Africa's Ambitious Machines: Unseen, Undervalued, Undeniable

Africa's Ambitious Machines: Unseen, Undervalued, Undeniable

The most significant AI gap is in our collective understanding of what it is and what is possible.

By: Osman Siddiqi, Co-Founder & CEO at Acclimate, Board Member of Legado Initiative and Manna Tech Ltd

A blog featured on the MIT Kuo Sharper Center’s website was shared with me—it had a headline designed to hook the audience it is intended for: “Africa is not ready for the AI era, and few are willing to say it.” And it did hook me. The author gets real things right: the infrastructure gaps, the compute deficit, the sovereignty risks around data. All valid.

The diagnosis in the piece, however, in trying to cut through false optimism, ends up reinforcing conceptual errors that produce that optimism in the first place. It treats AI as a single thing, with a single set of preconditions required to participate, and a single race worth running. A more important question is which sport should we be playing in the first place? And it matters because the strategy you build, based on the wrong frame or understanding, will be wrong in ways that will be difficult for stakeholders to see until the investment has already gone somewhere it shouldn’t have, costing us decades of potential and opportunity.

First, hear Arvind and Sayash out before hearing me out. It will help.

Drs. Arvind Narayanan and Sayash Kapoor, authors of AI Snake Oil, have a useful analogy for the root problem. Picture a world with no words for different forms of transport, just the single collective noun “vehicle.” Debates about whether vehicles are environmentally friendly become incoherent fast, because one side is talking about motorbikes and the other about aircraft. As they put it, “in everything we call AI, the underlying technologies often have nothing to do with each other.”

The Africa AI readiness debate runs on the same confusion. Image recognition systems, crop disease classifiers, food security early warning tools, GPT-4 — they are all “AI” but rarely spoken about in the same way, with AI often being conflated with the capabilities of the latest large language model. The aforementioned list shares little in common in terms of infrastructure requirements, data needs, compute costs, or being the right model for the right use case fit. Benchmarking Africa’s readiness against US and Chinese frontier model development is, in this sense, a categorical and strategic error. The article is probing whether the continent can keep up with aircraft manufacturing when the more relevant question is whether it is building good bikes and cars and buses.

There is a second component that the piece misses entirely. The techno-policy ecosystem only calls something “AI” when it sits at the cutting edge, when the implications and use cases are still being argued over and the optimists and pessimists take center stage. However, once an AI tool works well enough, we stop noticing it. Perhaps this is also a cultural zeitgeist or temporal frame—anything before AI became accessible to the masses through apps and software is not seen and understood as AI. This is despite the progress that has been made in many African countries.

I’ll elaborate on this. Three domains make the point concrete, and each works through a different mechanism. The examples below are by no means exhaustive.

• Agrifintech: Apollo Agriculture’s credit scoring engine, trained on satellite imagery, mobile data, and farm repayment histories, now reaches over 350,000 smallholder farmers in Kenya and Zambia. In Tanzania, Tigo Nivushe has extended over 130,000 loans to farmers based on mobile money activity rather than land ownership or formal employment.
• Health diagnostics: South Africa and Zambia are running computer-aided TB detection from chest X-rays via CAD4TB—and automated malaria detection is under evaluation in Kenya and Ghana. Makerere University’s AI Health Lab in Uganda built algorithms that read blood smears and diagnose malaria at accuracy comparable to a specialist, and it built them in-house. A federated learning project across eight African countries lets hospitals collaboratively interpret chest X-rays without sharing raw patient data.
• Geospatial Analysis: Conflict and food security, a global effort (FEWS NET) has deployed machine learning (ML) across decades of accumulated reports and real-time global media scanning to catch early signals of emerging crises, feeding directly into humanitarian response and government decisions at continental scale. Food insecurity in the Horn of Africa is now being reasonably preempted up to three months in advance. National-scale land cover mapping at 30-metre resolution, using random forest classifiers on satellite imagery, is feeding natural capital accounting in Liberia and Gabon. Much of this model is internationally funded, but it is operationally deployed in constrained environments and informs sovereign decisions for countries.

