The Global South Is Building AI on Someone Else's Infrastructure

Most countries deploy artificial intelligence without owning any part of it. The models, the chips and the data centres belong to companies headquartered in a handful of wealthy states, and the practical consequences of that arrangement are substantial.

What dependency looks like:

A company in Nairobi, Lagos, Dhaka, Lima or Manila building an AI product typically uses a foreign model accessed through an API, running on foreign cloud infrastructure, billed in dollars.

Each layer creates exposure. Pricing changes are set elsewhere. Service availability depends on decisions made in another jurisdiction. Terms of service can change. And billing in hard currency is punishing where local currencies depreciate, which describes much of the world.

Foreign exchange is the constraint founders in emerging markets mention most often, ahead of technical challenges. In Nigeria, Argentina, Egypt, Pakistan and Sri Lanka, currency crises have directly limited what companies could afford to run.

Latency and data residency:

Physical distance from data centres adds latency, degrading interactive applications. A user in East Africa querying a model hosted in Europe or North America experiences delay that a European user does not.

Data protection laws increasingly require that certain data stay within national borders, which conflicts directly with reliance on foreign infrastructure and forces either local hosting, which may not exist, or legal workarounds.

The annotation asymmetry:

A visible manifestation is data work. Thousands of people in Kenya, the Philippines, India, Venezuela and elsewhere label images, transcribe audio, evaluate model outputs and moderate content that trains and improves systems owned elsewhere.

The work provides employment and foreign exchange, and it has drawn sustained criticism over pay, working conditions and the psychological effects of content moderation.

The structural point is simple: labour from lower-income countries improves systems whose value accrues to firms in higher-income countries, and the wage gap between those two facts is the business model.

What countries are doing about it:

Several responses are visible. Local data centre investment is growing across Africa, Latin America and Southeast Asia, often funded by international operators rather than domestic capital, which addresses latency and residency without changing ownership.

Sovereign infrastructure projects, such as Senegal's national data centre or India's subsidised compute access, attempt domestic control.

Open-weight models are the most consequential development. A downloadable model can be run on modest local infrastructure, adapted to local languages and deployed without ongoing dependence on a foreign API. This does not remove hardware dependence, but it removes one layer of it.

Regional cooperation, pooling compute or coordinating regulation across neighbouring markets, has been discussed more than implemented.

What is realistic:

Frontier model development is not a realistic goal for most countries; the capital required is prohibitive.

What is achievable: local language datasets and models, domestic hosting for sensitive data, negotiating better terms collectively rather than individually, taxing and regulating foreign providers effectively, and building the engineering capability to adapt and deploy rather than only consume.

The outlook:

The dependency is real and unlikely to disappear, because it reflects genuine concentration in capital, energy and manufacturing.

The question for most countries is not independence but terms: whether they participate as customers with leverage and domestic capability, or as markets and labour pools with neither. That difference is decided by policy choices being made now.