The World Bank Told the Global South to Adopt AI Fast. It Forgot to Mention Who Owns the Receipt.

The number that matters is not a price. It is a year. According to the World Bank's January 2025 Global Economic Prospects report, the global economy is on track to post its weakest half-decade since the early 1990s. The diagnosis is not new. The prescription is. The bank is telling developing economies to adopt artificial intelligence quickly, not as a pilot program, not as a research initiative, but as a growth strategy. This is the kind of announcement that usually gets reduced to a paragraph and a link. Crypto Briefing compressed it into a news brief. The blockchain-native readers who saw it were encouraged to imagine that 'AI adoption' and decentralized compute were about to collide. But the report itself contains two sentences that should have stopped that excitement cold. Faster adoption, it says, could widen inequality. It could also deepen dependence on foreign technology. I have written obituaries for more bull markets than I care to remember. The pixel wasn't the asset in the NFT boom; the community didn't understand that until the floor vanished; and the JPEG did depreciate, along with the promises. The World Bank's AI recommendation carries a similar structure: a shiny object, a global audience, and a hidden ledger. This is not a report about technology. It is a report about ownership. And ownership is the subject no one wants to put in the headline. Context: What the World Bank Actually Said Let me be precise about what the World Bank was actually saying. The report is a macro document, not a technology policy paper. Its central recommendation, as summarized and reported, is that developing economies should embrace AI applications quickly. It is not telling countries to build their own GPT models. It is telling them to use the tools that already exist. That distinction matters more than it appears. 'Adoption' means using AI for crop disease detection, administrative document processing, medical triage, translation, and financial inclusion. 'Development' means building the model, training it, hosting it, and owning it. The World Bank, based on the public summary, chooses adoption. This is consistent with its traditional toolkit. The bank does not build power plants or data centers; it helps countries create the conditions for private capital to build them. It provides loans, policy advice, and a very powerful seal of approval. In the 2000s, the seal was for microfinance. In the 2010s, it was for digital infrastructure. In the 2020s, it is for AI. Each of these themes was real. Each of them also created a patronage pipeline: if the World Bank says a thing matters, consultants will write report after report about it, governments will create ministries for it, and technology vendors will suddenly discover a deep concern for poverty reduction. The bank's underlying assumption appears to be leapfrogging. Emerging markets skipped landlines and jumped to mobile phones. They skipped desktop software and went straight to mobile apps. Why not do the same with AI? Why not skip the mature IT layer and build an AI-native public sector? The logic is attractive. The problem is that the previous leapfrogs did not require heavy hardware at the point of consumption. Mobile phones are cheap enough, and cell towers eventually became ubiquitous because the network economics worked. AI services, in contrast, require massive central computing resources, data storage, and a reliable electrical grid. In low-income countries, those foundations are missing. That is the elephant sitting on the report, and the report does not seem to have a name for it. Core: The Adoption Economy's Hidden Procurement Let's talk about the money first, because everything else follows. A low-income country that follows the World Bank's advice will not train a frontier model. The compute cost alone for a 10-billion-parameter model is measured in millions of dollars. Very few national AI budgets in the global south can absorb that, even before accounting for talent, data curation, and regulatory overhead. So 'fast adoption' inevitably becomes procurement. The country either buys API access from a foreign cloud provider, or it downloads an open-weight model and finds someone to host it. Both paths route through someone else's data center. This sounds like a technical detail. It is actually the beginning of a balance-of-payments problem. Every AI query sent to a U.S. data center is a small payment flowing north. Every model downloaded from a Chinese open-source repository is a soft dependency that can be adjusted at a licensing level. When the World Bank inflates the 'growth at 30-year low' fear, it is making countries more willing to spend their scarce foreign reserves on these subscriptions. I have seen this pattern in stablecoins. Tether claims to back its coins; the market accepts the claim; and independent audits remain oddly rare. The AI cloud is the new Tether. You cannot open the source code of a hosted frontier model. You cannot inspect the training data. You cannot run your own audit. You can only trust the data center, the billing algorithm, and the terms of service. For a developing country trying to build a trustworthy public sector, that is an enormous risk to park under the label 'adoption.' The phrase 'AI readiness' is already starting to circulate in development finance circles. It will not be long before it becomes a loan condition. A health ministry seeking concessional funding might be asked to demonstrate its AI adoption