The World Bank Just Told the Global South to Skip the AI Race — But Who Pays for the Leap?

Over the past decade, I’ve watched the World Bank’s policy papers land like weather systems over emerging markets. They arrive with gravity, full of charts and conditional urgency, and then the rain either comes or it doesn’t. In January 2025, the Bank released its Global Economic Prospects report with a forecast that should have stopped every finance minister in the Global South mid-sentence: global growth is heading toward a 30-year low. And buried inside that same document was an unusual prescription — not austerity, not debt restructuring, but a call for developing economies to rapidly adopt artificial intelligence. Not build it. Not regulate it. Adopt it. Fast. That word — adopt — is doing more work than it appears. It is the World Bank’s way of saying: you don’t need to train your own models. You don’t need your own chips. You need to plug into the existing AI stack and let the productivity gains flow through your agriculture ministries, your health clinics, your tax collection systems. On paper, this is leapfrogging logic — the same logic that let Kenya skip landlines and go straight to mobile money. Skip the legacy software era, skip the data center build-out, skip the PhD pipeline. Just consume AI as a service and grow. But I’ve spent enough time in this industry to know that leapfrogging has a shadow. When a policy institution as powerful as the World Bank tells dozens of low-income countries to adopt foreign AI infrastructure, it is not just offering a growth hack. It is quietly redrawing the map of who builds the future and who simply rents it. Here is the uncomfortable technical reality the report’s summary glosses over: the infrastructure precondition for AI adoption is not software. It is electricity, bandwidth, and data sovereignty. According to ITU data from 2024, only about 36 percent of people in low-income countries have internet access. Sub-Saharan Africa’s electricity coverage still hovers below 50 percent. You cannot prompt your way out of a power outage. The Bank’s recommendation implicitly assumes a baseline of digital infrastructure that most of these economies simply do not have. The gap between the policy ambition and the on-the-ground friction is enormous. Based on my experience auditing projects in emerging markets, I’ve seen this mismatch play out in miniature: a well-funded government portal that operates only in the capital city, a mobile health tool that dies on 2G networks, an AI crop advisor that runs on a server three continents away. None of these are malice. They are physics. And yet, let me steelman the World Bank’s position before I take it apart. The alternative to adoption is not indigenous model development. Training a 10-billion-parameter model costs somewhere in the low millions of dollars in compute alone — a sum that would absorb the entire AI budget of most low-income nations. The realistic choice is not between homegrown AI and imported AI. It is between imported AI and no AI at all. The Bank knows this. Its recommendation is pragmatic in the most brutal sense of the word. It is saying: you cannot afford to be left out of the productivity gains that American and Chinese firms are already harvesting from their models. So plug in. Use the APIs. Use the open-source weights. Just get moving. Here is the trap embedded in that pragmatism: the market structure of AI adoption is not neutral. When a developing country adopts AI quickly using foreign platforms, its data becomes the raw material for someone else’s model. Its health records, crop yields, and citizen behavior flow outward. In return, it receives intelligence shaped by the priorities of a company in Palo Alto or Shenzhen. This is what scholars have called data colonialism — a structural arrangement where the periphery exports raw data and imports finished cognitive services. The World Bank’s report acknowledges the risk of “technological dependency,” but it does not resolve it. It cannot. Because the only path that truly resists dependency — investing in open-source stacks, local fine-tuning capacity, and domestic talent — is slower, less flashy, and produces less immediate GDP headline movement. Let me now say something that will be unpopular in both Washington and the crypto conferences I used to speak at. The adoption-first strategy might actually be the correct one for countries that are utterly infrastructure-starved. But it is not a strategy. It is a stopgap. It is a way to get some benefit without building the underlying capacity to generate it. And the World Bank knows the difference. You can see it in the report’s language: the same document that urges rapid adoption also warns about widening inequality. Inside the Bank, there is a policy tension between the growth economists and the social development staff. The growth people see AI as a multiplier. The social people see AI as a magnifier — it amplifies whatever inequalities already exist. Both are right, which is why the recommendation is so hard to operationalize. The contrarian angle every technology optimist hates to hear: fast adoption without absorptive capacity is the surest way to deepen dependency. Take the example of business process outsourcing. For years, countries like the Philippines and India built robust industries on human data processing — document review, transcription, customer support. Generative AI is already eating into these entry-level digital jobs. The World Bank’s recommendation, if implemented naively, could accelerate the very displacement it claims to defend against. The countries that get the worst of it are those that adopt AI for their own public services but lack the domestic ecosystem to absorb the displaced labor into new roles. It is a textbook case of creative destruction with no trampoline underneath. But here is the thing I keep returning to as I read this recommendation through the lens of decentralization, which is my own bias and my own conviction: the World Bank is actually endorsing a form of AI outsourcing that mirrors the early cloud computing playbook. And cloud computing did eventually create local value in emerging markets — not from the infrastructure itself, but from the applications built on top of it. M-Pesa would not exist without the telecom rails. The future version of this story is open-weight models deployed on shared infrastructure, maintained by regional consortiums, customized by local developers. That is not a fairy tale. It is the technical path that Meta’s Llama and China’s Qwen models have already made possible. If a country cannot train its own frontier model, it can still fine-tune an open model on its own agricultural data in its own language. That is a fundamentally different relationship to technology than consuming black-box APIs. But it requires initiative, training, and a policy environment that rewards experimentation over procurement. The World Bank has not yet pushed hard enough in that direction. Its statement is a weather system. It is not yet rain. So what is the actual play for leaders in the Global South over the next eighteen months? If I were advising a finance minister — or a community organizer, which is closer to my current role — I would say this: treat the World Bank’s endorsement as an opening, not a prescription. It gives you license to put AI on the national agenda. It gives you cover to negotiate for better terms on digital infrastructure. But do not interpret “adopt quickly” as “rely solely on imported services.” Use the urgency to create the institutional foundations: AI governance frameworks, data protection laws, digital trust infrastructure, and a deliberate investment in open-source capacity. Because the only AI strategy that will survive a decade is one where the community that uses the technology also has a say in how it is built and governed. Trust is the only protocol that matters. Code is law, but people are the context. And community over coin, always. These are not slogans I deploy for comfort. They are operational principles that distinguish a society that uses technology from a society that is used by it. The World Bank has given the global south an invitation to the AI party. The wise guest will not just show up and consume — they will ask for a seat at the table where the menu is written. The uncomfortable question we should all be sitting with is not whether developing economies can afford to adopt AI. It is whether they can afford to adopt it in a way that leaves their own people and their own data as the asset, not the resource. Who builds the future matters. And in this future, I suspect, the countries that will actually thrive are not the ones that bought the most API credits. They are the ones that built the most local capacity to say no when the tech needs to serve a different master. The World Bank has called for fast adoption. I will not disagree with the speed. But speed without direction is just centrifugal force. The next hundred years of economic development are going to be written in the context of machine intelligence. Trust is the only protocol that matters. And the protocols we choose now — whether open or closed, participatory or extractive — will determine whether that future is shared or hoarded. The report is out. The weather system has arrived. The question remains: will the rain grow food, or just move the topsoil?