The lottery of the century does not exist: AI will not sort the world by luck, but by exam

THE QUESTION OF THE CENTURY. For the past fifty years the world converged: poor countries grew, on average, faster than rich ones and the gap narrowed. The UN Development Programme (UNDP) now warns that AI could break that trend and open “a next great divergence”. The World Bank argues the exact opposite: its flagship 2026 report describes AI as “a lifeline” that developing countries “should grasp before it slips away”. Two serious institutions, two opposite verdicts on the same technology. Our thesis: both are right, because they are measuring different things. AI diverges on capital — who owns the data centers and collects the rents — and can converge on access — who uses the intelligence, which already rents for cents. The outcome is not a lottery you win or lose. It is an exam: countries that deliver electricity, connectivity and skills pass; countries that wait and see fail.

THE CASE FOR CONVERGENCE. The optimistic argument rests on simple economics: AI makes expertise cheap precisely where it is scarcest. The best evidence came from an NBER study co-authored by Erik Brynjolfsson (Stanford) covering more than 5,000 customer-support agents, mostly in the Philippines: an AI assistant raised average productivity by 14%, but by 34% among novices, and barely at all among veterans. With AI, an agent with two months on the job performed like a six-month agent without it. In short: AI levels workers upward — and translated to countries, that favors those starting from behind. In healthcare the potential effect is larger: the WHO projects a shortage of nearly 6 million health workers in Africa by 2030, and in Kenya one physician can be responsible for more than 10,000 patients. Programs like Babyl in Rwanda signed up over 30% of adults for digital consultations — though its parent company, Babylon Health, went bankrupt in 2023: the technology arrived before the business model did. The precedent that encourages optimists is M-Pesa, Kenya’s mobile money: over 70 million users gained financial services by leapfrogging the bank branches that never came. The World Bank puts a number on the hope: AI could let developing countries achieve in a decade what once took a century, and only 4.5% of jobs in low- and middle-income countries are at risk of automation, versus 14.2% in rich ones. Even Dani Rodrik, the economist who documented “premature deindustrialization”, sees AI-boosted services as a possible new ladder.

THE CASE FOR DIVERGENCE. The pessimistic argument also has hard numbers, starting with the IMF: 40% of global employment is exposed to AI — 60% in advanced economies, 40% in emerging ones, 26% in low-income countries — but the paradox is that rich countries, though more exposed, are better positioned to capture the gains, and the IMF itself warns the cross-country gap could widen. The second pillar is compute, the processing capacity needed to train and serve models. According to Epoch AI, which catalogs more than 500 AI supercomputers, the US hosts roughly 75% of the world’s GPU-cluster performance and China about 15%; the entire rest of the planet shares the remaining 10%. Africa, home to about 19% of the world’s population, hosts around 0.6% of global data-center capacity. The third pillar is power: UNCTAD estimates AI will be a $4.8 trillion market by 2033 — the size of Germany’s economy — that 100 companies, almost all American or Chinese, account for 40% of private AI R&D investment, and that 118 countries, mostly in the Global South, are absent from the tables where AI is governed. Usage is already diverging: per a 2026 Microsoft report, 27.5% of adults in developed countries used generative AI in the first quarter, versus 15.4% in the developing world — a gap that widened by 1.5 points in nine months.

THE SAWN-OFF LADDER. Here lies what we consider the most concrete and least discussed risk. The classic development route was to sell cheap labor to the world — textiles, manufacturing, call centers — and climb that ladder into the middle class. AI and robotics automate precisely that first rung. Bangladesh is the warning: its garment exports grew from $12.5 billion to nearly $40 billion in a decade, yet the sector lost 1.4 million jobs between 2013 and 2023; a local study puts the workforce decline in automated production processes at 30.6%, and 90% of the blow fell on women. More output, fewer workers: the ladder still stands, but its bottom rungs have been sawn off. The Philippines, call-center capital of the world, is the second front: so far reality contradicts the panic — the sector added 60,000 jobs in 2025, reaching 1.68 million, up 4% — but advisory firm Avasant estimates AI could eliminate up to 300,000 positions within five years, against roughly 100,000 new ones in algorithm training and data curation. Back in 2016 the ILO already warned that some Asian countries could lose more than 80% of their garment jobs to automation. The trap is not that AI destroys a poor country’s jobs overnight: it is that it disables the route others used to stop being poor.

OUR READING. Three ideas organize this debate. First: do not confuse producing AI with using it. Frontier clusters — the data centers capable of training the best models — will remain in very few hands, just as phone manufacturing concentrated in Asia. But Kenya never built a single handset and still got M-Pesa: value is also captured by users. Second: the real asymmetry is rents versus wages. AI’s capital rents will flow to shareholders in the US and China; wages, however, can rise in any country where AI makes doctors, teachers or civil servants more productive. Here, open-weight models — those whose parameters can be downloaded and run on your own servers, like GLM, Qwen or DeepSeek — are the best news for the Global South: they slash the entry price and grant sovereignty without a frontier cluster. Third: the real bottlenecks are boring. It is not the model, it is the plug: per the International Energy Agency, some 600 million Africans still lack electricity, and fewer than a third of developing countries even have a national AI strategy, according to UNCTAD. That is why we reject the lottery metaphor: divergence is indeed the default path, but it is not destiny. It is what happens if you skip the exam.

THE IMPLICATIONS. For developing countries, the World Bank’s prescription strikes us as correct and urgent: adopt existing tools now, adapt them to local languages and data, and only later aspire to build your own; and invest in the basics — electricity, connectivity, skills — the way past generations invested in literacy. Negotiating as blocs also matters: initiatives like “OpenAI for Countries” or Abu Dhabi’s Stargate (200 MW planned for 2026) offer infrastructure in exchange for dependency, and a small country bargains better in company; whoever controls the plumbing captures the rent. For rich countries, closing the gap is self-interest: a world where AI enriches only twenty nations is a more unstable world with fewer customers. And for the long run, our editorial line stands: in the short term the problems are real — concentrated rents, sawn-off ladders, a closing window — but in the long term this is the first general-purpose technology whose marginal copy costs almost nothing. The first Industrial Revolution created the Great Divergence because machines could not travel. Artificial intelligence travels through fiber optics. Whether it reaches everyone will be decided not by luck, but by the homework each country does this decade.