Africa's AI Moment Is Local Languages and Small Bandwidth

The wrong question

Most conversations about artificial intelligence in Africa are held in the wrong frame. The question people ask is whether Africa can catch up: whether we will train frontier models, whether we will have the data centres, whether a lab in Kampala can beat a lab in California on a benchmark.

That framing guarantees a losing answer, and it is not even the interesting question. The interesting question is what a model has to do to be useful to a farmer in Pader, a head teacher in Kitgum or a treasurer in a savings cooperative, because whatever answers that is a product with tens of millions of potential users and almost no competition.

The gap is not compute. It is language and delivery, and both of those are solvable from here.

I want to make this argument with numbers where I can, because the topic attracts a lot of enthusiasm and very little arithmetic. Where I could not verify a figure I have left it out rather than repeat something that sounded right.

I should say where I am standing. I write this as somebody who builds systems for schools, cooperatives and clinics in Uganda, not as somebody with a view of a research roadmap. That narrows what I can claim and sharpens what I can see, because the products I am describing either work in a district office next term or they do not, and there is no version of that judgement I get to postpone.

What delivery actually looks like from here

The International Telecommunication Union’s April 2025 report on the state of digital development in the Africa region puts hard numbers under the problem. By 2024, mobile broadband covered 86 per cent of the continent’s population, but only 38 per cent used the internet, against a global average of 68 per cent. Internet use reached 57 per cent of urban populations and 23 per cent of rural ones, the widest such gap in any of the ITU’s regions. And 16 per cent of the population was still relying on 3G, while 5G had reached only 11 per cent.

Then there is cost. The same report puts the median price of two gigabytes of mobile broadband a month at 4.2 per cent of gross national income per capita, the highest of any region and more than double the UN Broadband Commission’s affordability target of two per cent.

Africa in 2024: two gigabytes a month costs more than double the affordability target, and a sixth of the continent is still on 3G. Source: ITU, April 2025.

Read those figures as a product specification and they say something blunt. A useful African AI product cannot assume a fast connection, cannot assume a connection at every moment, and cannot assume the user will happily spend megabytes on a conversation. Anything that streams tokens over a shaky link to somebody counting their bundle is a demonstration, not a product.

This is not a constraint to apologise for. It is a design brief, and it is close to the brief that produced mobile money, which succeeded precisely because it worked on the phones people already had rather than the phones a designer wished they had.

Language is the interface, not a feature

Linguists count somewhere above two thousand living languages in Africa, with Ethnologue’s tallies varying according to how you draw dialect boundaries. Almost none of them are well represented in the text these models are trained on.

The consequences are concrete. UNESCO’s Courier reported on 2 April 2026 that ChatGPT recognises only around 20 per cent of written Hausa sentences, a language with tens of millions of speakers in Nigeria alone. The same piece described the African Next Voices project digitising 9,000 hours of spoken language from Kenya, South Africa and Nigeria, which tells you both that the data gap is being worked on and how much patient manual effort a single tranche of it costs.

I feel this personally rather than academically. I am fluent in Acholi and I translate between it and English. When I test a general-purpose model in Acholi, it does not fail loudly. It fails politely, producing fluent, confident sentences that any speaker immediately recognises as nonsense, which is a worse failure mode than an error message because somebody who does not speak the language cannot tell the difference. A monolingual product manager reviewing that output would sign it off.

For most people in this region, language is not a preference layered onto a working product. It is the difference between the product existing and not existing. A school reporting system a head teacher can only operate in English has silently added a translation task to their evening. A cooperative dashboard in English tells a farmer-member that the information inside it belongs to somebody else.

Sunflower, and Acholi first

The most encouraging work I have seen on this is coming out of Uganda. Sunbird AI released Sunflower, a pair of open models at 14 billion and 32 billion parameters, built on Qwen 3 and covering 31 Ugandan languages, published with a paper on arXiv in October 2025 that the model is named after. Their reported results place Sunflower ahead of both GPT-4o and Gemini 2.5 Pro on translation for 24 of those 31 languages.

