Mustafa Ehsan has spent the past year turning something as unglamorous as a spreadsheet into an argument about where AI economics is heading, and what that means for Africa.
Ehsan runs Convly, a research firm that tracks the price and performance of 33 AI models with the discipline of someone who has watched too many companies get the economics wrong.
His numbers tell a story of dramatic collapse: a workload that would cost a Nigerian startup roughly $1,750 a month on a premium American model can be run for about $21 on one of the cheapest capable alternatives today.
Model access, in his account, has stopped being the barrier it was two years ago.
As model access becomes cheaper, the question shifts from who can access AI to who can make it useful in their own context. Jephte Ioudom explored that question through a classroom.
In the Republic of Benin, the World Bank Group ran a study on how AI could affect education.
As part of it, an AI mathematics tutoring platform was built for the country’s pre-service mathematics teachers to test something largely unproven on the continent: what happens when an AI system is built around African teachers and African classrooms, rather than adapted to them afterward.
Ioudom, the founder of FoubsLabs who worked on the platform, came away from it with a conviction that now sits at the center of everything he says about Africa and artificial intelligence.
The teachers using that platform weren’t just testing software.
They were generating something Ioudom came to see as strategically important: data that could help “understand user patterns, shape AI products and discover new use cases.” And the question of who ends up owning that has come to define his entire theory of how Africa should approach this moment.
Both men spoke separately to Nairametrics: Ehsan about the economics reshaping who can afford to build AI at all, Ioudom about what it takes for Africa to build AI on its own terms.
The wager underneath both: Africa does not need to own the frontier model first. It needs to fight to own the layers around it, while building the capability to eventually build more of its own.
Nairametrics asked Ehsan whether African companies should build their own foundation models, and he didn’t hesitate.
His reasoning is blunt: training a frontier model costs hundreds of millions of dollars, and the result is often obsolete within months.
He points out that Kimi K3’s own predecessor was state of the art a year ago and is now three times cheaper and superseded. Spending scarce African capital to arrive third at something the market will commoditize anyway is, in his words, “a poor trade.”
Ioudom, asked almost the identical question, gives an answer that sounds different on the surface and turns out to be the same underneath.
That’s a process, he adds, that’s similar to what China has done: build the digital infrastructure and the talent pipeline first, and let the ambitious model follow. Not a rejection of the goal. A rejection of doing it first.
The distinction is worth holding onto rather than smoothing over. Ehsan doesn’t give a timeline for when building a frontier model might make sense; his objection is about economics that don’t obviously expire.
Ioudom does give a timeline: build the foundations now, build the frontier model later. They agree on what Africa should do next. They don’t fully agree on where it ends up.
Both men see open-weight models as an important bridge, for different reasons. For Ehsan, they deliver much of the capability without the frontier training bill.
For Ioudom, they let African engineers build expertise and “contribute to the development of these open-source models” rather than merely consuming them.
Their arguments meet most precisely on data.
Ioudom is direct to the point of being unambiguous: “Africa must own the data layer.”
He sees this not as an aspiration but as a precondition for meaningful AI sovereignty. He rejects the framing of AI as a new colonialism, but not because he thinks the risk is imaginary.
He rejects it because he thinks the solution is procedural: stronger data policy, sovereignty infrastructure, an African Union-led equivalent of Europe’s GDPR. The goal, in his words, is that “we no longer provide valuable data to foreign companies for free.”
The alternative he describes with real precision (external organizations taking African user data at no cost, building products from it, and “selling those products back to Africans at a premium”) is the exact outcome his entire policy argument is built to prevent.
Ehsan reaches the same conclusion from the language of business strategy rather than policy.
His prescription for where African capital should actually go, after ruling out training frontier models, is startlingly close to Ioudom’s. He points to local financial, legal and health data.
And to evaluation sets that test whether a model actually works for Yoruba, Igbo, Hausa or Nigerian Pidgin speakers, not just for someone else’s benchmark.
Talent, however, is one area where the argument becomes less abstract.
Handed a hypothetical $100 million for an African government, Ioudom gives the same answer at national scale: policy first, education second, infrastructure investment third.
He would also use specialist support from countries further along and make a deliberate effort to draw home nationals working in AI abroad. He points to China’s success at pulling its diaspora back as the template worth studying.
Neither is treating talent as a soft social priority. Both are describing it as productive infrastructure.
But the underlying infrastructure gaps Ioudom lists, energy and connectivity chief among them, have not closed just because a model got cheaper.
What remains as the harder problem is talent capable of building and evaluating these systems properly, the kind of institutional judgment that is, in Ehsan’s phrase, a “human and commercial problem rather than a capital one.”
