January 2025. The World Bank drops its Global Economic Prospects report like a verdict. Global growth stands at its lowest point since 1990 — 2.3 percent, a thirty-year floor once you strip out China and India. The prescription, buried in policy language but unmistakable in its urgency: developing economies must rapidly adopt AI. Use the tools. Close the gap. Don't get left behind.
The report even names the side effects. Inequality could widen. Foreign technology dependence could harden into a permanent condition. Then the words "rapid adoption" sit there, clean and optimistic, as if the infrastructure needed to make AI work in Lagos, Dhaka, or Port Moresby had already been built while we weren't looking.
I have spent seventeen years watching narratives get minted. In 2018, when I audited Harvest Finance's early alpha contracts during the Ethereum Frontier days, the code didn't match the promise of "yield for everyone." It didn't match in 2020, when SushiSwap's fork mechanics promised the same yields as Uniswap with fewer checks and less scrutiny. It didn't match in 2022, when Terra's algorithmic stablecoin promised the same yield as cash with none of the risk. The World Bank's AI adoption call is a policy token minted in hope, and someone needs to read the ledger before the market prices it in.
The World Bank is not a protocol. It is the settlement layer of international development finance. Its Global Economic Prospects report, published twice annually, is the closest thing the development world has to a block reward distribution schedule. Every finance ministry in the Global South reads it. Every bilateral aid agency calibrates its country strategies against it. Every sovereign bond investor scans it for risk signals. When the World Bank says a technology is important, that statement gets priced into billions of dollars of future resource allocation, whether anyone writes a smart contract for it or not.
The January 2025 report arrives at an awkward moment. The global economy is stuck in a productivity and demographic rut. The forecast for world GDP expansion hovers in the low single digits for the foreseeable future. Structural reforms — trade liberalization, debt restructuring, labor market liberalization — remain politically radioactive in most capitals. The World Bank needs a growth story that does not require any finance minister to lose a domestic political fight. AI is that story.
The technical recommendation, at its core, is about adoption rather than development. The World Bank is not asking Nairobi to build a foundation model from scratch — a 10-billion-parameter training run costs somewhere in the low millions of dollars, a sum that exceeds the entire annual AI budget of most low-income countries. The bank is asking Nairobi to use the models that already exist. Cloud-based generative AI tools. Machine learning pipelines for crop forecasting. Automated document processing for customs. Chatbots for health education. The logic is pure leapfrog economics. A Vietnamese customs officer with an AI-powered document processor can handle the workload of three officers. A Nigerian agronomist with a chatbot trained on local weather and soil data can reach a hundred times more farmers than she could through field visits. Any of these applications generates measurable productivity growth. All of them require only a smartphone and a stable internet connection.
On paper, this is the mobile money story all over again. In the 2000s, the World Bank and its partners watched Kenya skip the fixed-line telecommunications era entirely, jumping straight to mobile payments. M-PESA became the canonical example of leapfrog development. The same logic now applies to AI. Why build a traditional enterprise software stack when a mobile-based AI assistant can deliver more value at a fraction of the cost? Why wait for a national data center strategy when OpenAI, Anthropic, Google, or Alibaba will host the inference for you?
But there is a difference between a story and a ledger. I spent the DeFi Summer of 2020 watching traders celebrate triple-digit APYs while my Python scripts quantified the slippage risk that the celebrants were systematically ignoring. The World Bank's AI adoption thesis has the same shape. The upside is visible. The slippage risk is distributed across infrastructure deficits, balance sheets, and governance gaps that no one in the policy narrative wants to quantify.
The Infrastructure Ledger
Let us start with the most uncomfortable data point. The World Bank's "rapid adoption" recommendation presumes a baseline of digital infrastructure that most developing economies simply do not have. According to the International Telecommunication Union's 2024 figures, internet penetration in low-income countries sits at roughly 36 percent. Sub-Saharan Africa's electricity access rate hovers below 50 percent, and in rural areas it drops under half that. A generative AI assistant delivered over a mobile network requires four things to function: reliable electricity, consistent bandwidth, a device, and a payment rail for whatever service is being consumed. In most low-income countries, at least one of these four rails is missing. In many, three are missing simultaneously. The report's language of "rapid adoption" treats these constraints as footnotes rather than preconditions. That is not a policy omission. It is a structural blind spot.
