Redistributed

AI could narrow global inequalities. It could bring medical diagnostics to clinics without doctors, agricultural advice to farmers without agronomists, educational resources to students without schools. The technology is scalable. The marginal cost of deploying a trained model approaches zero.

In many domains, AI is widening the gap instead.

The same technology that promises universal benefit extracts from the Global South and concentrates returns in the North. Training data comes from everywhere; the models stay in San Francisco. Click workers in Nairobi label data for systems they'll never use. Researchers in Lagos publish papers that improve products built in Palo Alto.

This is not incidental. It's structural.

The Geography of AI Capability

AI capability clusters in a handful of places. Map the locations of compute infrastructure, research talent, and training data pipelines, and you'll find the same names: San Francisco, Seattle, New York, London, Beijing, and a few others.

Compute concentration:

As of 2026, roughly 80% of AI training compute exists in three countries: the United States, China, and the United Kingdom. The hyperscale data centers required for frontier model training are almost exclusively located in wealthy nations with reliable power grids, favorable regulations, and access to capital.

This isn't just about money. It's about infrastructure that takes decades to build: electricity generation, cooling systems, fiber optic networks, skilled maintenance workforces. Nigeria has a larger population than the US. It has a fraction of the data center capacity.

Research concentration:

The top AI research institutions — measured by citations, talent, and funding — cluster similarly. Stanford, MIT, Berkeley, Carnegie Mellon, Oxford, Cambridge, Tsinghua, Peking. These institutions feed into the companies that build AI systems. The pipeline is direct and self-reinforcing.

As Kate Crawford documents in Atlas of AI, this concentration isn't natural; it's the result of deliberate investment choices, immigration policies, and institutional development over decades (Crawford, 2021). The geography of AI reflects the geography of power.

Talent flow:

Brain drain is real. A researcher trained in Accra faces a choice: stay and build with limited resources, or move to where the compute is. Many move. This depletes local capacity while enriching already-rich institutions.

According to the OECD, over 60% of AI PhDs trained in Africa work outside the continent (OECD, 2023). The pattern holds for South Asia, Latin America, and much of Southeast Asia. Talent flows upward on the wealth gradient.

Training Data: Whose Language, Whose Knowledge?

Large language models are trained on text scraped from the internet. The internet is not neutral. It overrepresents English, wealthy countries, and Western perspectives.

The language distribution:

Common Crawl, the most widely used web scraping dataset, is roughly 45% English. The next largest languages — German, Russian, French — each account for 5-6%. Languages spoken by hundreds of millions — Hindi, Bengali, Yoruba, Swahili — are barely present (Bender et al., 2021).

This isn't just about translation. It's about what knowledge gets encoded into AI systems. A model trained overwhelmingly on English text learns English concepts, English framings, English assumptions about how the world works.

The fairness mismatch:

Nithya Sambasivan and colleagues studied algorithmic fairness in India and found that Western fairness frameworks simply don't map (Sambasivan et al., 2021). Categories that matter in the US — race as a binary, individual-level discrimination — miss the forms of inequality that matter locally: caste, religion, region, language. AI systems designed with American assumptions fail to address — and may reinforce — Indian inequalities.

This pattern repeats globally. Abeba Birhane's work on "algorithmic colonialism" documents how AI systems built in the Global North impose Northern categories on Southern realities (Birhane, 2020). Classification systems trained on Western data misclassify African faces, African names, African diseases.

The extraction pattern:

When AI companies need data from the Global South, they extract it. When they build products, they serve the North. Ruha Benjamin calls this "the new Jim Code" — technology that encodes inequality while appearing neutral (Benjamin, 2019).

Medical AI trained on scans from Kenyan hospitals improves diagnostics — but the resulting products are priced for American markets. Agricultural AI trained on data from Indian farms optimizes yields — for crops and conditions that match Western industrial agriculture.

The value flows one direction.

Infrastructure Gaps

AI requires infrastructure that most of the world lacks.

Bandwidth:

Running AI models — whether locally or via API — requires fast, reliable internet connections. The International Telecommunication Union reports that as of 2025, only 36% of people in Africa have internet access, compared to 90% in Europe (ITU, 2025). The connections that do exist are slower, more expensive, and less reliable.

A farmer in rural Senegal might theoretically benefit from AI-powered agricultural advice. They're unlikely to have the connectivity to access it.

Electricity:

Data centers are power-hungry. Training a single large language model can consume as much electricity as 100 American homes use in a year. Countries with unreliable power grids cannot host training infrastructure.

