Across the Global South, governments are building AI models trained on their own languages and for their own needs. None of them are trying to match the frontier. Instead, attempts are being made to train AI for specific jobs and local languages. Nigeria created one. Chile created one. Ghana approved hundreds of millions for the computers to train them. The pitch for this is to build AI that is sovereign, independent of American or Chinese labs, built with our data, our languages, and our infrastructure

Nigeria’s N-ATLAS is Meta’s Llama 3 model fine-tuned on Yoruba, Hausa, Igbo, and Nigerian English. Latam-GPT, built by Chile’s national AI centre with fifteen countries, does the same for the region. India’s Sarvam-M did the same on France’s Mistral Small, though Sarvam has since released larger models trained from scratch, and Southeast Asia’s SEA-LION family now builds on Google’s Gemma and Alibaba’s Qwen after dropping Meta’s Llama. These models are modest, but no company will cover these languages for free, so the state has to. The same goes for identity, payments, welfare and defence. A country cannot rent those forever. They have to run on domestic infrastructure, compliant to domestic law and audit.

These models are modest, but no company will cover these languages for free, so the state has to. The same goes for identity, payments, welfare and defence. A country cannot rent those forever.

This is not just a collection of isolated projects. From 19 to 22 October 2026, the World Bank and South Korea will host the Global AI and Digital Summit in Seoul under the banner “Small AI, Big Impact”. Earlier this month, a World Bank hackathon asked young innovators to build low-cost AI tools for health, agriculture and tourism that could work with weak internet connectivity and basic phones. Three winners will present their work in Seoul. This reflects a shift towards AI built for specific local needs.

We should expand this work. It is necessary, but it will not be enough on its own.

The trouble starts outside these lab projects. More than 40 percent of ChatGPT traffic comes from middle income countries, led by Brazil, India, Indonesia and Vietnam.

The trouble starts outside these lab projects. More than 40 percent of ChatGPT traffic comes from middle income countries, led by Brazil, India, Indonesia and Vietnam. The state can build any small model it wants, but the citizens still open a foreign app. That is because no sovereign model attempts to beat the frontier. Price alone will not close the gap. The cost of running an AI model with roughly GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024. This is not the cost of frontier AI itself, but it shows why capable AI services remain affordable even as the best models advance.

Moreover, methods developed at the frontier travel quickly to specialist fields. Take weather forecasting. In 2023, DeepMind’s GraphCast outperformed the European Centre for Medium-Range Weather Forecasts’ flagship forecast on about nine in ten of the measures tested. Its later GenCast system beat the centre’s ensemble forecast, which estimates a range of possible outcomes, on 97% of the targets tested. The centre has since incorporated AI forecasting into its own AIFS system.

The point is not that AI has replaced conventional forecasting. It is that advances in AI research can transform specialist fields too.

The point is not that AI has replaced conventional forecasting. It is that advances in AI research can transform specialist fields too. Similar breakthroughs have happened in image recognition and even in matrix multiplication, where a new method improved on an approach that had stood for 56 years.

Leading AI labs in the West are pursuing artificial general intelligence (AGI), systems that could perform a wide range of skilled tasks. Some researchers also seek systems that can help improve their own capabilities. Whether or not AGI arrives, this research keeps producing methods and technologies that can be applied elsewhere. Countries that only adapt existing models risk depending on innovations developed by others.

Brazil offers a warning. From the late 1970s, it restricted foreign competition in its computer market. Its domestic industry grew from nine firms to 306 by 1987, but much of that growth involved copying foreign machines rather than developing the technology behind them.

When protection ended, and the market opened, the domestic industry shrank sharply. By 1994, employment in the computer industry had fallen 55%, nine out of ten newly purchased software programs were foreign, and Brazilian-controlled research had been virtually eliminated. Much of what survived was assembly, while advanced chip fabrication remained abroad.

India’s parliamentary Standing Committee on Communications and IT reached a related conclusion in December 2025. It found that joining the WTO’s Information Technology Agreement early had exposed domestic firms to import competition before the industry was ready, deepening dependence on imports.

History does not repeat; it rhymes. The computer race was never about the computer itself. It was about who could design and fabricate the silicon inside it, and losing that pushed almost everyone out. Brazil kept assembly and lost the capability that mattered. The AI race is the same, only worse. It runs on model designs and training methods only a few labs master, locked to chips only one supply chain can build.

Brazil kept assembly and lost the capability that mattered. The AI race is the same, only worse. It runs on model designs and training methods only a few labs master, locked to chips only one supply chain can build.

The advantage of established labs also compounds. They already have millions of users, large revenues, and years of experience deploying models. When companies can turn user feedback and revenue into better models, those improvements attract still more users, strengthening their lead.

