Leapfrogging Opportunities

In the retinal imaging room of the Aravind Eye Hospital in Madurai, a technician photographs the back of a patient's eye. Within seconds, software flags whether the image shows signs of diabetic retinopathy — the leading cause of preventable blindness among the roughly 100 million Indians living with diabetes. The point of the system is not that it out-diagnoses a specialist. It is that India does not have enough specialists. There are tens of thousands of ophthalmologists for a population of 1.4 billion, most of them clustered in cities, while the disease grows fastest in villages. An algorithm reading a photo taken by a nurse on ordinary equipment does something no expansion of medical schools could do on the same timeline: it puts a screening decision where the patient actually is.

That is the whole idea behind what researchers and the World Bank have started calling "Small AI" — and it is a useful corrective to the way most people picture artificial intelligence. The frontier version of AI is enormous: models with hundreds of billions of parameters, trained in data centers that draw as much power as a small city, accessed over fast broadband on expensive hardware. Small AI is the opposite discipline. It means designing applications to run in constrained environments — low bandwidth or none, intermittent electricity, cheap Android phones, users who are not engineers — and doing so not as a compromise but as the point. A tuberculosis-screening model that reads a chest X-ray on a laptop in a clinic with no reliable internet is not a lesser version of AI. In the place where it runs, it is the only version that matters.

The central question of this chapter is whether Small AI lets the Global South leapfrog — skip the slow, expensive intermediate stages of development that rich countries climbed one rung at a time — or whether the constraints that make leapfrogging necessary also make it fail. The honest answer, laid out over the following pages, is that leapfrogging is real, it is documented, and it is nowhere near automatic.

The evidence that adoption is genuinely happening

Start with the fact that developing countries are not waiting to be given AI. They are already using it, at a scale that surprised even the institutions tracking it. By mid-2025, more than 40 percent of ChatGPT's global traffic originated in middle-income countries, led by Brazil, India, Indonesia, and Vietnam (World Bank, 2025). Job markets tell a similar story: globally, vacancies mentioning generative AI skills rose roughly ninefold between 2021 and 2024, and about one in five of those postings were in middle-income economies. These are not the numbers you would expect if AI were a luxury good confined to the wealthy world.

What the traffic data cannot tell you is whether this engagement is deep or shallow — whether a student in Jakarta using ChatGPT to draft an essay represents a development gain or just a new consumer habit. Adoption is a precondition for leapfrogging, not proof of it. The interesting evidence lies not in how many people use a chatbot but in whether purpose-built tools change outcomes on the ground. That evidence exists, but it is thinner and more uneven than the enthusiasm around it, and separating the two is most of the work.

The precedent everyone reaches for — and where it breaks

The word "leapfrogging" entered development economics through the mobile phone. Across much of Africa and South Asia, countries never strung copper landlines to every village; when cheap mobile handsets arrived, they skipped that stage entirely and went straight to wireless. Kenya's M-Pesa is the canonical case: mobile money spread precisely because the traditional banking infrastructure was absent, so there was nothing entrenched to defend and enormous unmet demand to fill. Studies later credited M-Pesa with lifting hundreds of thousands of Kenyan households out of poverty by making it cheap and safe to move money.

The mechanism there is worth naming precisely, because it is the engine of every leapfrogging argument. Incumbent infrastructure is not just an asset; it is also a commitment. A country with a national landline network has sunk costs, employed workforces, billing systems, and regulatory habits all organized around the old way. A country with none of that has no such gravity holding it in place. The absence of legacy systems removes the switching cost. This is why a country without a formal credit bureau can jump directly to algorithmic credit scoring built on mobile-airtime and utility-payment data, reaching people who never had a bank record to begin with — there is no entrenched underwriting establishment to displace.

But the mobile analogy flatters AI in ways worth being careful about. A phone call is a phone call anywhere; the technology was largely context-independent, and once the tower was up, the service worked. AI is not like that. It is data-dependent — a diagnostic model trained on European skin or American retinas can quietly underperform on the populations it is deployed to serve, so the technology often has to be rebuilt locally rather than merely imported. It is compute-intensive in ways a voice call never was, so even Small AI leans on some cloud connection for updates and retraining. And it carries governance complexity — questions of privacy, liability, bias, and surveillance — that a mobile handset simply did not raise. The mobile precedent proves that leapfrogging can happen. It does not prove that AI will leapfrog the same way, because the thing being leapfrogged with is far more demanding of local data, institutions, and oversight.

