The Leapfrog That Actually Happened

The word “leapfrog” has been abused so thoroughly in development economics that serious researchers mostly stopped using it. Every new technology becomes, in the pitch decks and policy papers, the thing that will let poor countries skip a stage of industrial development and arrive somewhere better. Mobile banking was supposed to do it. Solar microgrids were supposed to do it. The leapfrog almost never happens the way the optimists describe.

What actually occurred between 2025 and 2029 with AI adoption in lower-income countries is stranger, messier, and in some ways more impressive than anything the optimists predicted. It is also, in other ways, a cautionary story about dependency that the same optimists are still mostly ignoring.

Start with the basic mechanism. The earlier leapfrogs that actually worked — mobile money in Kenya, cheap smartphones in Bangladesh and Indonesia — worked because the underlying infrastructure was genuinely different. Africa didn’t need to build landline telephone networks before building mobile ones. The old infrastructure was the obstacle, and its absence turned out to be an asset. M-Pesa didn’t have to fight legacy banking systems. It built something new on top of cheap SMS.

AI adoption in 2025-2029 followed a similar logic but at a different layer of the stack. The relevant infrastructure isn’t copper wires or bank branches; it’s computation and connectivity. And here, something genuinely unexpected happened. The proliferation of cheap Android devices in the 2018-2023 period left a population of several billion people accustomed to smartphones — often quite powerful ones — with no particular attachment to desktop-era software paradigms. When capable AI models became available at low cost via API, or (more importantly) when capable small language models became runnable on mid-range devices without connectivity, these populations could engage with them immediately. There was no installed base of enterprise software to protect, no IT department pushing back on change.

The healthcare case in Nigeria illustrates the mechanism. By 2027, the Nigerian government’s partnership with three local AI companies had put a diagnostic support tool on the phones of roughly 40,000 community health workers — people with limited formal training operating in rural areas where physicians are scarce. The tool doesn’t diagnose. It surfaces questions, flags symptoms that warrant escalation, helps workers fill out referral forms accurately. The outcomes data published in early 2029 showed a 23% reduction in late-stage malaria presentations in the pilot districts. That is not a small number. That is not a rounding error. That is a technology doing something measurable in a healthcare system that had no realistic path to achieving the same result through conventional means — training enough physicians, building enough clinics — within any politically realistic timeframe.

But here’s what the leapfrog optimists tend to elide: the diagnostic tool runs on servers in Lagos and Amsterdam. The model was trained primarily on data from high-income countries, fine-tuned on a comparatively small Nigerian dataset. When the partnership’s terms were renegotiated in 2028, the Nigerian health ministry had limited leverage. The technology was embedded in their health system in ways that made switching costly. That’s not a catastrophe, but it’s not sovereignty either.

The contrast with India’s trajectory is instructive. India’s AI development between 2024 and 2029 was, by most metrics, the most successful in the developing world — and it succeeded precisely because India refused to treat AI as a consumer good to be imported. The government’s investment in domestic compute infrastructure (the National AI Compute Initiative, announced in 2024 and substantially delivered by 2027) meant that Indian companies were training models on Indian data with Indian hardware. The result is a set of language models capable of handling India’s genuine linguistic complexity — not just Hindi, but Tamil, Telugu, Bengali, and a dozen other languages — in ways that no foreign model matches. When Indian agricultural extension services now give smallholder farmers advice about pest management, the advice comes from a system that understands local crop varieties, local soil conditions, local market prices. That specificity matters enormously.

The Indian path required something most countries can’t easily replicate: a large domestic market capable of justifying the investment, a pre-existing base of technical talent, and a government willing to make long-term commitments in the face of short-term pressure to just buy cheaper foreign solutions. Most of sub-Saharan Africa, most of Southeast Asia, most of Latin America lacks at least two of those three conditions.

What this creates is a bifurcated outcome. Countries with the capacity to develop domestic AI capability — India, Brazil, South Korea (already rich but instructive), and to a lesser extent Indonesia and South Africa — have done so. Countries without that capacity have become consumers of AI infrastructure built elsewhere. The consumption is often genuinely valuable. A Ghanaian farmer using an AI-powered market price tool is better off than a Ghanaian farmer without one. But the value chain terminates in Accra, not in Mountain View or Shenzhen.

The Chinese influence story complicates this further in ways that Western analysts have been slow to engage with seriously. Chinese AI companies — Baidu, Alibaba Cloud, and a cluster of smaller firms — have been extraordinarily aggressive in East Africa, Southeast Asia, and parts of Latin America. Their pitch is explicitly differentiated from American offerings: cheaper, willing to operate in local regulatory environments without complaints, often bundled with infrastructure investment. The terms of those relationships deserve scrutiny. Data from user interactions flows to servers subject to Chinese law. In several documented cases, governments that accepted Chinese AI infrastructure deals also accepted provisions about data sharing that would be politically unacceptable in European markets.

None of this is simple. An Ethiopian farmer who gets better crop yield predictions from a Chinese-built system is getting a real benefit, whatever the geopolitical context of that system’s funding. Development economists have learned, mostly the hard way, to be suspicious of arguments that reject real improvements in the name of abstract principles. The question isn’t whether people in lower-income countries should use AI — they should, and they are — but whether the current configuration of dependency is the best achievable outcome, or whether it’s a path dependency that will be difficult to escape later.

The mobile money analogy is worth holding onto here. M-Pesa was built by a Kenyan company (technically majority-owned by Vodafone, but operationally Kenyan) on infrastructure that Kenyans controlled. The success of that model — and it was a genuine success, measurably improving financial inclusion and enabling small business formation — gave Kenya both the economic benefit and meaningful control over the terms. That combination is much harder to achieve in AI, where the capital requirements for frontier model development are orders of magnitude higher than for SMS-based payment systems.

The countries that have navigated this best are the ones that identified specific, bounded domains where domestic capability was achievable and worth the investment, rather than trying to compete across the board. Rwanda’s investment in AI for its national health insurance system is a reasonable example: a specific application, built on open-source models fine-tuned with local data, operating within a regulatory framework that Rwanda controls. The ambition is smaller than India’s national compute initiative, but it’s proportionate to Rwanda’s actual capacity, and it produces something Rwanda owns.

The leapfrog happened. Just not in the shape anyone drew in the PowerPoint.

The more honest framing is that AI created a new surface area for the old development challenge: how do you generate economic value from new technology without becoming permanently dependent on the foreign entities that control it? That challenge doesn’t have a universal answer. It has country-specific answers that depend on market size, technical capacity, political will, and — not to be underestimated — luck in timing. The countries that started building domestic AI capability in 2023 and 2024, when it was expensive and uncertain, are in a fundamentally different position than the ones that waited for the technology to prove itself.

Waiting was the rational choice at the time. It almost always is. That’s why the leapfrog is always harder than it looks from a distance.