plan. That sounds progressive. But in practice, the easiest way to demonstrate AI readiness is to sign a contract with an established global vendor. The World Bank's report therefore becomes a recommendation to use American or Chinese AI infrastructure, packaged as a neutral development policy. That is not neutrality. It is a commercial outcome, quietly pre-written. Core: Data Dependency Is the Real Structure The most overlooked asset in the entire AI conversation is data. Not models, not chips, not electricity. Data. And the data of the global south is already flowing outward, one prompt at a time. Consider a simple agricultural extension program. A farmer uploads a photo of a diseased plant to an AI assistant that recognizes the disease and recommends a pesticide. The farmer receives a valuable answer. But the photo, the location metadata, the soil conditions, and the identity of the farmer all travel to the server that processed the request. That server is almost certainly in North America, Europe, or China. The benefit of the answer remains local. The value of the data becomes global. This is the structure of data colonialism: raw data leaves the periphery, processed intelligence returns at a price. The World Bank's report names 'dependence on foreign technology' as a risk, but it does not name the mechanism. It does not discuss data localization. It does not mention open-weight models. It does not propose a legal framework for digital sovereignty. It simply warns and moves on. The crypto world likes to claim that decentralized networks solve this problem. I have tested enough of these protocols to say the ambition is real and the execution is incomplete. Decentralized inference is slow. Verifiable computing is expensive. The governance of a distributed AI network is unresolved. And the tokens that fund them are notoriously correlated to the market, not to the infrastructure. Still, the direction is correct. A blockchain-verified model weight is a proof of provenance. A decentralized storage layer keeps data local. A decentralized compute market allows a government in West Africa to buy processing from a provider in South America instead of a hyperscaler in Virginia. This is not a fantasy. It is an engineering problem that has only half been solved. The World Bank could accelerate that work by funding projects that build public, distributed compute infrastructure. Instead, it is likely to fund the easier path: cloud adoption, and dependency. The pixel wasn't the innovation. The community didn't get the valuation model right. But the infrastructure bills never depreciate as quickly as the hype does. That is why the next ten years of AI development will be decided by ledgers, not by whitepapers. Core: Infrastructure Is the Hard Constraint The greatest misreading of 'fast AI adoption' is the assumption that it can happen without physical construction. Let's use the hard numbers. In low-income countries, internet penetration is around 36 percent. In sub-Saharan Africa, electricity access is below 50 percent. No bright algorithm can run without power. No brilliant language model can answer from a disconnected village. The mobile-first model is the only realistic path for scale. Smartphone penetration in emerging markets is high enough to dream. But every AI interaction is a request that travels to a compute node and waits for a response. If the node is on another continent, the latency is not just a user experience issue. It is a reliability issue. A crop diagnostic system that takes twenty minutes to respond on a bad connection will not be used. A medical triage tool that cannot reach the cloud during a power outage is not a tool. The geographical distribution of compute remains staggeringly uneven. Of the world's roughly 800 hyperscale data centers, Africa holds less than two percent. Southeast Asia and Latin America are expanding, but they are starting from a base designed for consumer internet, not for AI training. The cloud providers are building new facilities, but their commercial incentives follow the clients who can pay in dollars or euros. The poorest citizens, the ones the World Bank claims to care about, are the least attractive customers. This is why the phrase 'leapfrogging' deserves scrutiny. Mobile leapfrogging worked because the infrastructure was distributed to the edges of the network. AI leapfrogging demands the opposite: a massive centralization of compute resources. The user does not need a PC, but someone somewhere needs a billion-dollar data center. The countries most encouraged to adopt AI are the least positioned to own that data center. They will become consumers of a utility that they neither control nor regulate. The report, at least in its public summary, does not address this. It treats AI as a cloud-based service whose cost is measured in tokens or API calls, not in gigawatts and submarine cables. But development economics is not tokenomics. The price of a society's future can be measured in concrete, copper, and electrons. Core: The Governance Vacuum There is also a governance gap that goes almost unmentioned. Most developing economies do not have an AI policy framework. According to the Stanford AI Index, only around 10 percent of African countries have a national AI strategy. The World Bank is telling countries to run toward a technology that most of their own legal systems have not defined. There is no data protection regime. No accountability mechanism. No rule for what happens when a model makes a decision that destroys a small business or denies a health