Why designing for constraints makes solutions more valuable, not less

There is a counterintuitive claim buried in the Small AI idea: that engineering for scarcity produces tools that are more useful in the Global South than the frontier systems Silicon Valley builds. The logic holds up. A model optimized to run offline on a $80 phone reaches the clinic 200 kilometers from the nearest specialist; a model that assumes fiber broadband and a modern GPU does not. A crop-advice service delivered over SMS reaches the smallholder with a feature phone; a slick app assuming a data plan excludes exactly the farmers who most need advice. When the binding constraint is connectivity, hardware, and literacy — not raw model intelligence — the tool that respects those constraints wins, even if it is technically "worse." Capability you cannot deliver is capability that does not count.

This is also why large-scale, frontier AI does not automatically serve these markets. The frontier optimizes for the metric that impresses in Northern benchmarks: more parameters, longer context, harder reasoning. The Global South's binding problems are not usually solved by a smarter model; they are solved by a good-enough model that runs where the electricity flickers and the signal drops. Small AI closes that gap by making appropriateness, not raw power, the design target.

India: the closest thing to a proof of concept

If any country demonstrates what leapfrogging at scale looks like, it is India — and its example is instructive precisely because of what it required. India's national AI computing capacity has crossed 34,000 GPUs, one of the largest such buildouts in the developing world, much of it offered to startups and researchers at subsidized rates through the IndiaAI Mission, approved in March 2024 with roughly $1.25 billion in funding over five years. The country hosts one of the world's largest pools of AI and software talent, with millions employed in the sector.

The more important part is the delivery layer beneath the models — what India calls its "digital public infrastructure." Aadhaar gave more than a billion people a digital identity. The Unified Payments Interface (UPI) turned that identity into a financial rail processing well over ten billion transactions a month by 2025, letting a street vendor accept a payment with a QR code and no card machine. Onto these rails, AI applications can be layered cheaply: identity-verified telemedicine, direct-benefit transfers, agricultural advisories reaching smallholders, retinal and TB screening in public clinics. These are operational systems serving hundreds of millions of people, not pilots.

The uncomfortable footnote is that almost none of this was cheap or fast or replicable by wish. It rested on sustained public investment, more than a decade of institution-building, unusual political continuity behind the digital-identity project, and the sheer structural advantage of scale — a population large enough that the data and demand needed to make AI systems viable were simply there. That last advantage is not a policy choice another country can adopt. It is why the honest question is not "can others copy India?" but "what can others do who cannot?"

How well are the success stories actually documented?

The book's standard is to distinguish what we know from what we hope, and Small AI is a domain where that discipline matters, because the marketing runs well ahead of the evidence. The rigor varies enormously by sector.

Agriculture has the firmest ground. Randomized evaluations of mobile agricultural advisory services — Cole and Fernando's study of Precision Development's SMS and voice service in Gujarat (2021) among them — found that farmers who received the advice were meaningfully more likely to adopt recommended practices, with measurable effects on input use and some yield gains. That is the kind of evidence, a controlled trial with a real counterfactual, that most Small AI claims lack.

Healthcare sits in the middle. There is strong published work on the diagnostic accuracy of AI retinal screening and on computer-aided detection for tuberculosis — the World Health Organization has recommended CAD software for TB screening in settings short of radiologists. What is much thinner is evidence on downstream patient outcomes: whether earlier detection actually translates into less blindness or lower TB transmission at population scale, once you account for whether flagged patients receive follow-up care. The algorithm's accuracy is documented; the health system's ability to act on it often is not.

Education is the most hopeful and the least proven. A widely discussed World Bank study in Nigeria (2025) found that an after-school program using a large language model as a tutor produced learning gains of roughly 0.3 standard deviations over six weeks — an effect the authors placed among the larger ones in the education literature. It is genuinely encouraging. It is also a single, short, small study, not yet the replicated, at-scale evidence base that would justify confident claims about AI transforming schooling across the Global South.

