World Bank Urges 'Small AI' for Developing Economies

By Marcus Bennett

The World Bank has called on developing countries to start building "small AI" models — compact artificial intelligence systems capable of running on a mobile phone with low connectivity — arguing the technology could compress economic development timelines from centuries into decades.

The call came from Paschal Donohoe, the World Bank's chief knowledge officer, in remarks reported by The Guardian on October 18, 2026. His central claim is stark: AI could accelerate development for emerging economies, but only if the technology adapts to local needs rather than arriving as a one-size-fits-all export from wealthy markets.

That is the whole of the argument. It is also a redirect of one of the most consequential resource questions in global AI policy.

What is 'small AI'?

In the World Bank's framing, small AI means models engineered to run directly on a mobile phone, even where bandwidth is scarce or unreliable. The constraint defines the design. A model that must operate on a low-cost handset, on a weak network, cannot depend on the data-center-scale infrastructure that underpins flagship systems from Silicon Valley firms.

The distinction matters. The report positions small AI against the current trajectory of the big technology companies, which, as The Guardian notes, are focused on ever more sophisticated consumer products. That race rewards compute, scale and capital. A phone-native model rewards efficiency, local-language data and adaptation to specific economic conditions.

Donohoe's proposition is that developing countries should not wait for that race to produce something they can import. They should build for themselves.

Why does the World Bank's position carry weight?

The World Bank is not a technology vendor. It is one of the largest sources of development finance and policy advice for low- and middle-income countries, and its chief knowledge officer shapes how the institution thinks about the role of information and technology in growth. When that office says AI could turn development processes measured in centuries into processes measured in decades, the statement functions as a signal to governments, ministries and development agencies worldwide.

The framing also lands at a moment of sharp divergence in how AI's benefits are distributed. Frontier model development concentrates in a small number of firms and countries with access to massive computing clusters. Most of the world's population accesses the internet through mobile devices, frequently on limited connections. If AI's economic value only arrives through high-end consumer products and cloud dependencies, large parts of the globe will receive it late, expensively or not at all.

Small AI is the counter-proposal. Fit the model to the phone the user already owns. Fit the model to the connection the user already has.

What did Donohoe actually say?

The Guardian's report carries the core of the position in Donohoe's own terms: AI could speed up economic development "from centuries to decades." The word choice is doing deliberate work. It does not promise leapfrogging, and it does not promise parity with advanced economies. It promises compression — the same distance covered in a fraction of the time.

The condition attached is equally deliberate. Donohoe said AI could accelerate development for emerging economies "but only if it was adapted to local needs," according to the report. Adaptation, not adoption, is the operative requirement. A model trained overwhelmingly on the data, languages and use cases of wealthy markets may arrive in a developing economy technically functional and practically misaligned — unable to serve a farmer, a clinic or a small business operating in conditions its training never represented.

How does this differ from Silicon Valley's roadmap?

The contrast the World Bank draws is explicit in the source. While big technology firms in Silicon Valley chase sophistication — larger models, richer consumer features, more demanding hardware — the development case runs in the opposite direction: smaller models, fewer resource requirements, deployment at the economic margins where a smartphone may be the only computing device available.

The stakes are structural. If AI follows the pattern of earlier general-purpose technologies, the countries that build and adapt it early will capture disproportionate gains, and the gap between early and late adopters will widen before it narrows. The World Bank's message to developing countries is that they are not obliged to be late adopters. The technical requirements for small AI — a phone, a model sized to it — are within reach in a way that data-center-scale AI is not.

What happens next?

The immediate implication of Donohoe's call is an agenda, not a product. Developing countries are urged to "kickstart the creation" of these models — language that puts the initiative with national governments, local research communities and regional developers rather than with foreign vendors.

The open questions are the ones the source leaves unanswered: which countries move first, what local data and languages those initial models serve, and whether the development-finance apparatus that the World Bank sits atop will fund the effort. What is already fixed is the argument's shape. As The Guardian reported on October 18, 2026, the World Bank's chief knowledge officer has put a name — small AI — and a timeline — decades instead of centuries — on a proposition that developing economies can build their own path into the AI era, one phone-sized model at a time.