The Consequences of the Wrong Conceptual Frame

The AI/ML applications delivering real value across the world have become invisible as “AI.” That invisibility has consequences. If the previous examples running food security early warning, tuberculosis diagnostics, and smallholder credit ratings do not register as AI, the policy conversation therefore concludes that Africa either has to build enormous capabilities or realistically has no technological basis to drive AI use cases. That conclusion is wrong. And it produces the wrong strategy.

These systems are the foundation from which the next layer develops. Practitioners building crop disease classifiers are simultaneously developing the data systems, training pipelines, domain expertise, and institutional relationships that make more ambitious applications possible later if we can accelerate the right culture of learning. Even so, the learning curve is active. Treating it as invisible means systematically undervaluing what is being built and redirecting investment toward a frontier race that is both unwinnable and, more to the point, beside the point.

The wrong question is whether Africa (in itself an arbitrary boundary of nations with extraordinary variation) is ready for the AI era. The right question is longer and needs a breath: which AI techniques produce the highest value against real needs given current constraints, and what should be built now for dividends in twenty years? That question is specific, tractable, answerable, and already being responded to on the ground.

The sovereignty layers, as described in the article, and GPU count frame answers a different question entirely; one about competing at a massive, large-language-model layer, which was never a realistic ambition for most of the world other than a handful of countries.

For example, DeepSeek showed what is possible to build under significant constraints, imposed or otherwise. While it is true that its ingredients—elite engineering talent, institutional resources, and state adjacency—are not automatically replicable, DeepSeek demonstrates that the costs of the underlying models are drastically reducible. Training and inference costs have dropped enormously since 2022. Streamlining and narrowing existing models for specific use cases, local languages, crop varieties, and environmental and climate patterns is now a materially different possibility than it was a few years ago.

This is not an argument that Africa’s deficits are trivial — they are genuinely quite real. It is an argument against treating organizational and infrastructure gaps between a leading superpower and a low GDP per capita continent that is more vast than most places on the planet combined as the problem to solve.

Many developing countries today sit in a position not entirely unlike China in the late 1990s or early 2000s—behind the frontier, infrastructure constrained, but with the potential to build capability against real problem sets with real needs. The solutions built for constrained, high-stakes, low-infrastructure environments are not just locally valuable and genuinely needed by Africans and other denizens of the global south. They are exportable to a large share of the global market facing structurally similar conditions. That is an extraordinary opportunity. Whether frontier capabilities follow in ten years or thirty is genuinely uncertain and not the point; a race is truly not the right frame of mind or framework.

There is a further pull at work driving invisibility of the wider scope of what is possible with AI. Technology that is technically sharp, genuinely fit for purpose, and unambiguously AI in reality starts to look insufficient, or even unsexy, when held against what is seen as the much sexier frontier. Much of this is coloured by the experience of everyday users—which includes serious decision-makers of policy and finance—and the cultural weight of a DeepSeek or Claude. For most people now using AI with those tools, the breadth of what actual tools and methodologies cover are rendered invisible. 

This is where the vehicle analogy matters even more, not just as a definitional point but as a psychological one. When the only vehicle anyone talks about is the aircraft, a very good bus that Kenya produces may start to feel uninteresting, and we are made to think we must try to make the bus fly for it to be worthwhile, rather than make them cleaner, safer, more connected, and more comfortable. That pull shapes what gets financial and cognitive attention, what gets built, and of course what gets dismissed before it is even tried.

Which brings us back to the article’s specific prescription. The six layers of sovereignty in the article are a legitimate agenda that should be pursued. But those layers are not preconditions to deliver technologically driven value now. Regional integration lowers costs, aggregates resources, improves negotiating positions, and accelerates all that we want. All extremely valuable. But framing those layers as preconditions implies telling decision-makers on and off the continent to wait for a foundation that may take years to decades to materialize.

