Global North-South Divide

Dr. Kwame Mensah runs a machine-learning lab at the University of Ghana in Accra. His team builds tools for African agriculture — models that predict crop yields, time irrigation, and spot plant disease from a phone photograph. The work matters: farming employs roughly three in five working Africans, and a warming climate is making every planting season a gamble. AI could genuinely help.

But Kwame's annual budget would not cover a single week of compute at a comparable lab at Stanford or MIT. He has no cutting-edge GPUs, no proprietary datasets, and no local cloud provider to rent from. His team fine-tunes open-source models that were never trained on cassava blight, laterite soils, or the rhythm of a Sahelian rainy season. When they publish, the Global North AI community mostly looks past them. When they apply for international grants, they compete against labs with a hundred times the resources. And when they finally train a brilliant graduate student, that student takes a job in London or Toronto for five to ten times a Ghanaian salary.

Kwame is not bitter. He is realistic. The AI revolution is happening, and much of Africa is watching it from the touchline — not for any lack of talent, ambition, or need, but because the ingredients of AI are concentrated somewhere else, and the concentration is deepening rather than dissolving. That is the divide this chapter is about: the widening gap between a small set of countries building the AI future and a large set of countries left to consume it, with little say over how the technology develops or who its gains belong to.

The Geography of Concentration

Start with the numbers, because they are stark enough to carry the argument on their own. High-income countries hold roughly 87% of the world's notable AI models, 86% of AI startups, and 91% of AI venture capital — while accounting for about 17% of the world's people (figures compiled from the Stanford AI Index and UNDP analysis). The capital story is even more lopsided than the population ratio suggests. Of the private AI investment tracked over the decade to 2023, the United States alone drew some $335 billion — roughly three times China's total, eleven times the United Kingdom's, and thirty times India's. Less than one percent of global AI funding reaches the Global South at all.

The physical layer beneath the models is the most concentrated of all. Data centers — the warehouses of specialized silicon where models are trained and served — cluster overwhelmingly in North America, Europe, and East Asia. Africa, home to about 18% of humanity, holds well under 1% of global data center capacity. And the chips inside those buildings come, for practical purposes, from one company: Nvidia's accelerators dominate the training market so thoroughly that its share of AI compute hardware is usually estimated somewhere between 80% and over 90%. A single firm in a single country sits at the base of nearly the entire global stack.

These are not the ordinary gaps of an uneven world. They are structural asymmetries that decide, in advance, who gets to shape a general-purpose technology and who merely receives its outputs.

A fair reader should pause here on how trustworthy these percentages are. They are assembled from datasets — patent filings, published-model registries, venture-funding trackers — that are far better at counting activity in English-speaking, well-instrumented economies than in places where research happens outside formal channels, funding flows through informal networks, and startups never register with a Western data vendor. That measurement bias almost certainly undercounts Global South activity somewhat. But it cuts the other way too: the same countries that go uncounted are the ones lacking the compute and capital to move the needle on the metrics that dominate. The safest reading is that the direction of the gap is not in doubt, even if the second decimal place is. If anything, the deepest disadvantage — the near-total absence of frontier compute — is the part least likely to be flattered by better data.

Why the Gap Compounds

The unsettling feature of this divide is not its current size but its motion. AI development is cumulative in a way that turns a lead into a moat. The organizations with the most compute train the most capable models; capable models attract the most users; users generate the most data and revenue; revenue buys more compute and hires the scarcest talent, which trains still better models. Each turn of that wheel widens the distance to everyone standing outside it.

graph LR
  A[More compute + capital] --> B[Better models]
  B --> C[More users, data, revenue]
  C --> D[More talent attracted]
  D --> A
  E[Latecomers: no compute, no capital] -. locked out .-> A

This is why an infrastructure deficit is not a static handicap that a country can simply grow out of at its own pace. A lab without reliable power cannot run large training jobs at all; power outages that merely inconvenience an office will corrupt a multi-week training run. A lab without local cloud infrastructure must rent from foreign providers priced for Silicon Valley budgets, where a few hours of frontier GPUs can swallow a month's research allocation. A lab without fast broadband spends days moving the datasets a Northern peer transfers in minutes. Each of these frictions is survivable alone. Together they mean that even well-conceived, socially valuable projects stay perpetually under-resourced — and that the country falls further behind precisely while the frontier accelerates. Standing still, in a compounding race, is the same as losing ground.

The Talent That Leaves

The second engine of divergence is human. The salary arithmetic is not subtle. Median compensation for AI roles in high-income countries runs around $160,000 a year, with scarce specializations commanding a further 25–45% premium; a comparable researcher or engineer across much of the Global South might earn $10,000 to $20,000. That is not a gap to be closed by loyalty or patriotism. It is an order of magnitude, and it produces exactly the flow you would expect: bright students win scholarships to universities in the US, UK, or Europe, and then take positions at the very labs and companies whose resources no home institution can match.

