I grew up drinking powdered milk. It’s the only milk that existed in my world in Kinshasa, until I moved to South Africa at 19 years old and tasted liquid/fresh milk for the first time. I remember that moment clearly, because it wasn’t neutral. It didn’t feel like an upgrade. It felt strange. The surprising part to this day is that even with having had every opportunity to switch, I still prefer powdered milk. Not out of nostalgia I perform for effect but out of actual taste preference, formed early and never really displaced.
I tell that story because there’s a version of the African AI adoption conversation happening right now that keeps circling the same handful of barriers: device affordability, connectivity, skills, and infrastructure, and the literature also repeatedly adds data availability, policy/governance, and trust/ethics as major constraints. Those are real, and they’re documented, smartphone penetration across lower-income countries often sits close to a half-level of what it is in wealthier ones, and the cost of a device or a data plan genuinely keeps people offline. What’s striking is what’s missing from that list entirely. Search widely for coverage of AI adoption barriers in Africa and you’ll find affordability, connectivity, skills, compute and almost nothing on billing infrastructure. Nobody’s writing “AI companies won’t accept mobile money” as a named barrier. It’s not that the affordability story is wrong. It’s that an entire layer of the problem, the one my own relationship with powdered milk turns out to explain rather well, isn’t being told at all.
Kenyan entrepreneur Tonee Ndungu recently reignited this conversation with a sharp observation: AI platforms lock Kenyan businesses out not through the technology itself, but through billing. This applies to the continent too. Card-tied monthly subscriptions, built for economies where people get paid biweekly into a bank account, simply don’t match how money moves for a continent running on mobile money and daily micro-transactions. His argument isn’t about casual users dabbling with a chatbot. It’s about creators and small business owners trying to scale, to move from trying a tool to depending on it and hitting a wall that has nothing to do with vision or ambition. It’s a fair point and nothing that follows here weakens it. If anything, the argument holds up better once you see it’s about a choice rather than a constraint, a choice is changeable in a way a hard constraint isn’t. But “can they afford it” smuggles in an assumption worth examining, and my own milk preference is where I’d start examining it.
There’s more than one African milk story.
In the DRC, the domestic dairy industry remains small relative to demand, and the country leans heavily on imports to meet consumption. Powdered milk tends to win because it is shelf-stable, doesn’t require refrigeration, travels easily, and is usually the most practical and affordable way to supply dairy in markets without a strong fresh-milk cold chain. Nigeria tells the same story for the same reason: powdered milk dominates because it is cheaper, shelf-stable, and easier to distribute in a market where fresh-milk infrastructure is weak and refrigeration is unreliable.
Kenya breaks the pattern completely. Most households buy raw, unprocessed milk through informal channels and boil it at home; it is preferred because it is cheaper, fresher, and better tasting, while powdered milk remains a minority product rather than the default choice.
South Africa is different again. A mature domestic dairy industry makes fresh milk the default, while UHT plays a big role because it is shelf-stable and cheaper than fresh milk. Powdered milk is present too, but much of its use is tied to infant formula and processing rather than everyday drinking milk.
The clearest outlier is pastoralist Ethiopia’s Somali region, and it’s worth pausing on this one rather than rushing past it. In a study of camel milk, there was no meaningful willingness-to-pay gap between fresh and powdered milk, and about 31% preferred the powder anyway, mainly because it was more practical in a cold-chain-poor, mobile, arid setting. Price wasn’t steering the decision at all. Fit was.
Five countries, five different relationships between milk, money, and infrastructure, not a single African story about powder being the practical, thrifty choice, and not a single story about fit beating price either. My own life sits inside one version of it: a market with almost no fresh-milk industry (powder, genuinely practical) followed by a move into a market with a fully developed one (fresh, genuinely default) and a taste preference that stuck around anyway, price and infrastructure be damned. That’s path dependence: an early choice, made under a real constraint, keeps shaping what you do long after the constraint disappears. I’m one data point in a five-country pattern that doesn’t reduce to a single sentence, and that’s the point.