But the problem extends beyond training. Running inference — actually using AI systems — also requires power. Intermittent electricity makes always-on AI applications impossible for large portions of the world's population.

Hardware:

The smartphones and computers needed to access AI services remain expensive relative to local incomes. A device that seems cheap by American standards can represent months of income for a worker in Bangladesh or Nigeria.

Even when hardware exists, it's often older, slower, and less capable. AI applications optimized for the latest devices run poorly or not at all on the devices that most of the world actually uses.

Economic Barriers

Even where infrastructure exists, economics create barriers.

API pricing:

Access to frontier AI capabilities means paying API fees — typically priced in US dollars, at rates set for American budgets. A cost-per-token that's trivial for a San Francisco startup is prohibitive for a Lagos one.

OpenAI, Anthropic, and Google price their APIs globally but set prices locally. The result: American companies can experiment freely, while companies in poorer countries must ration access carefully.

Compute costs:

Training competitive models requires GPUs that cost tens of thousands of dollars each, available primarily through cloud providers who charge premium rates. The barriers to entry keep rising. Each generation of frontier model requires more compute than the last.

Researchers in wealthy institutions can access compute through grants, institutional resources, and industry partnerships. Researchers elsewhere cannot.

Business model capture:

The dominant AI business model — free services supported by advertising and data extraction — optimizes for users who generate ad revenue. Users in wealthy countries generate more revenue. So AI products optimize for wealthy-country users.

Features that would benefit lower-income users — offline functionality, low-bandwidth modes, local-language support — don't justify the investment when the revenue comes from elsewhere.

AI Reinforcing Inequality

When AI systems do reach the Global South, they often reinforce existing inequalities rather than challenging them.

Credit and finance:

AI-powered credit scoring has expanded financial access in some cases. But systems trained on historical data encode historical discrimination. Patterns of redlining, colonial-era economic exclusion, and structural inequality become automated predictions.

As Virginia Eubanks documents in Automating Inequality, algorithmic systems tend to concentrate risk on those already vulnerable (Eubanks, 2018). This pattern isn't limited to the US — it applies wherever AI systems make consequential decisions based on historical patterns.

Labor platform dynamics:

AI-powered labor platforms have created work opportunities across the Global South — but often on exploitative terms. Platforms use algorithmic management to extract maximum value from workers while providing minimal protections.

The "ghost work" that Mary Gray and Siddharth Suri describe — the hidden human labor behind AI systems — is disproportionately performed by workers in the Global South (Gray & Suri, 2019). Kenyan workers moderate traumatic content for Facebook. Filipino workers label data for autonomous vehicles. Indian workers train chatbots through RLHF. The work is precarious, poorly paid, and psychologically damaging.

Surveillance and control:

AI-powered surveillance technologies developed in China and the US find ready markets in authoritarian governments worldwide. Facial recognition, social media monitoring, and predictive policing systems flow from wealthy countries to less-wealthy ones — extending state control capabilities without the civil liberties protections that (sometimes) constrain their use in origin countries.

The pattern Crawford identifies — AI as "extractive industry" — applies globally (Crawford, 2021). Extract data, extract labor, return surveillance and control.

Efforts to Redistribute

Against this backdrop, significant efforts are underway to redistribute AI capability. These efforts are real, growing, and worth understanding in detail — because they demonstrate that the current distribution is a choice, not an inevitability.

Masakhane:

Masakhane is a grassroots research community focused on natural language processing for African languages. Founded in 2019, it now includes over 2,000 researchers across the continent working on machine translation, language models, and NLP tools.

The approach is deliberately distributed: researchers work from their home institutions, share resources and datasets, publish collectively. The name means "we build together" in isiZulu.

Masakhane has produced translation systems for languages — Yoruba, Igbo, Swahili, Amharic — that commercial AI largely ignores. More importantly, it has built capacity: researchers who can continue developing AI tools locally, for local needs.

Lelapa AI:

Based in South Africa, Lelapa AI is a commercial venture building AI for African languages and contexts. Unlike Masakhane's research focus, Lelapa aims to build products — but products shaped by local needs rather than Northern assumptions.

Lelapa's Vulavula platform provides speech-to-text and text-to-speech for African languages, trained on data collected ethically from speakers across the continent. The model: AI built in Africa, for Africa, with African ownership of the value created.

Distributed research networks:

Beyond Africa, similar efforts are emerging: AI4Bharat for Indian languages, SEACrowd for Southeast Asian NLP, and numerous university-based initiatives across Latin America.