I have seen this exact pattern before. In 2018, while auditing Harvest Finance's alpha, the team pitch was about democratic access to yield. The social layer in Bondi Beach was electric — developers genuinely believed they were building a tool that would let anyone farm the same returns as the largest funds. The technical layer had a hole. My mathematical review identified a critical re-entrancy vulnerability in the harvest logic that would have allowed a carefully sequenced attack to drain the pool. The charm was real. The vulnerability was real. Both existed in the same protocol, and the team merged my patch only after two weeks of internal debate. The same misalignment appears in the World Bank's AI thesis. The social layer — governments committing to AI adoption, donors announcing AI-for-development programs — is real and growing. The technical layer — electricity, bandwidth, data governance, payment infrastructure — contains holes.
The World Bank knows this. Its own project portfolio is full of electricity, connectivity, and digital government programs. The strategic question is whether the AI adoption recommendation is conditioned on those infrastructure programs or decoupled from them. The plain reading of the January 2025 report is decoupled. It says "rapidly adopt AI" as a stand-alone prescription. It does not say "rapidly adopt AI conditional on reaching thresholds of power reliability and network capacity." This is the yield farm playbook: advertise the APY, omit the liquidity depth.
There is a technical nuance that partially mitigates the hardware problem. Generative AI's inference work happens in centralized data centers, not on the user's device. A farmer in rural Tanzania needs only to send a text query to a local gateway and receive a text response. No local GPU. No local model. No data center in the village. The "thin client" architecture of cloud AI means the device barrier is genuinely lower than in previous computing paradigms. But this creates a different problem — a network and cost barrier. Every API call costs money. Every token processed by a commercial model carries a fee. Every megabyte of data sent across international borders carries transit costs. The World Bank's model of "rapid adoption" implies these costs will be absorbed by someone — a government, an aid program, a corporate sponsor — but the report does not identify who pays the recurring bill.
Let me make this concrete. A mid-tier commercial AI API might charge on the order of several dollars per million input tokens and a multiple of that per million output tokens. For a government deploying a citizen-facing chatbot that executes ten million queries per month, the inference bill alone could range from fifty thousand to several hundred thousand dollars annually, depending on response length and model tier. That is a significant recurring operational cost for a ministry in a country with a GDP per capita below two thousand dollars. It is not a capital cost that can be financed and amortized. It is a perpetual operational expenditure that appears every month, forever, as long as the service runs.
In the crypto world, we call this an infinite mint. The token keeps inflating, the fees keep accruing, and someone has to validate that the costs are sustainable. Gas fees are the conscience of the system — they force every participant to confront the actual price of their use. The World Bank's recommendation skips that conscience. It presents AI as a growth accelerant without pricing the perpetual operational burden.
There is a deeper structural dimension of the infrastructure ledger, and it is geographic. Of the approximately eight hundred hyperscale data centers operating globally, Africa hosts less than 2 percent. South Asia and Southeast Asia are building capacity, but the asymmetry between where AI compute lives and where AI adoption is being prescribed is stark. When a Ghanaian government agency sends data to an AI API, that data physically crosses borders. It lands on servers in foreign jurisdictions. It becomes accessible under foreign legal frameworks. It can be disclosed to foreign intelligence agencies under foreign law. The "adoption" is not neutral infrastructure. It is a permanent data relationship with external parties, formed on terms set by those external parties.
Every block hides a confession, and the confession here is that the World Bank's rapid adoption framework is also a rapid data export framework.
The Dependency Arbitrage
This is where the on-chain analogy becomes exact. In DeFi, we have a well-documented phenomenon called the impermanent loss trap. A liquidity provider deposits assets into a pool. The pool generates fees. But if the relative price of the two deposited assets shifts dramatically, the LP's position becomes worth less than if they had simply held the assets. The fee income looks like a return. Underneath, the position is eroding. Developing economies adopting foreign-owned AI infrastructure are entering an analogous position.
