← HomeUnderstand sovereign AI
Short, plain answers to the questions behind the tracker. Each answer links to its sources.
What is sovereign AI?
A country's ability to build, run and govern AI on its own terms: where the computers sit, who owns them, whose data trains the models and whose laws apply.
No country controls every layer. Most depend on foreign suppliers for chips and cloud services, often US companies such as Nvidia. So sovereignty is a range, not a yes-or-no. Critics warn that "sovereign" can become a label for what is really a foreign product hosted locally.
A useful test when you see the word: who owns it, who can switch it off, and whose law governs it?
Sources: Paperjam (Nvidia interview), 2026 · Singh & Sengupta, arXiv, Nov 2025 · The Register, Mar 2026
What is an open AI model?
A model whose trained parameters (its "weights") anyone can download, run on their own computers and adapt. Strictly, that is an "open-weight" model; "open-source" means more.
Think of a recipe. A closed model, like most chatbots, lets you order the dish but never see the recipe. An open-weight model gives you the finished recipe to cook and change at home. A fully open-source model also shares the method: the training code and enough detail about the training data to rebuild it.
For South Africa this matters because local teams can build on an open model without sending data abroad or paying per question. Two local examples:
- MzansiLM (UCT, May 2026): an open model covering all 11 written official languages. It is a base for developers, not a chatbot.
- InkubaLM (Lelapa AI): a small open model for five African languages, including isiZulu and isiXhosa.
Sources: Engineering News, 6 May 2026 · Lelapa AI, 2024 · PBS NewsHour
What does South Africa need?
Six building blocks. Weakness in one holds back the others.
- Rules: a national AI policy, and privacy law (POPIA) that applies today.
- Computing power: GPUs that local researchers and businesses can actually use.
- Data centres and energy: buildings with reliable, affordable power.
- Data and languages: local text and speech, in the languages people speak.
- Skills: people who can build, run and govern AI.
- Adoption: government and businesses using AI well.
The World Bank groups similar ideas as the "4 Cs": connectivity, compute, context (data) and competency.
Sources: World Bank, 2025 · Oxford Insights, 2025 · See where each block stands →
Why does language matter?
Most South Africans do not speak English at home, and most AI tools are built mainly in English.
Only 8.7% of South Africans speak English as their main home language. isiZulu (24.4%), isiXhosa (16.3%) and Afrikaans (10.6%) are each larger. AI tools trained mainly on English work worse for everyone else. A clinic line, municipal service or bank chatbot that only works well in English leaves most people out.
Sources: News24, Census 2022 · Engineering News, May 2026
What does POPIA mean for AI?
South Africa has no AI law yet. The privacy law, POPIA, applies whenever AI uses personal information.
The Information Regulator says AI and automated decision-making bring new privacy risks; its chair, Pansy Tlakula, says the lack of a national AI policy makes it "a free for all". Section 71 limits decisions with serious effects that are made only by automated profiling, for example on credit, work or health. In practice, if a loan or job is refused solely by an automated system, you generally have the right not to be bound by it. Where an exception applies, you must be allowed to make representations and be told the logic behind the decision.
Sources: ITWeb, Sep 2026 · Bowmans
How do data centres stay cool, and how much water do they use?
Computers turn almost all their electricity into heat. How a data centre removes that heat decides how much water it uses.
- Evaporative cooling (cooling towers) removes heat by evaporating water. It is energy-efficient but uses about 1.8–4 litres of water for every kWh the computers use.
- Closed-loop cooling fills the pipes once and keeps the water circulating, so it does not need constant top-ups. Teraco says that, like many local operators, it uses closed-loop "free cooling".
- Dry coolers reject heat through air, as a car radiator does. They use no water but are less effective in hot weather.
- Seawater cooling: the Sines campus in Portugal uses Atlantic seawater and is designed to use no fresh water for cooling.
Microsoft says its newer zero-water design removes the more than 125 million litres a year a typical data centre uses. Useful questions for any planned site: which cooling does it use, and where does the water come from? Missing water information is one ground of the appeal against the rezoning for Equinix's planned Cape Town data centres.
Sources: Datacentres.com · TechCentral, Mar 2024 · Jacobs (Sines) · Bloomberg via Data Center Knowledge, Dec 2024 · Power and water in numbers →
Could data-centre heat make drinking water?
It is technically possible. Studies and patents exist, but we found no working data centre that turns its waste heat into drinking water.
There are ways to desalinate seawater with heat. One, membrane distillation, can run on low-temperature waste heat like a data centre's. But researchers say it remains uncertain whether it can match the energy efficiency of reverse osmosis, the usual method, which pushes seawater through a membrane using electricity.
What does exist: Google's Hamina data centre in Finland has a heat-recovery project to supply the town's district heating, and Sines in Portugal cools with seawater. Both reuse the sea or the heat, but neither makes water.
The scale in South Africa: Cape Town planned a permanent desalination plant at Paarden Eiland to make 50–70 million litres a day, at a cost of about R5 billion. In August 2026 the City said it was looking for another site after Transnet said it needs the land for port expansion; the plant is now expected around 2036. Mossel Bay's 15-million-litre-a-day plant, built in 2010 during a drought, is kept on standby.
A more likely overlap than turning heat into water is shared infrastructure. A coastal data centre could cool with seawater, as Sines does, in the same places that desalination plants draw it.
Sources: Energy & Environmental Science review, 2018 · Business Finland, May 2024 · Jacobs (Sines) · Cape Argus, Dec 2025 · Tygerburger, Sep 2026 · Western Cape Provincial Parliament, May 2026
What does this mean for my business?
You do not need to wait for the national policy. Three things apply now.
- Know where your data goes. If an AI tool sends customer information abroad, POPIA's rules on cross-border transfers (section 72) apply.
- Keep a person in the loop for decisions that seriously affect people, such as credit, hiring or claims.
- Check language fit. If your customers do not speak English at home, test tools in their language before you rely on them.
This is general information, not legal advice.
Glossary
GPUGraphics processing unit: the chip that trains and runs most AI.
Data centreA building full of computers, with its own power and cooling.
MW (megawatt)A unit of power. A large data centre can draw as much as a small town.
Open-weight modelAn AI model you can download and run yourself.
POPIAThe Protection of Personal Information Act, South Africa's privacy law.
CHPCCentre for High Performance Computing: the state's national supercomputer centre.
HyperscalerA very large cloud company, such as AWS, Microsoft Azure or Google Cloud.
CabinetThe President, Deputy President and ministers, who approve national policies.
How we decide a status
We rate what has happened, not what was promised.
A status changes only when a public source supports the change. Where sources disagree on a date, we say "expected" and give the range.
Corrections
Errors are corrected as soon as they are confirmed.
Each correction is listed here with its date and a description of what changed. No corrections have been published.
Where sources disagree, the page states which source is used and why. For example, reports give both 26 and 28 April 2026 as the date the draft National AI Policy was withdrawn; Robot Circle uses 26 April, the date of the government's announcement.
About Robot Circle
Robot Circle tracks the development of sovereign AI in South Africa: the computing capacity, data centres, policy and skills the country needs to build and govern artificial intelligence on its own terms.
Robot Circle is independent. It is not affiliated with, funded by or acting on behalf of any government body or any organisation it tracks.
Every entry is based on publicly available sources, which are linked alongside it. Research is assisted by AI tools. A status changes only when a public source supports the change.
Editor: Charl Dettmer, Cape Town