The systems running quietly today, predicting floods, diagnosing TB, scoring credit for farmers are AI. And more such systems as well as more sophisticated systems that solve real problems are possible to build now given the expected trajectory of capabilities and cost. The debate about whether Africa is ready for sophisticated solutions needs to be led by what is needed and what is possible to build given what we know, where we are, and where we can go.

The tools being built across Africa are not just technically appropriate responses to constraint. They are choices. Choices being made about what problems matter, whose data gets used, which communities benefit, and what AI is actually for. That is not a consolation prize, it is sovereign, even if fragmented, decision-making. It is a more interesting, more politically responsible, and frankly more important set of questions and what we should all be asking. 

The frontier should be defined not just by technical advancement, but by what the technology is used for, specifically who shapes that use, and importantly, who absorbs the costs. An enormous, diverse continent with a young, innovation-hungry, at times struggling population is more than capable of expanding that imagination on its own terms. It does not need influential institutions to define the ceiling.

The deeper issue is the very real skepticism and costly directions the wrong narratives generate, a skepticism that causes people who could be building something worthwhile to halt and people who could be funding what is worthwhile to look elsewhere for shinier objects that resemble the frontier they are familiar with. If this topic is not addressed accurately, looking at the wider suite of what is possible technologically through AI or otherwise, we will find ourselves on a pathway where we are not valuing real solutions by real expertise we already have to drive the next phase of African innovation.

A brief note of immense gratitude to all the reviewers who helped make the article clearer and richer.

References

Narayanan, A. & Kapoor, S. (2024). AI Snake Oil: What Artificial Intelligence Can Do, What It Can’t, and How to Tell the Difference. Princeton University Press. SSIR excerpt: https://ssir.org/books/excerpts/entry/ai-snake-oil

GSMA Mobile for Development. “AI-driven smallholder farmer lending in Africa: Insights from Apollo Agriculture.” December 2025. https://www.gsma.com/solutions-and-impact/connectivity-for-good/mobile-for-development/programme/agritech/ai-driven-smallholder-farmer-lending-in-africa-insights-from-apollo-agriculture/

OptimusAI. “AI Credit Scoring: How Mobile Money is Lending to the Unbanked.” May 2025. https://optimusai.ai/ai-credit-scoring-mobile-money-unbanked/

Maruta, C. et al. “The role of artificial intelligence in diagnostics: A new frontier for laboratory medicine in Africa.” African Journal of Laboratory Medicine, 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12505451/

Rech, D. “The Ethics of AI-Driven Health Projects in Africa.” Think Global Health https://www.thinkglobalhealth.org/article/the-ethics-of-ai-driven-health-projects-in-africa

Busker, T. et al. “Predicting Food-Security Crises in the Horn of Africa Using Machine Learning.” https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2023EF004211

Brandt, M. et al. “Cloud-computing and machine learning in support of country-level land cover and ecosystem extent mapping in Liberia and Gabon.” https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0227438

Independent reports - DeepSeek’s R1 reasoning model runs 20–50× cheaper than OpenAI’s comparable model. Stanford HAI data shows AI inference costs dropped over 280 times between 2022 and 2024. Sources:

Tunguz, T. “The AI Cost Curve Just Collapsed Again.” tomtunguz.com, 2025. https://tomtunguz.com/deepseek-sputnik/

IntuitionLabs. “DeepSeek’s Low Inference Cost Explained.” https://intuitionlabs.ai/articles/deepseek-inference-cost-explained

Alan Turing Institute. “Brief analysis of DeepSeek R1 and its implications for Generative AI.” arXiv, February 2025. https://arxiv.org/pdf/2502.02523

Carbon Brief. “AI: Five charts that put data-centre energy use and emissions into context.” September 2025. https://www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context/

Zhu, Wang, et al “China’s model of technology leapfrog: A case study of electric vehicle policies and the development of green technology.” Renewable and Sustainable Energy Reviews, 2025.https://www.sciencedirect.com/science/article/abs/pii/S1364032125010871