The easy move here is to map my milk story straight onto AI. Constraint forces adaptation, adaptation becomes tomorrow’s infrastructure, and today’s billing workarounds (virtual dollar cards, informal payment bundling) settle into permanence the way DRC’s and Nigeria’s powder markets did. It’s a decent story. It’s just not quite the right one, and the gap between the two versions is where the real insight lives.
Milk can’t prove anything about AI billing on its own. No analogy carries that kind of weight, and I don’t think it should have to. What it does is give a concrete, lived sense of two ideas that matter here: that path dependence is real, and that “African AI adoption” was never going to be one story with one cause. The AI argument has to stand on its own evidence from here, so here it is.
Powdered milk’s various roles across the continent were all, in their own way, responses to a scarcity that was internal: thin or absent domestic dairy infrastructure in places like the DRC and Nigeria, a fully developed one in South Africa, something else again in Kenya’s informal markets. Whatever the shape, nobody outside Africa was withholding fresh-milk logistics from any of these countries. The constraints were real, structural, and nobody’s fault in particular.
The AI billing wall is a different kind of problem. Mobile money isn’t a gap in African infrastructure and in many markets it’s more advanced and more widely used than the equivalent systems in the countries where these AI companies are based. Kenya has been running M-Pesa since 2007. The infrastructure billing would need to plug into already exists, proven at scale for almost twenty years.
To be fair to the companies here: building toward it isn’t trivial. Integrating with dozens of fragmented telco and fintech systems across markets (M-Pesa in Kenya, MTN MoMo across Ghana and Uganda, Orange Money across Francophone Africa), each with its own compliance regime, forex handling, and settlement rules, is a genuinely harder problem than accepting one global card network. That’s a real cost, not an excuse invented to dodge blame. But harder is a different claim from impossible, and Google Play’s own decision to accept Airtel Money as payment in Kenya, no bank card, no foreign wallet required, is proof the second claim doesn’t hold even where the first one does. What’s missing isn’t capability on either side. It’s priority and priority is the kind of thing that changes when enough people name it as a choice rather than accept it as a law of nature.
There’s another assumption buried in this conversation, and it’s easy to miss because it sits one level below the billing question: what actually counts as adoption?
The instinct is to measure it by tier. Free-tier use looks like browsing. Paid, API-level use looks like real, dependent, economically meaningful adoption. But that split falls apart the moment you picture an actual person using the tool. A freelancer in Lagos drafting a client brief on a free ChatGPT account, getting paid for the finished proposal, has generated real income from AI, inside the tier the billing conversation treats as the unscaled, not-yet-adopted layer. A small business owner in Kampala using a free account to write ad copy that sells product has done the same thing. The money is real. The tier is free.
Which means the question isn’t really “what counts as adoption” at all. It’s a narrower, more useful one: is AI output actually earning someone money, on whatever tier that happens to be? Nobody is tracking that number. Not the AI companies, and not the people writing reports about barriers to AI adoption in Africa. It’s invisible because no one is set up to see it, not because it isn’t there. And that has a real cost, it lets the whole conversation about African AI adoption run on the wrong scoreboard (subscriptions, sign-ups, revenue per user) while missing the most resourceful, least visible layer of actual use.
None of this means the free tier is enough, or that the billing wall doesn’t cost people something real. Rate limits and weaker models are a genuine ceiling, not a workaround, and nothing here should be read as evidence that people are fine on the free layer. What it does mean is that the current way of measuring the cost of the wall can’t see the floor of value already happening beneath it. A freelancer running her entire client pipeline through a free account this month may be getting more real value out of AI than a funded startup paying for API access it barely touches. Counting by tier can’t see that. Counting by affordability can’t either. Both are measuring the wrong thing, from opposite directions. The wall is still a problem. It’s just not the problem the affordability chart thinks it is.
Which brings the billing question back into focus, sharper than before. If real economic value is already moving through free-tier accounts in spite of the wall, then the wall was never actually stopping adoption. It was stopping the version of adoption AI companies know how to count. People kept building around it anyway, on the tier that isn’t supposed to count as real use. That’s worth sitting with before getting to what mobile money already proved about what happens when infrastructure ignores how people actually use things.
Mobile money itself is the precedent, and it’s not a small one.
