The Countries Being Left Behind by AI (And Why That Should Terrify Everyone)

In 2023, a small startup in Nairobi built a credit scoring AI using data from mobile money transactions. The model was clever, genuinely useful, and funded by an accelerator program backed by a major US venture fund. When the fund examined the model’s intellectual property, it decided to bring the core technology to the United States, where it could be scaled and monetized more effectively. The Kenyan engineers were hired, at US salaries, to work remotely on a product that was now headquartered in San Francisco. The data about Kenyan borrowers was owned by a Delaware company. The productivity of that Kenyan engineering team now accrued to US investors.

This story is not exceptional. It is, increasingly, the template for how AI development interacts with the developing world: local talent and local data are extracted into global platforms owned and controlled elsewhere, and the economic surplus follows the ownership structure rather than the geography of production.

The global distribution of AI capacity is becoming one of the most consequential geopolitical facts of the current decade, and it is receiving a fraction of the analytical attention it deserves.

The concentration is stark. By nearly every metric — number of foundation model developers, AI research publications, AI venture capital investment, compute infrastructure — AI development is concentrated in the United States, China, and, to a lesser extent, the United Kingdom and European Union. Countries in these blocs have access to frontier AI through their domestic companies or through partnerships with US and Chinese tech firms. The rest of the world — South Asia, Southeast Asia, Latin America, Africa, the Middle East — has AI deployed into it but almost no meaningful role in shaping what that AI is, how it works, or who controls it.

This is not primarily a function of human capital. India produces world-class AI engineers at scale; many of them work at Google, Meta, and OpenAI. Nigeria has a growing tech ecosystem with genuine AI capability. Brazil has universities doing serious AI research. The gap is not talent. The gap is capital and compute. Training a frontier large language model requires hundreds of millions to billions of dollars in GPU compute. The world’s GPU capacity is heavily concentrated in US and Chinese data centers. The capital to fund AI ventures at the scale required to compete globally is concentrated in US venture markets and Chinese state investment.

The result is that countries outside the three main blocs face a structural choice: accept AI services built by and for US or Chinese contexts, or develop local capabilities at a fraction of the frontier scale. Neither option is attractive. Accepting US and Chinese AI services means accepting the values, the data practices, and the geopolitical orientations embedded in those services. Building local capabilities at small scale means building models that are less capable than frontier alternatives, which makes them hard to adopt in competition with free or low-cost frontier alternatives.

The colonial parallel is not a metaphor chosen for rhetorical effect. It has structural substance. Colonial economic extraction depended on a specific mechanism: raw materials were extracted from colonial territories, processed and manufactured into finished goods in metropolitan countries, and then sold back to colonial territories at a markup. The territories provided labor and resources; the metropolis captured the value-added stage of production.

The AI economy has a similar structure. Data — the raw material of AI — is generated everywhere. User behavior, medical records, transaction histories, communications, infrastructure sensor data: these are generated in proportion to population and economic activity, which is globally distributed. But the processing of data into AI models — the value-added stage — happens in a handful of metropolitan centers. The finished products (AI services, AI-powered software) are then deployed globally, often generating revenue that flows back to the metropolitan centers.

The mechanism by which this happens is not primarily coercion. It is economic gravity. US and Chinese AI platforms offer services of sufficient quality at sufficiently low cost that local alternatives are hard to sustain. A hospital in Ghana that adopts Google’s AI diagnostic tools is making a rational decision: the tools are good, they are affordable or free, and building comparable local tools would require capital and talent that is not available. But the data those tools generate, the models trained on that data, and the profits from that deployment accrue to Google’s shareholders in California and institutional investors in New York.

The sovereignty question is perhaps more important than the economic one, though they are related. Sovereignty, in any meaningful modern sense, requires the capacity to make consequential decisions about one’s own governance, economy, and society. When critical infrastructure depends on AI systems controlled by foreign companies subject to foreign law, sovereignty is constrained in important ways. The country that cannot independently build or operate the AI systems powering its credit markets, its public health surveillance, its education system, or its national security apparatus is dependent on foreign actors whose interests may not align with its own.