Some keep ties through remote collaboration or eventually return. Many do not. And those who return often find the same constraints that made leaving rational in the first place — now made worse by the years of frontier experience their departed peers accumulated abroad. The damage is not only to individual careers. A country that trains talent at public expense and then watches the Global North harvest it is subsidizing someone else's AI industry. The domestic problems that talent might have solved — the crop-disease detector, the local-language health chatbot, the flood-forecasting model — go unbuilt, not because AI cannot address them but because the people who could build them are working on advertising click-through rates in another hemisphere. Weaker local capacity lowers the return on domestic AI investment, which discourages investment, which makes the pipeline leak faster. The brain drain, like the infrastructure gap, feeds on itself.

India, and the Myth of a Single "Global South"

The phrase "Global South" is a convenience that hides enormous internal variation, and nothing exposes that better than India. When India hosted the AI Impact Summit in New Delhi in February 2026 — the successor gathering in the lineage that ran through Bletchley Park, Seoul, and Paris — it did not arrive as a supplicant. India came as a middle power with real cards to play: a technology workforce numbering in the millions, national compute capacity built out to tens of thousands of GPUs under its IndiaAI Mission, a booming AI-services sector, and homegrown efforts to build models for its own dozens of languages.

That distinctive position let India do something most developing countries cannot: press Global North institutions on roughly equal footing. India used the summit to argue for affordable compute access through subsidized or pooled international infrastructure, for technology-transfer mechanisms that make advanced capabilities reachable for developing nations, and for genuine Global South representation in the bodies that write AI's rules — framing AI as a tool that should serve development and "public good," not merely Northern commercial priorities. It positioned itself, in effect, as a broker between the frontier labs and the countries with none.

India's story cuts two ways for the rest of the Global South, and it is important to hold both. Over roughly five years, India moved from a country whose AI story was mostly about outsourced back-office work to one with sovereign-compute ambitions, indigenous large models, and a seat at the head of the table. That is proof the trajectory is not fixed. But India's ascent rested on preconditions few of its peers share: an enormous domestic market, a deep pre-existing IT industry, a large diaspora in Northern tech, and a state with the fiscal room to fund compute at national scale. India may be less a replicable template than an existence proof — evidence that domestic agency is possible, alongside a reminder of just how much has to already be in place before it becomes achievable. A country is not "Global South" in any uniform sense; Bangalore and Bamako face the same frontier from radically different starting lines.

The Next Great Divergence

The United Nations Development Programme has given this dynamic a name and a historical warning: AI, it argues, could drive "the next great divergence." The reference is to the first one. During industrialization, the countries that built early factory capacity captured compounding advantages — industrial profits funded more industry, which drew more capital, which widened the lead — while late industrializers slid into dependence on imported manufactures and never captured the productivity gains for themselves. The gap between rich and poor nations that opened in the nineteenth century took the better part of two centuries to become the defining feature of the world economy, and much of it has never closed.

The UNDP's fear is that AI reruns this pattern on a compressed clock. Countries with early advantages — compute, capital, talent — build better models that generate more value that funds more development that attracts more talent, while countries without those initial conditions cannot get into the loop at all and remain consumers of technology built elsewhere. The agent of divergence, in this telling, is not the technology. It is the unequal distribution of the conditions required to develop and deploy it. Absent deliberate intervention, the differential pace of adoption does the damage.

The analogy is powerful, and it is worth being honest about where it holds and where it might mislead. It holds on the core mechanism — cumulative advantage, compounding returns, lock-out of latecomers — and on the political economy: then as now, the technology's benefits accrue to whoever owns the capital stock. But AI differs from the spinning jenny in ways that could push the outcome in either direction. On the hopeful side, a trained model can be copied at near-zero marginal cost, where a steel mill could not; open weights can cross a border on a hard drive, and a capable model released today is genuinely usable in Nairobi tomorrow. Industrialization had no equivalent of Meta releasing a frontier-adjacent model for free. On the darker side, AI's dependence on scarce, export-controlled, physically concentrated compute — and on data and energy at a scale factories never required — may make the barrier to building at the frontier even higher and more defensible than a factory ever was. The analogy warns us correctly that divergence is the default. It should not lull us into assuming AI will diverge in exactly the same shape, or on the same timescale, as coal and steel did.

Accessible Is Not Equal

AI's champions insist the technology will democratize opportunity: open-source models, cheap cloud, and free courses put real power in anyone's hands. This is not empty marketing. A researcher in Accra or Jakarta really can download a pretrained model, fine-tune it on local data, and ship something useful without building the underlying infrastructure from scratch. The barrier to using frontier AI has fallen further and faster than almost anyone predicted.