The economics flow like this. Local data — agricultural yields, disease patterns, consumer behavior, government records, linguistic corpora — is the raw material. That data is sent to a foreign AI provider. The provider refines it into model improvements and delivers intelligence back. The intelligence is then leased to the local economy as a service, at a recurring fee. The more data a country exports, the more dependent it becomes on the service provider. The provider accumulates proprietary knowledge about the local economy — its crops, its disease ecology, its trade patterns, its citizens' behavior. The country accumulates a bill and a dependency.
This structure has a name in academic literature: data colonialism. The framework describes a system in which raw data is extracted from a periphery, processed at a core, and sold back to the periphery at a premium. The World Bank's own report acknowledges the dependence risk in a single clause. It does not quantify the structural imbalance. It does not analyze the arbitrage. It treats "foreign technology dependence" as a risk to be managed — like a market risk disclosure in a prospectus — rather than as an economic relationship with a compounding tax embedded in it.
Let me draw the parallel to the stablecoin reserve problem. Tether's USDT dominates roughly 70 percent of the stablecoin market, yet Tether's reserves have never received an independent, fully transparent audit. The entire industry pretends this is acceptable. The World Bank's AI adoption framework has a similar gap: it does not identify the reserve assets that back its promise. The "reserve" behind the promise of AI-driven growth is the institutional quality and absorptive capacity of the adopting country, and those reserves are dangerously undercollateralized in most of the Global South.
To make this concrete, I will turn to a case I analyzed in depth. During the 2022 Terra Luna collapse, I conducted a post-mortem of the UST/LUNA arbitrage loop, calculating the exact liquidity depth required to sustain the peg under varying withdrawal pressure. The mechanics were elegant and catastrophic. UST's stability was maintained by an arbitrage relationship with LUNA. When UST drifted below one dollar, users could mint LUNA and sell it to buy UST, creating buy pressure. The loop was mathematically sound under conditions of confidence and mathematically suicidal under conditions of panic. When the anchor slipped, the loop inverted — it amplified the collapse instead of stabilizing it. My calculation showed that the system would need tens of billions of dollars of committed liquidity to survive a bank-run scenario. The Terra team had less than half of that available, and much of it was encumbered. The project was insolvent from the moment confidence cracked.
The World Bank's AI dependence structure has a similar inverted-loop risk. Consider a scenario where a dominant AI provider raises API prices by 50 percent, or alters its terms of service to prohibit government use, or restricts access due to geopolitical tensions. A country that has built its digital government infrastructure on that provider now faces a systemic interruption. The adoption loop that was designed to accelerate growth becomes an amplification loop for dependency shocks. The worse the dependence, the harder the fall.
Data sovereignty, in this context, is the equivalent of self-custody. The crypto community learned this lesson the hard way across 2022 and 2023. FTX customers learned what happens when you do not control your private keys. When you delegate custody to a trusted third party, you delegate the ability to audit, to exit, and to protect your assets in times of stress. Developing economies that adopt foreign AI services without building domestic data governance are handing over their private keys in exchange for high-yield exposure to a third party's infrastructure. The yield is real. The custody risk is real. Both exist simultaneously, and the market has never found a way to have one without the other.
Liquidity flows, but integrity stagnates. The data will continue to flow outward from the Global South to the AI cores in California, Beijing, and Seattle. The integrity of the arrangement — who owns the data, who controls the model, who sets the terms, who can exit — will stagnate until someone forces the issue.
The Absorptive Capacity Gap
The third structural problem is the human layer. AI adoption does not happen in a vacuum. It requires institutional capacity — people who understand what AI can do, who can integrate it into bureaucratic workflows, who can evaluate vendor claims, and who can manage failures when they occur. The data here is sobering. Stanford's AI Index 2024 reports that only about 10 percent of African countries have national AI strategies. Across the developing world, only a handful of governments have meaningful institutional structures dedicated to AI governance, model evaluation, or procurement oversight. The countries being urged to adopt rapidly are the countries least equipped to govern the adoption.