benefit. The reason this matters is that AI is not a static tool. It is a dynamic system that makes decisions under uncertainty. In the developed world, regulators are scrambling to catch up. In the developing world, the scramble has not even started. To recommend 'fast adoption' without 'robust governance' is to recommend that railways be built before the law of contracts exists. In my audit experience, I learned to look for the owner key in a smart contract. The person who controls the owner key controls the funds. In AI governance, the owner key is hidden inside the cloud. The country using the system may have a logo on the dashboard, but it does not have administrative control over the training set, the drift of the model, or the override rules. That is not sovereignty. It is usage. Core: Labor, Inequality, and the Missing Safety Net The labor market implications are not theoretical. AI automates the tasks that developing economies have historically used to enter global supply chains. Data entry is already being automated. Basic customer service is already being automated. Translation, transcription, and simple document review are being automated. The sectors that suffer are the ones the World Bank has counted on as pathways out of poverty. The Bangladeshi garment worker, the Kenyan call center agent, the Filipino BPO employee: these workers do not appear in the report's growth projections. What appears is a macroeconomic estimate of GDP. But if a technology displaces the lower rungs of the labor market while rewarding the high-skill workers who can manipulate the models, the result is not 'growth'; it is more inequality packed into a different shape. The World Bank knows this. It has published research on premature deindustrialization and the dangers of automation in the global south. Its own report warns about widening inequality. But it does not offer a concrete mitigation plan. It does not explain how a worker who spent eight years learning to file claims or review contracts will be retrained to manage an AI system's exceptions. It does not estimate the fiscal cost of an unemployed youth bulge. I recall sitting in a conference room in Brussels during the DeFi summer of 2020. A founder with a beautiful deck and an un-audited contract told me his protocol would democratize finance. My article was among the first to spotlight it. The protocol lasted until the reentrancy bug was exploited, and my coverage was cited as an example of hype. That lesson followed me: enthusiasm without risk weighting is just the financialization of optimism. The World Bank's enthusiasm for rapid AI adoption is the same double exposure. It is not wrong to suggest that AI can improve farming yields or teacher training. It is wrong to suggest that those benefits can be captured without managing the distributional shocks that arrive in parallel. 'Fast adoption' without a buffer is a policy, not for the future, but against the most vulnerable. Core: The Open-Source Escape Hatch If the global south wants to avoid the foreign dependency trap, the most promising path is open-source AI. Models like Llama, Qwen, and Mistral are free to download. They can be fine-tuned on local languages and local data. They can be hosted in regional data centers. They represent something the proprietary AI world cannot offer: the possibility of digital ownership. Open source is not a silver bullet. It requires skilled engineers. It requires data pipelines. It requires the very infrastructure that is missing. But it is a much better starting point than an API subscription, because every API call is a lopsided exchange. The data leaves the country, the model remains opaque, and the billing resets every month. The World Bank's report, based on the public summary, does not emphasize this route. That is an important omission. If the institution is genuinely worried about 'dependence on foreign technology,' it should be championing open-weight model ecosystems and national compute clusters. Instead, it is likely to be captured by the vocabulary of cloud adoption and AI readiness, which is the vocabulary of the vendor economy. For the crypto industry, this is the moment to prove that decentralization is more than a meme. Federated learning, proof-of-inference, and decentralized storage are all relevant. But they need to be built to a standard where a ministry in Lagos or Jakarta could deploy them without a team of PhDs in cryptography. The community didn't lose credibility because it was wrong about the future. It lost credibility because it refused to do the boring work of making the future usable. Core: The Investment Signal Is Real But Delayed For investors, this report is not a buy or sell order. It is a signal that will take years to become revenue. The World Bank does not buy AI stocks. It influences sovereign risk assessments, development finance, and the thematic preferences of a vast ecosystem of contractors, NGOs, and multilateral banks. History suggests a timeline. When the World Bank embraced digital infrastructure as a theme, it took roughly two to three years before the first wave of related projects reached procurement. If the same happens with AI, the period from 2025 to 2027 will see a stream of 'AI for development' requests for proposals, feasibility studies, and pilot deployments. The vendors who benefit are not the frontier model labs. They are the cloud infrastructure providers, the system integrators, and the consulting firms that know how to navigate development procurement. The creators of open-source models also benefit, because