This is not a hypothetical concern. When the US government imposed export controls on advanced AI chips in 2022 and expanded them in subsequent years, the stated goal was to prevent China from building AI capabilities that could be used for military advantage. The effect was also to prevent every other country from accessing the most capable AI hardware — and therefore from developing advanced AI capabilities of their own. The chip export controls are a geopolitical weapon deployed in a US-China competition, but the collateral damage is the AI sovereignty of the entire rest of the world.

The European Union has recognized this dependency problem and has invested significantly in building domestic AI capacity: the OECD AI Initiative, national AI strategies in France and Germany, the AI Act as a regulatory framework that would apply regardless of where AI is developed. But even the EU’s ambitions are concentrated in a few countries, and the rest of the world — the populations that represent the majority of humanity — has no meaningful equivalent.

What happens to a country’s economy when AI-powered services from elsewhere automate its service sector? This question is not abstract. The service sector is the primary route through which developing countries have historically escaped agricultural poverty: first manufacturing, then services, then knowledge work, with each transition providing employment and building human capital for the next. Countries like India, the Philippines, and Kenya have built significant economic positions in business process outsourcing, software development, and digital services precisely because the global economy created opportunities for labor-cost arbitrage in cognitive work.

AI threatens to close that window. If the cognitive tasks currently performed by white-collar workers in lower-cost countries can be automated at near-zero marginal cost by AI systems owned by US or Chinese companies, the economic migration path from agriculture to services to knowledge work is disrupted. The jobs that would have trained the next generation of knowledge workers don’t exist. The foreign exchange earnings that would have funded domestic investment don’t materialize.

This is not an argument for preventing AI adoption in developing countries. The productivity gains from AI are real and would benefit these economies as well. The argument is about who captures those gains and under what conditions. An Indian data entry worker replaced by an AI tool controlled by a US company does not benefit from that productivity gain. A Nigerian startup that can afford to use GPT-5 but cannot afford to train its own competitive model is building on sand — its core capability can be switched off by a policy decision in Washington.

What would a more equitable global AI architecture look like? There is no simple answer, but a few structural changes would matter. Compute access — either through international institutions that provide compute resources to developing country researchers, or through open-source model training efforts that allow local fine-tuning without requiring frontier training runs — would partially address the infrastructure gap. Data sovereignty frameworks that ensure countries control the data generated by their citizens and can participate in the value created from it would partially address the extraction problem. International standards-setting bodies that include representatives from developing countries in designing AI governance would partially address the legitimacy problem.

None of these changes will happen without political pressure from developing countries themselves, and from the minority of developed-country institutions that recognize global AI concentration as a stability risk rather than just an equity concern. The historical parallel is the fight over international intellectual property rules in the 1990s and 2000s, where developing countries eventually won some meaningful concessions but only after sustained coalition building.

The terrifying aspect of the AI divide is not just the inequality it creates, which would be bad enough. It is the stability risk. A world where the most powerful technology in human history is controlled by two rival superpowers, where every other country is dependent on one or the other for its AI infrastructure, where the economic development pathways of billions of people are being closed by automation they don’t control — that world has dynamics that are historically associated with conflict and instability rather than with the cooperative management of shared challenges. The countries being left behind by AI are not a humanitarian afterthought. They are the majority of the world, and how they relate to a technology they don’t control will shape the next fifty years of international order.

There is a version of this story in which the concentration of AI accelerates a kind of technological colonialism that the post-WWII international order was supposed to have foreclosed. The colonial era was characterized by the extraction of raw materials from less powerful territories by more powerful ones, with the value-added manufacturing happening in the metropolis. The global AI economy risks reproducing this structure with data as the raw material, model training as the manufacturing stage, and API access as the finished good — sold back at whatever price the provider chooses to charge. The countries that provide the human data used to train frontier models may eventually conclude that they want data sovereignty arrangements that give them more control over this dynamic. When that political pressure builds, and it will, it will reshape the terms of global AI deployment in ways that companies currently operating without thinking about these dynamics are not prepared for.