The World Bank's AI Leap: Policy Without a Power Grid

The World Bank's January 2025 Global Economic Prospects report told developing economies to adopt artificial intelligence rapidly. Not cautiously. Not conditionally. Rapidly. The same institution that tracks global infrastructure deficits—36 percent internet penetration in low-income countries, sub-50 percent electricity coverage in sub-Saharan Africa—recommended that the world's poorest nations sprint toward a technology whose deployment floor is a functioning electrical grid. The ledger remembers what the hype forgets. Filtered through commercial crypto media and into investor feeds, the recommendation has been framed as an opportunity for the Global South. I read it as a liability transfer, executed one narrative at a time. The report landed in a peculiar moment. Global growth is projected at 2.7 percent—the weakest half-decade in thirty years. The World Bank needs a growth narrative, and AI is the cleanest one available. The logic is deceptively simple: developing economies can leapfrog traditional development stages by adopting AI tools directly, bypassing the institutional scaffolding that wealthy nations took centuries to build. The same logic animated mobile payments in Kenya and telephony in Bangladesh. That precedent is doing heavy lifting it may not survive. I have tracked World Bank policy endorsements for two decades. When the institution blessed financial inclusion in the 2000s and digital infrastructure in the 2010s, capital followed within 24 to 36 months. Bilateral aid agencies, philanthropic foundations, and private investors treat the Bank's quarterly output as a capital allocation signal. This AI endorsement is the beginning of a pipeline, not the end of a debate. The transmission mechanism is predictable: policy visibility, then institutional resource allocation, then national AI plans, then project tenders. The question is who holds the contracts when those tenders close. The vehicle matters as much as the message. A commercial crypto asset outlet surfaced the story, delivering it to an audience primed to hear rapid adoption as validation for decentralized AI tokens or blockchain-verified identity narratives. The framing conveniently omits that the World Bank's recommended path is centralized, foreign-owned, and paid for in data. The medium is part of the message. I do not cover the story; I follow the code. Let me dissect the recommendation across three structural faults. First, the electricity problem. AI is a physical technology with physical requirements. Training a ten-billion-parameter model requires megawatts of consistent power. Inference—the actual deployment of AI—requires networked devices, reliable bandwidth, and data storage. In low-income countries, energy access is below half the population. Rural areas, where agricultural AI applications would deliver the highest marginal value, are precisely where the grid is thinnest. Load shedding in Lagos is not a technical footnote; it is the binding constraint on adoption. This is not speculative. In 2025, I audited a protocol claiming to use zero-knowledge proofs to verify human identity. The algorithm relied on training data that systemically excluded roughly thirty percent of global users, overwhelmingly those in the Global South with intermittent connectivity. Engineers in Sydney cannot make a model work on a network that drops every nine minutes. The constraint was never the model. It was the infrastructure. The World Bank's recommendation omits the sequencing question that determines everything: does infrastructure precede usage, or does usage force infrastructure into existence? Mobile leapfrogging worked because a phone tower is one-tenth the capital cost of fiber infrastructure. AI infrastructure—data centers, high-bandwidth interconnects, reliable power—is a different cost class entirely. Kenya's M-Pesa rode a network of cheap radio towers to success. There is no cheap AI radio tower. A single data center costs upward of $500 million, more than the annual national digital budget of most low-income countries. Second, the absorption problem. The recommendation implicitly bets on buying AI rather than building AI. Technology acquisition without absorptive capacity produces import dependence, not development. Economists call this the capability gap, and it explains why fertilizer imports did not industrialize African agriculture and why solar panel imports did not create Asian manufacturing hubs. I have watched the same error play out in crypto markets. In 2018, I audited EtherCity, a virtual real estate project whose ownership records were stored off-chain without cryptographic proof. Forty million dollars evaporated when the market tested that design. The analogue here: a developing economy adopting foreign AI services without understanding their internal governance transfers risk, not capability. It imports outcomes it cannot audit. The World Bank knows this. Its own analysts named foreign technology dependence as a risk in the same report that urges rapid adoption. That is a contradiction, and it is the kind of contradiction that appears when an institution optimizes for the headline rather than the outcome. The report offers no financing mechanism for local AI capacity, no open-source mandates to reduce dependency, no governance framework for the data that will flow across borders. Silence in the code is the loudest confession: it names the disease and prescribes the vector. Third, the data colonialism structure. The commercial logic is simple. Developing economies contribute data and market access. Foreign vendors contribute algorithms and collect recurring fees. Data flows out. Intelligence flows in. The bill flows forever. Scholars have documented this structure under the label data colonialism, and the World Bank's framing actively legitimizes it. The winners are cloud providers expanding into emerging markets, AI API vendors, and the consultancies that will assemble deploying-AI-in-developing-countries service packages. The losers are local AI startups that cannot compete with subsidized foreign providers, and citizens whose data leaves without consent or compensation. The labor market consequences are equally unexamined. For economies built on low-skill services—Philippine call centers, Bangladeshi data pipelines, Indian back-office processing—AI adoption is not a neutral efficiency gain. It is a structural shock delivered without a safety net. The World Bank's growth models treat adoption as a rising tide. The distributional ledger is not attached. The report also collapses the distinction between economies at different development gradients. Vietnam's manufacturing base, Indonesia's digital economy, and Niger's subsistence agriculture face entirely different AI adoption equations. One-size-fits-all rapid adoption is development policy from the era of one-size-fits-all structural adjustment. That era ended badly. But the bulls have a point, and dismissing it would be forensic malpractice. Open-source AI genuinely changes the calculus. Llama and Qwen models are free to download, fine-tune, and deploy. A developing economy with modest cloud budgets can access frontier-adjacent intelligence that would have cost millions in 2022. This is real, and it is new. Cloud-based inference enables a thin-client model: smartphone penetration in developing markets now exceeds sixty percent, meaning a government ministry can use a language model to process permit applications that previously required bribes and weeks of delay. A farmer querying pest management models via SMS is adoption justified. There is also a defensive argument for urgency. If AI capability concentrates entirely in the United States and China, developing economies face a worse outcome: no seat at the table, no agency in training data, no voice in safety standards. Adoption, even imperfect adoption, is a form of participation. The World Bank's recommendation may be the only politically viable path to keep the Global South in the conversation. My critique is therefore not with adoption. It is with the absence of a governance spine. The recommendation lacks an institutional architecture for managing dependency, verifying claims, and distributing gains. It is a growth call with no accountability. Watch the 2025 project pipeline. If AI readiness becomes a loan conditionality—if financing windows appear within twelve months—the recommendation is operationalized, and scrutiny matters. If it remains a paragraph in a quarterly report, it is narrative management. And for the poorer economies that adopt AI on borrowed credibility, the bill arrives in data, in dependency, and in political autonomy. We traded value for visibility, and lost both. I hope the Bank's ledger reads differently this time. I do not expect it to.