they are cheaper to deploy and easier to localize. The report's phrase about a 30-year growth low is itself an investment signal. It creates urgency. It makes governments feel they cannot afford to wait for careful planning. The psychological effect is to lower the political resistance to accepting foreign technology and external financing. For prudent investors, the same report is a warning: the risk of failed AI projects in countries without infrastructure is high. The best positioned players are not the ones who promise the greatest theoretical leap, but the ones who can operate in low-connectivity environments with small budgets and tolerate high failure rates. Political strategy and the 'adopt fast' compromise Now let me say the thing that most coverage is afraid to say. The World Bank's report is not a technical recommendation, and it is not solely a development recommendation. It is a political strategy for a multilateral institution in a world where its traditional toolkit is failing. For twenty years, the orthodoxy included structural reform, debt management, and trade liberalization. Those reforms were painful, slow, and politically difficult. In a context of weak growth, the World Bank needs to offer a story that is not a repeat of austerity. AI is the perfect story. It is forward-looking. It is untested enough to be flexible. It suggests that growth can be generated by technology rather than by asking governments to do hard things. The report's internal tension — fast adoption, but beware of inequality and dependency — is the visible trace of an institutional compromise between the macro economics department and the social development division. The phrase 'absorptive capacity' is the one that should dominate the discussion. You can recommend adoption, but adoption is constrained by the ability of a society to absorb technology. That capacity includes digital literacy, administrative institutions, competitive markets, property rights, and energy security. None of those are importable. A country cannot order absorptive capacity from a cloud provider. This is where the crypto/Web3 counter-narrative becomes both useful and dangerous. Decentralized AI infrastructure is genuinely one of the only answers to the foreign dependence warning. But if the crypto industry treats this report as an excuse to pump tokens with the words 'AI plus crypto' attached, it will burn the trust it is trying to build. The global south has seen enough hype cycles. It does not need another exclusive community. It needs a submarine cable, a skill certificate, and an audit trail. The community didn't lose the last bear market because it was too idealistic. It lost because idealism was priced as a token, not built as infrastructure. The same failure is available now, at a much larger scale. What responsible fast adoption would look like It is not enough to say the World Bank is wrong. The better question is what a responsible policy would actually require. It would not begin with model selection. It would begin with sequencing. First, it would invest in the physical layer: electricity, broadband backbones, and regional data centers. Second, it would fund digital literacy at the institutional level, because a ministry that cannot use email will not safely use AI. Third, it would establish data governance before deployment, so that government records do not become the permanent property of a foreign cloud. Fourth, it would require open standards and exit clauses. Every AI adoption contract should allow the host country to migrate its data and workflows to another provider or to an in-house system. If that is impossible, the adoption plan is not a plan; it is a lease. A responsible AI fund would not measure success by the number of pilots or the number of prompts processed. It would measure success by resilience: what happens when the cloud provider raises prices, the model is discontinued, or the network is down for two days? A system that cannot survive those shocks is not a development tool. It is a recurring vulnerability. Takeaway: What to watch, and what to ask The next twelve months are decisive. If the World Bank opens a dedicated financing window for AI adoption, the recommendation has moved from rhetoric to allocation. If it does not, it remains a mood signal. Watch the national plans of India, Indonesia, Nigeria, and Vietnam. Are they budgets for imported AI services, or are they investments in local compute, data sovereignty, and open-source capacity? The difference will decide whether the global south consumes the AI revolution or owns a piece of it. The exact same question applies to the crypto industry: are we selling a token price, or building the infrastructure that lets a village in Kenya run a model on a server in Nairobi instead of sending its data to a cloud in North Virginia? The World Bank's report is not wrong to call for AI adoption. It is incomplete. The missing chapter should describe who owns the model, who controls the data, who pays for the electricity, and what happens to the worker who is replaced before the new jobs are born. Until that chapter is written, 'fast adoption' is not a growth strategy. It is a transfer of value, dressed up as progress. The pixel wasn't a revolution. The community didn't need a cheerleader. And the trust that gets spent when institutions oversimplify complex futures will depreciate long before the technology proves itself. The global south deserves better than a press release. It deserves an infrastructure plan. That is the story I intend to keep covering.