This concentration is visible in both consumer and enterprise markets. In August 2026, ChatGPT and Gemini accounted for about four in five visits to AI chatbot websites tracked by Similarweb. Separately, Menlo Ventures estimated that Anthropic, OpenAI and Google together accounted for 88 percent of enterprise LLM API spending in 2025, with Anthropic leading at 40 percent. These figures measure different markets, but both show how much usage is concentrated among a small number of providers.

The gap is not insurmountable. Epoch AI estimates that the best open-weight models trail the closed frontier by about four months. But countries that adapt existing models remain dependent on what upstream developers choose to release. Singapore’s SEA-LION programme, for example, moved from Meta’s Llama to Alibaba’s Qwen in 2025. Countries seeking to reduce this dependence by developing their own models face a different obstacle. They must compete against established firms that have spent years building models, data systems, and infrastructure.

Most countries therefore adapt models already available on Hugging Face.  While a practical approach, fine-tuning an existing model is common practice, not a unique technological advantage.

While a practical approach, fine-tuning an existing model is common practice, not a unique technological advantage.

Even if we use existing models and host them ourselves, we still depend on an industrial supply chain few countries can reproduce. ASML supplies the machines used to manufacture the most advanced chips, TSMC fabricates many of them, and NVIDIA dominates the AI-accelerator market.

Planned compute capacity does not always translate into operational infrastructure. India approved its AI compute mission in March 2024, but by September 2025 only 17,374 of the 34,333 GPUs proposed across its first two rounds had reportedly been installed. Kenya’s planned Microsoft-G42 data centre also ran into delays after negotiations over government payment guarantees broke down.

When demand exceeds supply, the largest buyers are better placed to secure capacity. In practice, those include the major Western cloud providers. This does not require an explicit decision to favour the West. It follows from who places the largest orders, while smaller buyers can be left waiting.

That leaves us with two broad paths. We can try to build frontier models of our own, but without a major breakthrough or enormous investment in hardware, talent and compute, this is far from viable. Even then, our models would have to compete in a market where capability, price and efficiency shape adoption. Or we can continue adapting existing models for specific sectors and local needs, while hosting them ourselves. This is practical, but it cannot by itself deliver independence. The base model is still developed elsewhere, the chips are expensive and age quickly, and the next generation is already being trained upstream.

We can try to build frontier models of our own, but without a major breakthrough or enormous investment in hardware, talent and compute, this is far from viable. Even then, our models would have to compete in a market where capability, price and efficiency shape adoption.

Dependence on foreign models and chips is one risk. A second is that AI may undercut services these countries export. The IMF estimates that around 40% of jobs globally are exposed to AI, rising to 60% in advanced economies. The World Economic Forum also finds that more than a third of young workers, and three-quarters of young workers in Eastern Asia, are in occupations with medium to high AI exposure.

While exposure does not mean that every job will disappear, it shows how widely AI could change the work people do.

The risk specific to the South sits in outsourcing. In the Philippines, an IMF paper found that around one in three workers is highly exposed to AI, though about 60 percent of these highly exposed jobs are also rated highly complementary, suggesting AI may assist rather than replace workers in those roles. A newer ILO study, using a different method, finds that more than one-quarter of Philippine employment, or 12.7 million jobs, is exposed to generative AI, the highest share among ASEAN countries with comparable data. Customer service, sales, and clerical roles are particularly vulnerable. The country’s IT-BPM industry employed about 1.9 million people and generated more than $40 billion in 2025. A 2025 report by Caribou Digital and Genesis Analytics, produced in partnership with the Mastercard Foundation, estimates that around 40 percent of tasks in Africa’s business-process outsourcing and IT-enabled services sector could be automated by 2030.

These figures describe exposure, not certain job losses. But they point to a real vulnerability. Many nations of the Global South are significant exporters of the office-based services most exposed to AI.

These figures describe exposure, not certain job losses. But they point to a real vulnerability. Many nations of the Global South are significant exporters of the office-based services most exposed to AI. Women face a disproportionate risk. The ILO finds that in the Philippines, 40.3 percent of jobs held by women are exposed to generative AI, compared with 19.3 percent of jobs held by men. Poor internet access and weaker digital infrastructure may also leave workers in poorer countries less able to benefit from AI’s productivity gains.

Building our own AI infrastructure and models eases the anxiety of this wave to an extent, but the gains do not stay home. They rise to the top of the chain, to the semiconductor suppliers, who are the only players positioned to profit whichever way the race turns. NVIDIA designs the accelerators, ASML builds the machines that print them, and TSMC fabricates them. Every lab and government in the race, winners and losers alike, ultimately depend on this supply chain, directly or through cloud providers. In a gold rush, the sure money was never in the gold, but in the shovels. The semiconductor industry sells the shovels. Every model we train sits on that silicon. Build our own models or not, unless that silicon is ours, we stay the prospectors.