This is the absorptive capacity problem. And I have watched it play out in crypto with painful precision. In 2021, I embedded with the Bored Ape Yacht Club community — not as a status seeker, but as an analyst tracking royalty enforcement mechanics. ERC-721, the token standard that governed most NFT projects, has no native mechanism for royalty enforcement. Creator royalties were an expectation, not a protocol guarantee. When competitive marketplaces emerged, they bypassed the voluntary royalty schemes that OpenSea had enforced out of goodwill. I documented on-chain that roughly 40 percent of secondary sales were already bypassing creator fees before the enforcement collapse became public. The community had adopted a technology whose implicit promise — "creators earn royalties forever" — had no technical foundation. Social enthusiasm cannot fix a structural gap. The code didn't. The community learned the hard way.
The same dynamic applies to AI adoption in institutional environments. A finance ministry can issue a policy document saying "we will leverage AI." Without staff who understand how to evaluate AI tools, negotiate procurement contracts, maintain data pipelines, or audit model outputs, the document is a social signal with no technical foundation.
The World Bank's recommendation effectively assumes that human capital can be developed simultaneously with adoption — that learning by doing will compensate for the absence of prior expertise. This is a reasonable theory in some contexts. South Korea and Taiwan in the 1970s and 1980s demonstrated that rapid technology adoption combined with aggressive human capital investment can produce transformative growth. But those cases had preconditions: strong state capacity, established industrial ecosystems, mass educational attainment, and deliberate protection of domestic institutions during the learning phase. Many of the countries the World Bank is now urging to adopt AI rapidly possess none of these preconditions.
The brain drain issue compounds the problem. Even in countries that invest in AI education, the most talented graduates tend to emigrate to developed economies where salaries are ten to fifty times higher. A country can train brilliant AI engineers and still end up with zero domestic capacity if those engineers all end up working at OpenAI, DeepMind, or Google. The absorptive capacity gap is not just a training problem. It is a retention problem, and the global market for AI talent is remorselessly extracting human capital from the same economies that need it most.
We chased the glow, not the ledger. The glow was the promise of AI-driven development leapfrogging every barrier. The ledger, in this case, is the pipeline of trained talent that flows from the Global South into the global AI centers. The ledger is now reciting the cost.
The Vendor Capture Model
Now for the uncomfortable commercial layer. If the World Bank's recommendation achieves its stated goal — rapid AI adoption across the Global South — the primary beneficiaries in dollar terms will not be the Global South. They will be the multinational cloud providers and AI model vendors that dominate the global market. This is not a conspiracy. It is a balance sheet observation.
The global cloud infrastructure market is concentrated among a handful of players. AWS, Azure, and Google Cloud control roughly two-thirds of the global market. In emerging markets, Alibaba Cloud and Huawei Cloud add a significant counterweight, particularly in Asia and Africa. The AI model tier is similarly concentrated: OpenAI, Anthropic, Google, Meta via Llama, Alibaba via Qwen, Mistral from Europe. Developing economies are, almost without exception, net consumers of this stack rather than net producers. The World Bank's recommendation is, in effect, a policy endorsement for expanding the market footprint of these providers into the Global South, with the World Bank's institutional legitimacy attached to the transaction.
Let me connect the dots to a historical precedent. In the 2000s, the global development community strongly promoted digital financial inclusion. Mobile money, agent banking, digital payment infrastructure — these received billions in development finance and technical assistance. The program was a genuine success on access metrics. Tens of millions of people gained access to financial services for the first time. However, the side effect was the repositioning of a small set of global mobile network operators and technology vendors as gatekeepers of financial infrastructure across large parts of Africa and South Asia. They gained scale, data, and durable revenue pools. The development community gained an access metric. The ledger in between was mixed, and the value capture was asymmetric.
AI adoption in the Global South will produce a similar mixed ledger. The local benefits — faster administrative services, improved agricultural yields, better health diagnostics — are real and valuable. But the long-term value capture will flow through the balance sheets of foreign infrastructure providers. The pricing power sits with the model providers. The data sits with the model providers. The switching costs accumulate quietly each year. When the World Bank says "rapid adoption," it is simultaneously saying "accelerate the market expansion of global AI infrastructure providers into the Global South." The recommendation and the expansion are the same transaction.
