On June 9, an American company shipped its most capable artificial intelligence model. Three days later, a letter from the US Commerce Department forced it offline for every user on the planet, with no notice and no appeal.
The model was Anthropic’s Fable 5. The mechanism was an export control directive that barred access by any foreign national, so the company disabled it worldwide to comply. Other models kept running. This one did not.
Whatever you make of the security rationale, which Anthropic itself disputes, the structural fact is plain. A government on another continent reached into a service that African ministries, startups, and clinics were beginning to build on, and switched it off. As Maathangi Mohan says, the plug was in someone else’s hand.
The dominant response to that problem treats it as an access gap. Close the “AI divide,” the argument goes, by building data centers and importing GPUs until Africans can reach the same frontier models everyone else uses. Access matters.
But access to a model a foreign regulator can revoke on a Friday afternoon is not capability. It is a subscription with a kill switch. The useful question is which layers of the stack African countries have to own so that someone else’s decision cannot take the floor out from under its public services.
Sign Up Now for more African AI ownership insights
Someone Else’s Model Is Not a Strategy
Start with the hardware, because everything sits on top of it. African countries hold less than 1% of the world’s data center capacity while housing roughly 18% of its people. The United States and China alone operate more than 90% of the world’s usable AI compute, and more than 150 countries, the whole continent among them, have almost none.
Most African data already sits on servers in Europe and North America, which means it sits under foreign law, including a US statute that lets American authorities compel disclosure from US companies wherever the servers physically are.
When capability lives in someone else’s data center, it can be withdrawn for someone else’s reasons: national security, trade leverage, a complaint filed by a rival, a change of administration. The people cut off are rarely the intended target.
An LMIC health worker whose screening tool runs through a frontier API is collateral in a contest between great powers and frontier labs. Dependence on a revocable service is a borrowed capability, and the lender can call the loan without notice and without recourse. Fable 5 was the demonstration. It will not be the last one.
“Made in Africa” is a List of Layers
This is where Varaidzo Matimba’s framing earns its keep. In her 2025 landscape study for the MERL Tech Initiative, she defines Made in Africa AI as systems that originate from, are accountable to, and serve African communities, rooted in methods that prioritize local definitions of success rather than imported metrics.
The question she keeps putting to practitioners: who holds the power to design, deploy, and interpret the system.
Read that way, ownership is not one decision. It is a stack of them: whose data trains the model, whose languages it speaks, which problems it is pointed at, who evaluates whether it works, where it runs, and who can switch it off.
Some of those layers Africa cannot realistically own this decade. Frontier pretraining is one, and chasing a national supercluster to get it is, as I have argued here before, mostly a way to sell Nvidia chips.
But most of the layers that decide whether a tool helps a Kenyan farmer or a Malawian nurse are exactly the layers that can be built and governed on the continent, and they are cheap by comparison.
Layers Worth Owning
Small, open, on-device models are not only cheaper, they are un-revocable. A model whose weights you have downloaded, fine-tuned on your own data, and deployed on a community health worker’s handset cannot be turned off by a directive in Washington.
There is no kill switch because there is no switch: no API in Virginia, no license server, no cloud account to suspend. The compute sits where the user already is, and so does the control. The escape from the Western AI trap is to fine-tune small, task-specific models and run them on the phones people already carry, at a fraction of the cost of the frontier.
Language is the clearest case. Africa is home to more than 2,000 languages, and a 2025 review of large and small language models found that only about 42 of them have any meaningful support, with more than 98% severely underrepresented or ignored and just three scripts broadly covered.
No amount of imported compute fixes a gap that exists because the data, the tokenizers, and the evaluation were never built for these languages. Africans are building that layer themselves.
- Masakhane, the grassroots pan-African collective, has produced translation benchmarks for more than 30 African languages with no sovereign budget at all.
- Lelapa AI’s InkubaLM is a 400-million-parameter model trained from scratch for isiZulu, Yoruba, Hausa, Swahili, and isiXhosa, small enough to run without a hyperscaler.
These are the floor: owned in Johannesburg and in community repositories, not licensed from California, and impossible to switch off from abroad.
The economics are moving the same direction. McKinsey projects that by 2030 inference, the running of models, will overtake training to become more than half of all AI compute, and that inference is highly distributable, pushed toward smaller nodes close to the users it serves.
The expensive, centralized layer is training. The layer that scales with every query, and that wants to sit near the people using it, is inference. That is the layer Africa can own.
Ownership Changes Dependence
Owning the model layer does not escape the compute layer. Cassava Technologies is building Nvidia-powered AI factories on the continent, beginning with GPUs in South Africa, which is real and useful and still runs on silicon designed and rationed abroad.
Lelapa ships InkubaLM and still rents its GPUs overseas, working when Silicon Valley sleeps to get the time. The base models, the chips, and the next security patch still arrive from elsewhere.
“Made in Africa” can be a goal, not a law, that changes dependencies, and that matters.
The risk shifts from losing tomorrow what you already depend on, the Fable 5 problem, to missing out on the next frontier upgrade. The first is a kill switch. The second is a slower disadvantage you can plan around.
“Made in Africa” can be a plan to own the floor, the un-revocable layers that decide whether a tool serves Africans, and borrow the frontier for the ceiling on terms you can walk away from. Ownership of the floor does not guarantee benefit. It returns the decision about benefit to Africans.
“Made in Africa” Realities
A single country has no leverage against Nvidia or US export policy. A bloc has some. Smart Africa and the continent’s finance ministers should pool demand for cheap devices and fund a shared commons of open models and local-language data, because twenty ministers negotiating together can set terms one cannot.
Donors should stop financing only application pilots and start funding the floor: open language datasets, shared regional compute, and the small models that run on the devices Africans already own.
And any funder backing an AI health or agriculture tool should be able to answer one question before the money moves: where does this run, who owns the data, and who can turn it off. If the answer is a single foreign vendor, then there is an inherent sustainability risk.
Build the floor on what cannot be switched off from abroad. Borrow the ceiling if you must. Be resilient overall.