These networks face resource constraints that Northern institutions don't. But they bring something else: researchers who understand local contexts, local needs, local constraints. AI built by and for the communities it serves.

Open models and compute access:

The rise of open-weight models (Llama, Mistral, Falcon) creates new possibilities. Researchers without access to frontier compute can fine-tune existing models for local applications. The computational barriers remain significant but are lower than building from scratch.

Organizations like EleutherAI and Hugging Face provide infrastructure — model hosting, fine-tuning tools, community support — that partially democratizes access. It's not enough, but it's something.

Policy Approaches

Closing the digital divide in AI requires deliberate policy intervention. Several approaches show promise.

Data sovereignty:

Countries are increasingly asserting control over their citizens' data. India's proposed data protection framework, the African Union's data policy framework, and similar initiatives aim to ensure that data generated locally creates value locally.

The goal: prevent the extractive pattern where data flows out and returns as products priced for foreign markets. Instead, require data to be processed locally, or ensure local benefit from processing elsewhere.

Compute sharing:

Public compute initiatives could extend beyond wealthy countries. The European High-Performance Computing Joint Undertaking provides a model: pooled resources, distributed governance, shared access.

An African or Global South equivalent — a shared compute infrastructure available to researchers across developing countries — could shift the balance. International development funding could support such initiatives, as could compute contributions from wealthy-country institutions.

Capacity building:

Investment in local AI education and research capacity pays long-term dividends. This means funding universities, supporting researchers who stay rather than emigrate, and building the institutional infrastructure that enables sustained development.

The World Bank, regional development banks, and bilateral development agencies are beginning to fund AI capacity building. The scale remains inadequate, but the direction is right.

Multilingual and local-context requirements:

Regulation can require AI systems deployed in a country to support local languages and contexts. The EU AI Act's documentation requirements could be extended to include language coverage disclosures. Procurement standards can favor systems that serve local needs.

Benefit-sharing frameworks:

When AI systems train on data from a population, that population should share in the benefits. This could take many forms: revenue sharing, technology transfer requirements, local hiring mandates, or public investment funded by AI company taxation.

Nick Couldry and Ulises Mejias's work on "data colonialism" suggests frameworks for understanding data as a resource that, like physical resources, should benefit source communities (Couldry & Mejias, 2019).

What Redistribution Means

The digital divide in AI is not inevitable. It reflects choices — about where to invest, what to build, whose needs to serve. Different choices would yield different results.

For researchers:

Redistribution means building AI that works in more languages, more contexts, more conditions. It means collaborating with researchers in the Global South as equals, not as data sources. It means sharing compute, datasets, and credit.

For companies:

Redistribution means pricing that accounts for economic differences. It means product design that works in low-bandwidth, intermittent-power conditions. It means investment in local capacity rather than extraction.

For policymakers:

Redistribution means data sovereignty frameworks that prevent extraction. It means public compute infrastructure accessible beyond wealthy countries. It means development funding for AI capacity building.

For all of us:

AI could be the technology that closes global gaps — bringing the world's knowledge to everyone, automating tedious labor universally, providing expert assistance wherever needed. Or it could be the technology that widens those gaps, concentrating capability in wealthy enclaves while extracting value from everywhere else.

Which it becomes depends on choices made now. The technology is not destiny. The infrastructure is not neutral. The distribution is a decision.

Sources

Bender, Emily M., Timnit Gebru, et al. "On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?" FAccT, 2021. ACM
Benjamin, Ruha. Race After Technology: Abolitionist Tools for the New Jim Code. Polity Press, 2019. Publisher
Birhane, Abeba. "Algorithmic Colonization of Africa." SCRIPTed 17, no. 2 (2020). SCRIPTed
Couldry, Nick and Ulises Mejias. The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Stanford University Press, 2019. Publisher
Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Publisher
Eubanks, Virginia. Automating Inequality: How High-Tech Tools Profile, Police, and Punish the Poor. St. Martin's Press, 2018. Publisher
Gray, Mary L. and Siddharth Suri. Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt, 2019. Publisher
International Telecommunication Union. "Measuring Digital Development: Facts and Figures 2025." ITU, 2025. ITU
Mohamed, Shakir, Marie-Therese Png, and William Isaac. "Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence." Philosophy & Technology 33 (2020). arXiv
OECD. "Talent for Innovation: International Mobility of STEM Students." OECD Publishing, 2023. OECD
Sambasivan, Nithya, et al. "Re-imagining Algorithmic Fairness in India and Beyond." FAccT, 2021. ACM

General

The Digital Divide in AI