Before M-Pesa launched in 2007, many Kenyans without bank accounts sent money through informal bus and matatu networks, along with family and friends. Those channels were common but risky, while the formal options were limited, costly, and inconvenient. Formal alternatives existed. Kenya Post ran a remittance service. What won wasn’t the cheapest option; it was the one that fit how people already moved money and gave them something the informal channels couldn’t: a trustworthy record of the transaction. Safaricom built M-Pesa around existing behavior rather than trying to move people onto a foreign banking model, and that’s the part banks in Kenya hadn’t done.
It didn’t just fill a gap, it became an export. Vodafone took the M-Pesa model to Afghanistan in 2008, India in 2013, and Romania and Albania in 2014 and 2015, targeting markets with large unbanked populations. CNN later covered M-Pesa’s entry into Europe, and Reuters reported that Snapchat’s cash-transfer feature drew inspiration from the same basic idea. And the localization Tonee is asking AI companies to build already exists as a live commercial pattern elsewhere in the ecosystem: in Kenya, Google Play itself accepts Airtel Money as a payment method for app and content purchases. If Google can build that plumbing into its own store, the absence of it in AI billing looks less like a technical wall and more like an unmade decision.
Worth being honest about the limits of this comparison too. Safaricom had something most African AI builders don’t: near-monopoly telecom infrastructure, deep state relationships, and years of runway before any real competitor arrived. That’s a real asymmetry, and it means “Africans will just build around it” isn’t a guaranteed outcome the way it can sound. Repeating M-Pesa’s exact conditions isn’t realistic. But the direction of travel, friction producing improvisation, improvisation hardening into infrastructure, doesn’t require repeating those exact conditions to be worth watching for. The virtual dollar card services and workaround billing patches emerging now to route around AI’s payment friction might be the seed of the same pattern in miniature, at a smaller scale than M-Pesa achieved, but a real one.
Which reframes the debate from whether AI companies will eventually notice and build the localized billing rails Tonee is asking for, to whether they’ll get there before African builders route around them entirely and build something better, even a smaller, scrappier version of “better” than M-Pesa’s.
The same logic that shaped five different milk markets is almost certainly shaping AI adoption too. Some countries will absorb billing friction through virtual card workarounds and keep building on top of whatever the major AI platforms offer. Others, with stronger fintech ecosystems and more institutional appetite for independent infrastructure, may end up constructing something more structural. Neither path is more authentic than the other. They’re rational responses to different starting conditions.
This is probably the wrong moment to look for a single African AI adoption story, in the same way there was never a single African milk story. What’s more useful, and more accurate, is thinking in terms of one continental ambition sitting above many national pathways: shared principles around digital sovereignty, responsible AI development, and cross-border collaboration, with each country free to build the specific infrastructure that fits its own economic texture underneath that shared frame. Not one AI ecosystem. A network of them, linked by cooperation and shared standards, diverging by design rather than by default.
The affordability story isn’t wrong, and it isn’t going away, devices and connectivity are real costs that keep real people offline. But it’s also the only story getting told, and that’s the problem. A barrier that lives entirely in a company’s product roadmap doesn’t show up in an affordability chart at all. It disappears into the same bucket as “can’t afford the device,” when it’s actually a different kind of obstacle: a choice, not a law of nature, even a difficult and defensible choice.
And the milk evidence points to something sharper still: cutting the price wouldn’t fix this even if price were the whole problem. South Africa’s most expensive milk tier is powder, and people still buy it. Ethiopian pastoralists paid the same for both and still chose powder. Price isn’t steering either outcome. So the obvious fix everyone often reaches for, just make AI subscriptions cheaper for African markets, would likely fail the same way. A discounted plan that still demands a foreign card and a rigid monthly cycle solves nothing, the same way cheaper fresh milk, where it exists, still loses out to whatever actually fits how people live.
Mobile money showed what happens when existing systems fail to meet people’s needs: new ones emerge, imperfectly, unevenly, but they emerge. AI now faces the same test. Whether global AI companies pass it may matter less than whether Africans once again decide to build the solutions themselves.
Thank you for reading!
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.