Cassava AI, set up in London in 2024, works with Google, Microsoft, Amazon Web Services and Anthropic. In October, Cassava became an authorised reseller of Anthropic’s Claude models through Amazon Bedrock. The chips are American, and most of the models running on them are American too. The African parts are the building, the power contract and the customer relationship. Cassava’s leaders are open about it. “What matters to our customers is being able to use AI in their business,” said Hardy Pemhiwa, the group’s President and CEO. It also means a country can own a lot of compute and still have very little say over the systems making decisions about its citizens because those are two separate things. The problem is people are treating them as one.
advertisement
What Kenya Did
Kenya has done what the festival was arguing for, and it didn’t need a single GPU; it has given people the right to have a human review a machine’s decision about them, and it didn’t need a single GPU to do it.
In July, the Office of the Data Protection Commissioner published a Guidance Note on Artificial Intelligence. Section 6.6 sets out five points:
- Decisions should not be made only by automated means without meaningful human involvement.
- A person affected by an automated decision must be told about it.
- They can ask for a human to review it.
- The organisation must be able to explain the logic behind the decision when asked.
- A contested decision must be looked at by a human who has the authority to reverse or change it.
That last point sums up the whole argument. On this view, sovereignty isn’t about owning the machine. It’s about having someone with the authority to overrule it.
Section 9 lists seven high-risk categories, including employment screening and staff monitoring. These need impact assessments and disclosure to candidates and employees. None of this needed a data centre. All it needed was a regulator and a drafting team.
The Countries Buying Chips Haven’t Signed The Rules
This is where the two ideas uncomfortably split. By April this year, Strathmore University’s CIPIT had counted 44 African countries with data protection laws and 37 with working data protection authorities. It sounds reasonable. Hold that thought. The continent’s own binding treaty looks worse. The African Union’s Malabo Convention came into force in June 2023. By February this year, only a measly 20 ratifications/accessions out of the 55 member states counted.
Now compare that with the AI factories. Cassava’s GPUs are going to South Africa, Kenya, Nigeria, Morocco and Egypt. None of the five has ratified Malabo. The very countries buying the compute are the same countries that haven’t even signed the continent’s own rules on data and cybersecurity. Immaculate Kassait stepped down as Kenya’s first Data Commissioner this week. She named ratification as one of the things she was leaving unfinished.
How Long Should A Right Take?
A right to challenge a decision doesn’t help much if it takes years to use. Take the May 2025 Worldcoin case. It’s collection of biometric data broke the Data Protection Act. Later, it was ordered that the data of more than 300,000 people be deleted. The deletion was confirmed in January 2026, more than two years after the first complaints. Granted, a right that takes two years to enforce is still a right, albeit a slow one.
Section 6.6 asks for something quicker and simpler than going to court: identify and name someone inside the organisation who can review a decision then change it. Whether that person actually exists anywhere yet is still the question.
The Same Argument, But Inside The Enterprise
On Day One, the enterprise version of this debate came up without anyone calling it sovereignty. Dean Erasmus, Microsoft’s Chief Data Officer for South Africa, told a session on enterprise data that organisations usually have far more AI agents running than they think. In his experience it’s 10-100 times more. He drew the same line the regulators have. “No agent should be allowed in your environment if it doesn’t have an identity that’s defined by your organisation,” he said.
Kadotien Alassane Soro, the Country Head at Yango CIV came at it from an opposing side in an agentic enterprise session. “Your agentic AI is as good as the system you have access to and the information you have access to,” he said. In his books, it is data integration and people, not models, that are the real limits. Curiously, neither of them was talking about national sovereignty, yet they both described it. Know what is acting on your behalf. Know what it’s acting on. Keep a human who can stop it.
What To Take Back To Your Team
If you lead technology in an African organisation, the Cape Town debate turns into a few plain questions:
- How many AI agents are running in your environment? Does each one have an identity your organisation defined and controls?
- Who can overrule the machine? If AI makes or shapes decisions about customers or staff, can you name the person who reviews a contested decision and has the authority to reverse it?
- Can you explain a decision when asked? Kenya’s guidance expects you to be able to explain the logic behind it.
- Do you use AI to screen candidates or monitor staff? Kenya’s guidance puts both in the high-risk category, with impact assessments and disclosure. It’s a sensible bar even if you work in another market.
- When a vendor sells you “sovereign” AI, what exactly is local? Is it only the building and the power contract, or do you also get control over the models and your data?
What To Watch
Three things will show whether the Cape Town idea of sovereignty survives once procurement starts:
- Rights written into the deal. When an African government signs up to host an AI factory, will the contract also give citizens the right to ask why a system made a decision about them, and to have a person review it? Or will building the data centre and protecting citizens stay as separate conversations?
- Malabo ratification. Will the five countries receiving Cassava’s GPUs move to ratify Malabo now that they host the infrastructure?
- The section 6.6 person. Can a single Kenyan organisation name the person who reviews a contested automated decision and has the authority to reverse it?
As you can see, 12,000 GPUs answers a question about capacity. Sovereignty, however, is an entirely different question.