Why African start-ups are increasingly building on Chinese AI

Ask African start-ups which artificial intelligence they build their products on, and the answer surprises those who imagine a market dominated by the big American names. According to Lexi Novitske, a partner at the pan-African venture capital fund Norrsken22, the trend is clear: open models, and Chinese models in particular.

« If you look at what local tech companies are building their systems on today, it is open source models, Chinese AI models, because they are cheaper and more available. »

Lexi Novitske, partner at Norrsken22, to TechCentral (paraphrased)

Lexi Novitske knows the ecosystem well. Based in Nigeria since 2012, she invested early in Paystack and Flutterwave, two of the biggest successes of African fintech. In 2023, Norrsken22 closed a first fund of $205 million, backed notably by founders of Skype, Delivery Hero, King and Mojang. The fund has invested about half its capital, with no exit so far.

Price first

The main reason is economic. An African start-up often bills its customers in naira, rand or shillings, in markets where purchasing power is limited. Paying for calls to an AI model in dollars, at the rates of the big American providers, can absorb a large part of its margin. Chinese models, offered at much lower prices, or free to download when their weights are openly published, change that equation.

Lexi Novitske cites Alibaba Cloud’s tools as less expensive, and the Qwen family of models, developed by Alibaba, as performing very well on the market. Other Chinese models, such as DeepSeek’s, have also gained worldwide popularity over the past two years, notably among developers who want to run a model on their own servers.

Open, therefore adaptable

The second advantage is availability. A model whose weights are published can be downloaded, installed on a local server or with a host of one’s choice, and adapted to one’s needs: trained further on African-language data, specialised in a field, slimmed down to run on modest hardware. For start-ups that want to control their costs and keep their data close to their customers, that is valuable.

Closed models, accessible only through a paid interface, often offer the best performance on the hardest tasks. But for many everyday uses, such as classifying messages, summarising documents, customer service assistants or translation, a well-tuned open model is more than enough.

Open weights, open source: what are we talking about?

An AI model is a very large file of parameters, its “weights”, obtained after costly training. When a company publishes these weights, anyone can download the model and run it on their own machines, without going through the original company or paying it for each use. This is called an open-weight model. People often speak of open source, even though the terms of use vary from one model to another and the training data is rarely published.

For an African start-up, this opens up concrete possibilities. It can host the model with a local provider or in a data centre close to its customers, reducing response times and transfer costs. It can also adapt it: teach it the vocabulary of its sector, or improve it in languages such as Yoruba, Swahili or Wolof, which the big models still handle poorly. Communities of African researchers, such as the Masakhane network, have been working on these languages for years and produce data used for these adaptations.

The “vacuum” left by American capital

Lexi Novitske links this trend to a broader phenomenon: the retreat of American venture capital from Africa. In her view, it is not so much a matter of government policy as of the choices of private investors, who reduced their exposure to African markets after the 2022–2023 correction.

« I think American private capital in particular is missing out. It is not so much a question of government policy as a question of private capital. »

Lexi Novitske (paraphrased)

She sees a huge vacuum that Chinese players are ready to fill. The fintech precedent is telling: OPay and PalmPay, backed by Chinese capital, have captured a large share of payments in Nigeria. According to the investor, Chinese companies can afford to accept losses for three or four years while they build their network effects, a luxury local founders do not have.

The risks of dependence

Building on Chinese AI is not without questions. First, compliance: some Chinese models, in their hosted versions, apply filters linked to the content rules in force in China. Downloaded versions can be configured differently, but companies need to check how the models behave on topics that are sensitive for their users.

Next, data: using a service hosted in China raises the same questions as sending data to any foreign country, under African data protection laws, which increasingly regulate international transfers. Running an open model on a local server avoids this problem, but requires skills and hardware.

Finally, geopolitics. Technology tensions between the United States and China can have unpredictable consequences: restrictions on certain services, on chips or on funding. A start-up that depends on a single provider, whoever it is, is exposed. Regulators, Lexi Novitske points out, are already uneasy about the dominance of Chinese-backed fintechs in transaction volumes.

Finally, there is the question of quality. A cheaper model is only worthwhile if it does the job. Serious teams compare several models on their own data before choosing, measuring the accuracy of answers, the handling of local languages and speed. Sometimes the cheapest model is also the best suited; sometimes a more expensive model cuts errors so much that it ends up costing less. For many start-ups, choosing Chinese AI is first the result of these pragmatic tests, not an ideological preference.

Competition that benefits developers

Competition between American, Chinese, European and open models has an immediate positive effect for African developers: prices fall and choice widens. The big American labs are in fact multiplying programmes and agreements in Africa, with governments as well as companies, aware that the continent will be one of the big growth markets for AI.

Lexi Novitske’s message is above all a warning to Western investors: the technology infrastructure of tomorrow’s Africa is being chosen now, in the technical decisions of thousands of small teams. They choose what is affordable and available. If American players want to carry weight, they will have to offer terms suited to the continent’s economic realities.