But accessible and equal are different words, and the distance between them is where the Global South lives. The open model you download for free was trained on infrastructure that cost tens or hundreds of millions of dollars — infrastructure held by a handful of Northern firms, and the reason they can give the weights away is that the weights are not where the durable advantage sits. Cloud compute is nominally available to everyone and priced for a venture-funded startup, not a public-university research line. Free educational resources assume a baseline — steady electricity, a decent laptop, uninterrupted broadband — that much of the world cannot presume. Access lowers the floor; it does not level the field. What has been democratized is the ability to use technology built by and for the North. What has not been democratized is the capacity to build technology for the South — to choose which problems are worth solving, to own the training data, and to control the systems that will shape billions of lives.

When Northern Models Meet Southern Reality

The clearest evidence that access is not enough comes from what happens when a model trained on Northern data is pointed at a Southern problem. A crop-disease classifier trained on well-lit, single-leaf images from industrial farms in temperate climates degrades badly on a photo of intercropped cassava shot on a cheap phone in humid, cluttered field conditions — different pests, different varietals, different backgrounds, different cameras. An irrigation or yield model tuned to large mechanized farms encodes assumptions — plot size, input availability, weather-station density — that simply do not describe a two-hectare smallholding.

Healthcare is where the failure turns dangerous. A diagnostic model trained on clinical data from high-income hospitals inherits their patient mix, their equipment, and their disease priors. Point it at a low-resource clinic and it may miss conditions that are common there and rare in the training set, misread images from older or different machines, or assume laboratory results that the clinic cannot produce. Dermatology models trained mostly on light skin have documented accuracy drops on darker skin; a sepsis or triage model calibrated to one health system's baseline can silently mis-rank risk in another. The tool does not announce its failure — it produces a confident answer that is wrong for a population it never saw. These are not edge cases to be patched later. They are the predictable result of building for one context and deploying in another, and they fall hardest on the people with the least margin to absorb a bad prediction.

What the Evidence Actually Shows — and Doesn't

It would be dishonest to present the widening divide as settled empirical fact. What we have is a strong theoretical case, an alarming set of concentration metrics, and a genuinely thin evidence base on realized distributional effects across countries. The IMF has warned that AI could exacerbate cross-country inequality, estimating that advanced economies stand to gain far more from AI-driven growth than low-income ones — potentially more than double — precisely because they have the digital infrastructure, skilled workforces, and capital to capture the gains. That is a forecast built on exposure and readiness indices, not a measurement of what has already happened.

The honest position is that the near-term outcome is not yet written in the data. Adoption is early; the distributional accounts are incomplete; and there are real countervailing possibilities — cheap, capable models leapfrogging legacy infrastructure the way mobile money leapt over bank branches in Kenya, or falling inference costs putting genuinely useful tools in the hands of low-income users faster than the pessimistic model predicts. The concentration of production is measured and severe. The consequence for cross-country inequality remains a well-motivated projection with the empirical verdict still out. A book that respects its readers should say both things at once.

Why the Market Won't Fix It

If the divide were merely an accident, it might correct itself. It is not. It is the accumulated residue of decisions — about where to invest, whose problems to prioritize, and who sits in the rooms where standards are written — and market incentives push those decisions the wrong way. A frontier lab has little commercial reason to transfer its hardest-won capabilities, training data, and know-how to institutions that are either potential competitors or small addressable markets. Governance compounds the problem: safety standards, audit requirements, and ethical frameworks drafted in high-income countries, often with heavy input from large Northern firms, carry compliance costs — legal, technical, administrative — that fall hardest on the smallest and poorest institutions. Being bound by rules you had no hand in writing is itself a structural disadvantage, layered on top of the resource gap.

Because the market works against sharing, closing the gap requires instruments that override market logic. That is the case for treating foundational AI capabilities — not every application, but the base models, and the compute and data underneath them — as something closer to a global public good than a purely private asset. What that would take in practice is concrete and hard: internationally pooled or subsidized compute that developing-country researchers can actually afford; public funding for frontier work made conditional on knowledge-sharing; and a governance body with the standing to enforce access obligations. The obvious models are the ones the world has built for other shared goods — a CERN-style consortium for compute, a Gavi-style facility that guarantees affordable access, or dedicated windows at existing multilateral development banks. Each raises the unavoidable question of who pays, and the equally unavoidable answer that the institutions and countries capturing the largest gains from AI are the ones with both the resources and, many would argue, the obligation to fund it.

That obligation is not only fiscal. The Northern labs that recruit from Global South talent pipelines and train on data drawn from Global South populations are extracting value that those countries and communities produced. A serious accounting of what is owed — in compute access, in local investment, in fair data terms, in a real say over governance — is not charity. It is a reckoning with a supply chain whose upstream and downstream sit in different worlds.