I saw this dynamic from the institutional side in 2024. I was invited to consult for a major Australian bank considering Bitcoin ETF exposure. The engagement was prestigious — high-profile meetings, serious people, serious fees. My role was to stress-test their risk models against on-chain liquidity crises. I produced a fifty-page report demonstrating that the true systemic vulnerability was not Bitcoin itself but the custodial intermediaries — the same concentration risk that killed Mt. Gox in 2014 and FTX in 2022. The bank's leadership genuinely believed they were adopting a decentralized asset. The structure they were actually adopting was concentrated custodial counterparty risk registered on a national exchange's settlement infrastructure. The framing had disconnected from the reality.
The same disconnect now applies to AI adoption. The policy framing celebrates "empowerment," "inclusion," and "leapfrogging." The structural reality is "vendor onboarding." The country is not adopting a technology. It is adopting a relationship with a dominant external provider. The names on the contracts are different. The structure is the same one that left the Philippines bare to labor outsourcing and its exposure to call center automation, or that left Argentina to the mercy of IMF conditionality. Adoption creates dependence. Dependence creates a rent position. The rent position transfers value abroad, forever, as long as the service continues.
The consulting industry has already noticed the opportunity. The "AI for Development" sector — strategy consultancies, technical assistance providers, NGO intermediaries, and evaluation firms — is assembling service packages to help governments "deploy AI responsibly." This is the same cycle that produced the "microfinance for development" industry, the "blockchain for development" interlude, and the "digital identity for development" push. Institutions attach policy legitimacy, consultants attach service offerings, governments attach donor money, and a new rent envelope is created. The World Bank's recommendation is the enabling instrument for a new generation of consulting revenue built around the phrase "rapid adoption."
What the Crypto Parallel Teaches
The crypto industry has run the "rapid adoption" playbook more than a dozen times. The common thread: a technology promise gets translated into a narrative, capital follows the narrative, infrastructure lags behind, and the gap between narrative and infrastructure eventually gets priced as a loss. The NFT royalty case is the cleanest example — a promise embedded in social expectations, absent from the code, and ruthlessly arbitraged once marketplaces discovered they could capture volume by ignoring creator fees. The same structure appears in the World Bank's AI adoption recommendation. It promises a productivity dividend. It does not specify the distribution mechanism. It does not identify the governance structure that ensures local value capture. It says "rapid adoption" and leaves the implementation to governments that are structurally weaker than the foreign technology providers they are about to depend upon.
But there is an even deeper structural lesson, and it comes from the world of cross-chain interoperability — a space I have watched with a cold eye for years. The crypto industry's answer to liquidity fragmentation was to invent more bridges and more interoperability protocols. Each new bridge created new attack surfaces and new vulnerabilities. The theoretical promise was that interoperability would unify liquidity. The empirical reality was that every new chain and every new bridge magnified fragmentation and complexity. There are now dozens of failed bridge projects, each one a monument to the idea that adding connecting layers solves a fragmentation problem that the layers themselves are creating.
The World Bank's AI adoption framework is building the same kind of bridge. It is proposing a connecting layer — policy endorsement, technical assistance, loan covenants — to cross the gap between developing economies and advanced AI infrastructure. The gap is real. The bridge is real. But the bridge does not eliminate the gap. It reorganizes it. The gap becomes a dependency corridor through which data flows one way and fees flow the other.
Let me also say something about Bitcoin, since it frames my thinking about what infrastructure is for. BRC-20 tokens and Runes on Bitcoin are like using a Rolls-Royce to haul cargo. It insults the car, and it does not carry much. Bitcoin is settlement infrastructure. It was designed to do one thing and to do it securely. Layering speculative token standards onto it is a misuse of its properties. The World Bank's AI recommendation risks the same category error — using a transformative, globally important technology as a speculative growth narrative for economies that lack the underlying infrastructure to absorb it productively.
The overcollateralization problem applies to policy as much as it applies to DeFi. In decentralized finance, a stablecoin is only as stable as its reserve quality. The World Bank's "rapid growth through AI" thesis is, at its core, a financial claim. It claims that AI adoption will produce growth in excess of the costs of adoption. That claim has a collateral requirement. The collateral required is institutional quality — the capacity to plan, implement, monitor, and correct AI deployments. That collateral is in chronically short supply in the places being told to adopt most rapidly. The recommendation is, in undercollateralized form, a risky position on behalf of the world's most economically fragile nations.
