World Bank’s Small AI Case: $5 Tutors vs $775B in AI Capex

The World Bank’s case for “small AI” rests on a cost gap. Its World Development Report 2026 describes a maths tutor in Ghana, delivered by text message on basic phones, at a cost of about five dollars per student. The same report charts the projected 2026 capital spending of five US AI hyperscalers at $775 billion, above the 2025 GDP of 12 of the 15 economies it shows, by our count.

CNBC reported on 13 November that “small AI” has big potential, according to the World Bank. The argument comes from the Bank’s World Development Report 2026: The Promise of Artificial Intelligence, released on 4 August. We read its press release, its publication page and its Overview; we did not read the CNBC article and quote nothing from it.

Business Pill · Tokens Explained

An evergreen Business Engineer explainer on tokens, the unit AI reads, writes and bills in. It teaches the general idea only and does not discuss the World Bank report.

What the World Bank Means by Small AI

The report’s Box O.1 says that in the tech industry small AI refers to model size: small language models are trained to approach the performance of a large language model on far fewer parameters. Small AI can also be small in scope, built for one narrow task such as identifying a crop disease or screening a medical image. The report says small AI models can run offline on a phone or laptop.

A model can also stay large in the cloud, the report says, and still reach users in low-resource settings through a basic phone over a voice call or a text. It adds that small AI models are typically adapted from larger foundational models but can perform better on a narrow task.

The Examples and Their Price Tags

In Ghana, according to the report, Rori provides AI-powered maths tutoring by text message on basic phones and weak networks. It generated nearly one full year of learning from two 30-minute weekly sessions over eight months, at a cost of about five dollars per student. Cough Against TB, built by Wadhwani AI with India’s Central Tuberculosis Division, correctly flagged 9 in 10 people when screening for tuberculosis from the sound of a cough.

The report says Nuru, which runs on a phone and even offline, outperformed agricultural extension officers at diagnosing a diseased leaf. In Kenya, it says, the judiciary used AI to assign more than 10,000 court cases each year to more than 1,500 mediators. In India, weather forecasting systems that AI now supports let some smallholder farmers reap net savings of as much as $560 per farmer.

The Jobs Numbers Behind the Pitch

The press release says jobs in high-income countries are more than three times as likely to be at risk of automation by generative AI: 14.2% of existing jobs there, against 4.5% in low- and middle-income countries. It says 16.2% of jobs in developing economies could see productivity meaningfully boosted by AI, close to the 18.7% expected in high-income countries.

By our arithmetic, that is 3.6 jobs boosted for every job at risk in low- and middle-income countries, against 1.3 in high-income countries. Indermit Gill, the World Bank Group’s chief economist, said in the release: “They do not need large models or big data centers to reap its benefits”. “By adapting small, low-cost AI tools to local conditions, they can bring better medical care, education, judicial services and agricultural extension within reach of millions”, he said.

Big Tech’s Capex as the Yardstick

The report’s Figure O.5 sets projected 2026 capital expenditure by Alphabet, Amazon, Meta, Microsoft and Oracle, $775 billion, against 2025 nominal GDP. Brazil ($2,280 billion), Mexico and Indonesia sit above that bar. Argentina ($681 billion), Thailand, Viet Nam, Bangladesh, South Africa, Pakistan and six others sit below it, 12 of the 15, by our count.

The report’s framework is adopt, adapt and advance. It says advancing, building frontier models and the infrastructure that powers them, is by far the hardest and most expensive path, and that for most developing countries building frontier AI models is not a realistic goal in the near future. It says countries can make models like ChatGPT, Claude and Gemini more useful by connecting them to local documents, and names Latam-GPT and SEA-LION as regional language models.

The cheaper end of the model market is where we have tracked prices: see our report on open-weight token prices and Gemini 4 Hydrogen’s list price.

The Constraints the Report Names

The release says that in Sub-Saharan Africa nearly one-third of rural schools still lack reliable electricity and more than two-thirds lack dependable internet access. The Bank’s Mission 300 aims to provide energy access to 300 million people in the region by 2030. The report also says an AI tool to screen for a diabetes eye complication rejected more than 20 percent of nurses’ images because clinic lighting did not match its training conditions.

Gaurav Nayyar, the report’s director, said in the release: “The window to get this right is narrow”.

The Structural Read

As we read it, the report frames small AI as a cost choice: adapt cheap, narrow tools that run on basic phones instead of paying for frontier models and data centres. Only one of the examples it cites, Rori, comes with a cost per user, and the report itself names the missing electricity and internet access that limit such tools.

Three Implications

FOR GOVERNMENTS The report says governments should use their buying power to test, evaluate, buy and expand successful AI solutions in public services such as agricultural extension, health care and education.

FOR MODEL VENDORS The report says countries can make models like ChatGPT, Claude and Gemini more useful by connecting them to local documents and knowledge, or by building chatbots inside local messaging apps.

FOR THE JOBS DEBATE The release puts 4.5% of existing jobs in low- and middle-income countries at risk of automation by generative AI and 16.2% as jobs whose productivity AI could meaningfully boost.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of The Inference Economy.

The framework says: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”

As we read it, the World Bank’s small AI case is an argument about that marginal cost: the one cost per user it gives is about five dollars per student for a text-message tutor. The framework predates the report and does not discuss the World Bank or developing countries.

What Is Not Established

The Rori, Nuru and Cough Against TB results are the studies the report cites, not World Bank trials of small AI at national scale. The five-dollar figure is one programme’s cost per student, not a general price for small AI. The jobs shares are the Bank’s estimates of jobs amenable to automation or to being complemented, not forecasts of job losses. Figure O.5 compares one year’s spending with a year’s output; it is a scale comparison, not a like-for-like measure.

The Bottom Line

The World Bank’s report argues that developing countries can gain from small, low-cost AI tools without building frontier models. Its own figures put a text-message maths tutor at about five dollars per student and US hyperscaler capex for 2026 at a projected $775 billion; by our arithmetic, its jobs estimates give 3.6 jobs boosted per job at risk in low- and middle-income countries.