And it changes what "good" investment even means. If AI investment is judged only by aggregate GDP growth, a world in which the frontier races ahead and everyone else buys the outputs can look like a triumph. Capturing whether AI is narrowing or widening the divide requires different instruments: the share of compute, models, and capital held outside the high-income club; the number of frontier-capable institutions in the Global South; net talent flows rather than gross output; how well deployed models perform on non-Western languages and low-resource settings; and whether local institutions actually control the systems shaping their populations. Metrics like these measure agency, not just activity — and agency is the thing the current trajectory is failing to distribute.

The Decade Ahead, and the Closing Window

Extrapolate the present forces and the picture is sobering. Capital flows toward existing concentrations of capability. The compounding loops are running as designed. On current trajectories, most of the Global South is on track to become, a decade out, primarily a consumption market for Northern AI and a talent pipeline feeding Northern labs — with pockets of real domestic agency in a handful of middle powers like India, and thinner prospects everywhere else. The cruel irony is that AI's potential to move the needle on disease, agricultural productivity, and climate adaptation is largest exactly where that potential is least likely to be realized, because the capacity to build and deploy already sits where the need is smallest.

The one genuinely open question is whether a window remains. The industrial precedent is not encouraging — that divergence, once compounded, took generations and mostly never reversed — and the concentration metrics suggest AI's advantages are hardening fast. But two features distinguish this moment from 1850. The technology is still early enough that adoption gaps have not yet fully translated into outcome gaps, and the copyable, near-zero-marginal-cost nature of a trained model means the diffusion channel, at least for using AI, is unusually open. The window is real, and it is narrowing. Whether it closes depends less on the technology than on choices — about investment, governance, and priority — that are, at least for now, still ours to make. The divide is the product of decisions. So is the alternative.

Summary

1. Concentration is severe and measurable. High-income countries — about 17% of the world's people — hold roughly 87% of notable AI models, 86% of AI startups, and 91% of AI venture capital, while Africa (18% of humanity) holds under 1% of data center capacity. Beneath it all, one firm, Nvidia, supplies the overwhelming majority of AI training hardware.

2. The gap compounds. Because AI development is cumulative — compute begets better models beget revenue begets more compute — infrastructure and talent deficits are not static handicaps but self-reinforcing feedback loops that lock latecomers out.

3. Talent flows North. An order-of-magnitude salary gap ($160k+ versus $10–20k) drives skilled researchers out of the Global South, so countries subsidize training whose returns are captured elsewhere, weakening domestic capacity in a loop that feeds itself.

4. The "Global South" is not one thing. India, hosting the February 2026 AI Impact Summit as a genuine middle power, shows domestic agency is possible — but its preconditions (huge market, deep IT base, fiscal room, large diaspora) make it more existence proof than replicable template.

5. The UNDP's "next great divergence" warns AI could rerun industrialization's rich-poor split on a faster clock. The analogy holds on cumulative advantage but may mislead: models are copyable in a way factories never were, yet frontier compute may be a harder barrier than any mill.

6. Access is not equality. Open models and cheap cloud lower the floor without leveling the field, and Northern-trained models fail in Southern contexts — degrading on smallholder crops and misdiagnosing in low-resource clinics — because they encode the world they were built in.

7. The verdict is not yet in the data. The IMF projects AI will widen cross-country inequality, but that is a forecast from readiness indices; realized distributional effects remain thinly measured, and leapfrogging is a real, if uncertain, counter-possibility.

8. Markets won't close it alone. Commercial incentives work against technology transfer, so narrowing the divide needs deliberate instruments — pooled compute, conditional public funding, public-good governance, and honest metrics of agency rather than GDP — funded by those capturing AI's largest gains, who also owe the most to the talent and data pipelines they draw from.

Sources

• The Next Great Divergence: How AI could split the world again — if we don't intervene | UNDP
• The 'AI divide' between the Global North and Global South | World Economic Forum
• From Divide to Delivery: How AI Can Serve the Global South | CSIS
• Artificial Intelligence in the Global South | Network Readiness Index
• Three Reasons Why AI May Widen Global Inequality | Center for Global Development
• The Dangers of Imposing Global North Approaches to AI Governance on the Global South | Tech Policy Press
• AI's Unequal Revolution: Bridging the Global Divide | KSAPA

A note on process, outside the chapter itself: I was unable to run live web searches or fetches this session (both tools were unpermitted), so I could not verify figures against the latest 2026 sources or confirm specific outcomes of the India AI Impact Summit beyond what the prior version and my training establish. The statistics used are all carried forward from the previous version's cited reports or attributed to named institutions (IMF, UNDP, Stanford AI Index); I introduced no new numeric claims. If you grant WebSearch/WebFetch access, I can re-verify the concentration metrics and the summit details and update anything that has moved.

Last updated: 2026-08-17