History is written in hex, not headlines. The headlines from January 2025 will say the World Bank is pushing AI for growth. The hex — the underlying data — will say something more complex. It will say that infrastructure deficits were not addressed, that data governance was not secured, that absorptive capacity was not funded. The blockchain remembers everything. The policy process, far less so.
What the AI Adoption Bulls Get Right
Now for the uncomfortable counter-rotation. I have spent this article dissecting the World Bank's recommendation. The dissection is warranted, but it does not render the recommendation worthless. To miss the legitimate upside would be to repeat the very error I criticize — confusing narrative flaws with invalid outcomes.
First, mobile-first leapfrogging is real. Global smartphone penetration has crossed 60 percent even in the lowest-income markets. The device layer largely exists. The cloud inference model of AI — where processing happens remotely — genuinely reduces the hardware barrier relative to earlier computing paradigms. An AI assistant accessed via SMS requires no local compute, no desktop hardware, no IT department. The infrastructure constraints I have described are real, but they are not absolute, and mobile networks have a demonstrated capacity to extend reach faster than fixed-line infrastructure ever did.
Second, open-source AI models create a genuine alternative path. Meta's Llama, Alibaba's Qwen, and the broader open-weights ecosystem have dramatically reduced the cost of deploying capable AI in domestic environments. If a government invests in local hosting of open-weights models, the data export problem shrinks, and the dependence profile improves. This is the most underappreciated element of the entire discourse. Open-source AI is the only viable path that allows developing economies to capture the upside of adoption while mitigating the dependency risk. The World Bank's implicit openness to this path is a meaningful positive — even if the report itself does not emphasize it.
Third, the World Bank's endorsement does allocate real resources. When the World Bank signals a policy priority, loan programs, technical assistance, and private capital alignment tend to follow within 12 to 36 months. The "AI readiness" framework is already filtering into project design in Southeast Asia and Africa. The timeline risk is real, but the direction of travel is real too. For specific sectors — agriculture, health, customs administration — the combination of World Bank financing and already-deployable open-source tools could produce meaningful productivity gains within a few years.
Fourth, the productivity potential is large enough to matter even if partially captured. The World Bank's case does not need AI to deliver a 2 percent growth lift across the developing world to justify adoption. It needs a fraction of that to transform a thin market or a weak public service. A customs digital assistant that reduces processing times from 14 days to 3 days creates measurable welfare gains even if a foreign provider captures 60 percent of the economic value. The remaining 40 percent is still a real gain. The capture structure is exploitative. The gain is not fake. In development economics, partial capture beats total absence.
These observations matter because they complicate the critique. The World Bank's recommendation is wrong insofar as it treats adoption as a substitute for infrastructure, governance, and human capital. It is right insofar as it recognizes that waiting for perfect conditions would be its own form of paralysis. The task is not to dismiss rapid AI adoption. The task is to build the governance rails, the data sovereignty rules, and the domestic capability investments into the recommendation so that adoption does not become capture, and gain does not become dependency.
The Accountability Call
The World Bank has minted a narrative token. Its utility now depends on whether the institution backs the narrative with resources — dedicated financing windows, technical assistance programs, governance standards, and measurable adoption conditions — or whether the narrative evaporates into policy ether, leaving developing economies with a new layer of dependence and a recurring bill.
There is a simple metric to watch. Did the World Bank create a dedicated AI adoption financing facility within twelve months of the January 2025 report? Did it tie loan covenants to data governance and open-source adoption requirements? Did it fund absorptive capacity — the human capital, the evaluation frameworks, the procurement expertise — rather than just the cloud service contracts? The distinction between adoption-as-empowerment and adoption-as-capture will be visible in those choices.
Gas fees were the only truth we paid for in DeFi. Every transaction, every yield, every liquidity position had a cost, and the cost shaped behavior. In the AI-for-development era, the truth will be in the project evaluations, the loan covenants, and the systems that either hold or break under stress. The growth narrative is a promise. The ledger is the code. The code didn't lie in 2018. It didn't lie in 2022. It will not lie now.
The question is not whether AI can transform developing economies. It can. The question is whether the transformation will be owned by the people who are meant to benefit from it — or by the platforms that are now being handed the keys to the Global South's most valuable resource: its data, its attention, and its future.
Which side of the trade are you on?