Sunflower’s reported results across the 31 Ugandan languages it was evaluated on, against a general assistant’s grasp of written Hausa. Sources: Sunbird AI, October 2025; UNESCO Courier, 2 April 2026.

The argument in that work travels well beyond Uganda. Rather than adding the world’s largest languages to a model one at a time, pick a region, cover its languages properly, and accept a smaller footprint in exchange for real competence in the places you claim to serve. A model that handles 31 Ugandan languages well is more useful in Uganda than a model that handles 200 languages badly, and it is small enough to reach places the larger one never will.

We are building on it. In our own work the translation path goes into Ugandan languages with Acholi first, partly because I can verify the output myself and partly because Acholi is spoken across the region where our cooperative and school clients actually operate. Verification matters more than people expect. If nobody on the team speaks the target language, you have no way of knowing whether your translation feature is working or merely running, and the polite failure mode above means you will not find out from the logs.

Living inside a thousand requests a day

On 26 August 2026 we were granted an API quota of one thousand requests a day. That is a generous grant and I am grateful for it. It is also a hard architectural fact, and I want to describe what it does to a design, because this is the part that enthusiastic articles about African AI leave out.

A thousand requests a day is not a chat interface. It is a budget. It forces you to decide which requests are worth spending, to cache aggressively, to batch overnight rather than translate on demand, and above all to make the feature work when the quota is gone. Our answer is a local model fallback: a smaller model running close to the user handles the ordinary cases, with the hosted API reserved for the ones that genuinely need more. When the quota is exhausted or the network is absent, the product degrades instead of dying.

I have come to think of that pattern as the default for anything built here rather than a workaround for a temporary shortage. If your system stops working the moment an external service is unreachable, you have not built for this market. You have built for a market with better infrastructure and hoped ours would behave like it.

There is a second, quieter benefit. A quota makes you honest about which uses of a model are actually worth anything. Given unlimited calls, teams sprinkle intelligence over everything, including places where a lookup table would have done. Given a thousand, you find out very quickly which three features people care about.

The unglamorous problems

The other half of the argument is where to point all this. There is a strong pull towards impressive-sounding applications and a much better return in boring ones.

Scholaris, one of the two products I am building now, is an offline-first school reporting system. There is nothing exciting about a report card. But a head teacher assembling termly results by hand, then travelling to a district office to submit them, is losing days each term to a problem software solved elsewhere decades ago, and the reason it is unsolved here is not difficulty. It is that nobody built for a school with unreliable power and no dependable connection.

A secondary school examination room in Uganda. Photo: Zach Wear, Unsplash.

Labwor Fleet King, the other, does fleet telematics with fuel-theft detection. Fuel theft is not a research topic. It is a line item that quietly removes a large share of the operating margin of every transport and mechanization business I have worked with, including cooperatives whose tractors are the most expensive things they own. The intelligence needed to spot a suspicious drain pattern is modest. The value of spotting it is not.

Cooperative records are the third case. A savings and credit cooperative serving farmer members typically runs on paper plus one spreadsheet. Getting that into a system that answers a member’s question in the member’s own language is worth more to them than any general-purpose assistant.

None of these will be written up as a breakthrough. All of them change somebody’s week, which is a better test.

What this asks of builders

Building for this asks for a different taste: being pleased by a model small enough to run on a mid-range Android phone rather than impressed by one that needs a data centre, and treating a language nobody has benchmarked as an opportunity rather than an inconvenience. It also asks for honesty about what is not ready. Models in low-resource languages still make errors a fluent speaker catches instantly, which makes a human in the loop a design requirement for anything consequential rather than a temporary embarrassment to be engineered away next year.

Africa is not late to artificial intelligence. It is early to the version that will matter most: the version that runs small, speaks the language of the room it is standing in, and is aimed at a report card or a fuel tank rather than a leaderboard. The continent that had to make payments work on feature phones is now being handed a technology everybody else assumes will arrive over fibre. I have some confidence about who ends up building the more interesting thing.