Finance is documented mainly on the access margin. Alternative-data lenders and mobile-money platforms have unquestionably brought millions of unbanked people into the formal financial system. Whether that access improves their welfare is more contested — the same tools that extend credit can extend over-indebtedness, and the evidence on net well-being is mixed. Inclusion is real; its consequences are not uniformly benign.

The overall picture: leapfrogging is not a slogan without substance, but the substance is patchier than its champions imply. We know these tools can work. We know far less about how often, how durably, and at what scale they do.

Why accessible technology still doesn't leapfrog on its own

Here is the trap. Suppose the models are open, the phones are cheap, and the app runs offline. Development impact still does not follow automatically, and the reasons are not about the technology.

Leapfrogging redirects infrastructure requirements; it does not abolish them. A crop app that runs offline still needs periodic data updates, which needs some connectivity. Any of it needs electricity, which across large parts of sub-Saharan Africa remains unreliable or absent. It needs devices capable of running the software, and in the poorest settings handsets are older and scarcer. And it needs users with enough digital literacy to trust and operate the tool. Countries at "stage zero" — without basic power or connectivity — cannot leapfrog with AI any more than a village without roads could leapfrog with cars. AI can help a country move faster once the foundations exist; it cannot substitute for the foundations.

Above the infrastructure sits everything the World Bank and development researchers converge on as the real determinant: policy and ecosystem. Open access to models and data, so countries adapt existing work rather than build from scratch. Skilled local workers — not just researchers but technicians, implementers, and clinicians who can deploy and maintain systems. Regulation calibrated to protect citizens without strangling adoption, since both over- and under-regulation carry costs. Local innovation ecosystems tying startups, universities, and government together. None of these alone is sufficient; all of them are necessary; and assembling all of them at once, in a country juggling poverty and public health and thin institutions, demands a coherence of effort that is genuinely hard to sustain. The gap between what is possible in principle and what gets built in practice is not a gap in technology. It is a gap in electricity, training, governance, and follow-through — and that is exactly where the potential most often stalls.

What smaller, poorer countries can do without India's advantages

If India's model rests on public investment, institutional depth, and population scale that a Rwanda, a Nepal, or a Democratic Republic of Congo cannot replicate, then telling them to "be like India" is useless advice. The more realistic pathways run around scale rather than through it.

The first is to borrow the delivery layer instead of building it. India has begun offering its digital-public-infrastructure stack — the identity and payments architecture — as shareable, open technology through initiatives that other countries have started to adopt. A small country need not spend a decade inventing its own payment rail if it can adapt one that exists. The second is regional pooling: several small economies sharing compute, data standards, and regulatory frameworks can reach a viable scale together that none could reach alone, much as small nations pool resources for other public goods. The third is to specialize narrowly — pick the two or three problems where a Small AI tool has documented impact, such as agricultural advisories or TB screening, and deploy those deeply rather than pursuing a broad AI strategy the institutions cannot support. And the fourth is to lean hardest on open models and international technology transfer, so that the country adapts rather than originates. None of these substitutes for local capacity. But they lower the bar from "build an AI ecosystem" to "deploy proven tools on borrowed infrastructure," which is a bar a weaker state can actually clear.

Is the window opening or closing?

The trend cuts two ways at once, and both are true. On one side, the conditions for leapfrogging are improving fast. Connectivity keeps spreading, devices keep getting cheaper, and — crucially — the cost of running a given level of AI capability keeps collapsing. Open-weight models good enough for most Small AI tasks are now freely available, and the compute needed to run yesterday's frontier model on today's cheap hardware falls every year. In that sense the door is opening wider.

On the other side, the frontier is sprinting away from the periphery. The most capable AI is advancing faster than Small AI can be adapted and deployed, and it is advancing in directions — vast context windows, agentic autonomy, multimodal reasoning — that assume exactly the compute and connectivity the Global South lacks. If the economic value increasingly concentrates at a frontier that only a handful of firms and states can occupy, then the relative position of countries relying on Small AI could deteriorate even as their absolute capabilities improve. Leapfrogging closes yesterday's gap while a new gap opens ahead. Whether the net trend favors the Global South depends on which curve moves faster — falling deployment costs or rising frontier concentration — and that is genuinely unsettled.

On timelines, realism matters. India took well over a decade to build the infrastructure that made its AI-for-development layer possible. A country at a significantly earlier stage of digital development should expect a comparable horizon for anything at India's scale — closer to ten or fifteen years than to three — even though individual tools, a screening program or an advisory service, can show results in a single growing season or school term. The pilots are fast. The systems are slow.

Who is responsible, and what should they do

The normative question is unavoidable: if leapfrogging is possible but not automatic, whose job is it to close the gap? No single actor has both the responsibility and the leverage, which is precisely the problem.

National governments hold the most leverage over the things that matter most — electricity, connectivity, education, regulation — but they are also the actors most starved of resources and most distracted by competing crises. International institutions like the World Bank can fund infrastructure and broker technology transfer, and have real convening power, but cannot manufacture the domestic political will that sustains a decade-long buildout. Global North AI companies control the models and the compute, and could enable enormous leapfrogging through open licensing and subsidized access — but their incentives point toward paying customers, not toward the world's poorest markets. The most efficient interventions are reasonably clear: subsidized or pooled compute access, genuinely open model licensing, targeted digital-literacy investment, and international transfer of the delivery-infrastructure "recipes" India and others have proven. What is scarce is not the list. It is the alignment of the actors who could deliver each item.

The deepest governance point concerns whose problems get solved. The failure mode of AI-for-development is importing a Northern application into a Southern context and expecting it to fit — an English-language tutor for a Hausa-speaking classroom, a credit model trained on American spending for a cash economy. The alternative is demand-driven development, where local institutions define the problem and shape the tool, and the model is trained on local data by people who understand the context. That is not a matter of goodwill; it is a matter of arrangement — funding channels, data governance, and procurement rules that put local actors in the driver's seat rather than in the passenger seat of someone else's roadmap. Leapfrogging that merely replicates Northern products in Southern markets is not development. It is a new dependency wearing the costume of progress.

Summary

1. Small AI is a design discipline, not a downgrade. It means building AI to run in constrained environments — offline, on cheap phones, over weak signals, for non-expert users — and doing so makes the tools more valuable in the Global South, not less, because capability you cannot deliver does not count.

2. Adoption is genuinely underway. Middle-income countries generate over 40 percent of ChatGPT traffic, and generative-AI job postings there grew roughly ninefold from 2021 to 2024 — but usage is a precondition for leapfrogging, not proof of it.

3. The mobile-phone precedent both illuminates and misleads. Leapfrogging works when the absence of legacy infrastructure removes switching costs, as with M-Pesa. But AI's data-dependency, compute-intensity, and governance complexity make it harder to import than a phone call, so the analogy should be used with care.

4. India shows what is possible and why it is hard to copy. Its 34,000-plus GPU capacity, $1.25 billion IndiaAI Mission, and Aadhaar/UPI delivery layer serve hundreds of millions — but rested on public investment, institutional depth, and population scale most countries cannot replicate.

5. The evidence is uneven. Agricultural advisories have solid randomized-trial support; healthcare screening has strong accuracy data but thin outcome data; education tutoring is promising but barely tested at scale; financial inclusion is real but mixed on welfare. The tools can work; how often and how durably they do is less established.

6. Leapfrogging is never automatic. It redirects rather than removes infrastructure needs, and it stalls not on technology but on electricity, connectivity, skills, regulation, and institutional follow-through.

7. Smaller, poorer countries have real alternatives to the Indian model — borrowing open delivery infrastructure, pooling regionally, specializing narrowly, and relying on open models and transfer — that lower the bar from building an ecosystem to deploying proven tools.

8. The window is opening and closing at once. Falling costs and open models widen access even as a runaway frontier concentrates the greatest value where the Global South cannot yet reach — and which trend wins is genuinely unsettled.

Sources

• Strengthening AI Foundations: Emerging Opportunities for Developing Countries — World Bank
• Artificial intelligence for low income countries — Nature (Humanities and Social Sciences Communications)
• Artificial intelligence for development (AI4D): A contested notion — SAGE Journals
• Will Data & AI Cripple or Leapfrog Developing Nations' Growth? — Atlantic Council
• Artificial Intelligence in the Global South — Network Readiness Index

Last updated: 2026-08-18