BLOG POST

Developing Countries Can’t Afford to Buy Their Way Into the AI Race. They Can’t Afford to Sit It Out Either.

With public debt already elevated across much of the developing world, governments have limited fiscal space for costly AI industrial policies, whose full budgetary implications are often poorly measured. Yet an important opportunity remains within reach: adopting AI technologies by improving governance and creating an environment in which businesses can put them to productive use.

In May 2026, a $1 billion data center that Microsoft and the Abu Dhabi AI firm G42 had announced for Kenya stalled. The sticking point was not the technology. The companies wanted the government to guarantee purchases of computing capacity every year, and a Treasury that already spends roughly a quarter of its revenue on interest could not commit at the level requested. At the planned full scale of one gigawatt, President Ruto noted, switching the facility on would also mean shutting off power to half the country.

Finance ministers in developing countries increasingly face proposals to finance national large language models, sovereign data centers, or dedicated GPU capacity, often with financing offered by the same vendors or bilateral partners promoting the projects. These proposals can impose substantial costs on already constrained budgets. They also tend to present participation in the AI race as a choice between building domestic capacity and falling behind.

Kenya’s experience illustrates why affordability must be central to the decision. It also suggests that governments should question the premise behind many of these proposals: success in AI need not depend on building a national frontier capability.

The contest is adoption, not the frontier

In a new paper published in the Asian Economic Policy Review, Hao Chen and Meg Rithmire argue that the fixation on frontier models has obscured what China—one of the two AI superpowers—is actually doing: a state-directed effort to embed AI as a general-purpose technology across the physical economy, from manufacturing and supply chains to industrial robotics. By 2025, downstream applications made up 42 percent of the generative AI services registered in China, up from none two years earlier. The authors conclude: “The decisive contest may be less about which country achieves artificial general intelligence first and more about which political economy can more rapidly diffuse AI into productive activity.”

They are not alone in this reading: in Technology and the Rise of Great Powers, Jeffrey Ding argues that great powers have historically pulled ahead less by inventing general-purpose technologies than by spreading them across their economies.

For development policy, the distinction between inventing AI technologies and applying them is fundamental. Frontier innovation requires enormous capital expenditures and specialized infrastructure. Widespread adoption depends more on the capacity of firms and institutions to absorb new technologies, on regulatory predictability, and on access to computing power that can often be rented rather than owned. Open-weight models can further reduce entry costs, although running them still requires hardware. These are different policy challenges, and developing countries should allocate scarce resources accordingly.

Why “sovereign AI” is the wrong answer at the highest price

Consider first the economics of the underlying assets. A GPU cluster has a useful life of roughly five years, and older chips generally consume more electricity per unit of output than newer ones. This matters particularly in economies where electricity is scarce. Sovereign borrowing to finance a data center can create liabilities extending beyond the useful life of its equipment. Guarantees and offtake contracts can create contingent liabilities that are not fully reflected in headline debt statistics. Both exposures deserve scrutiny in countries already facing debt sustainability concerns.

Unless national security imperatives dictate a fully sovereign capability, the legitimate goals behind the pitch—models that work in local languages, control over sensitive data—could be met through pooled or regional arrangements at a fraction of the cost. The obstacle is political rather than technical, since most AI already runs on shared hyperscale infrastructure. Existing regional blocs such as ASEAN, Mercosur, or the African Union are the natural place to start, despite their internal frictions.

The fiscal context makes these choices especially consequential. UNCTAD estimates that 3.4 billion people live in countries spending more on interest than on health or education. New commitments to AI infrastructure therefore compete with essential public services and debt service, including where the eventual fiscal cost is not yet fully recorded.

AI policy reaches the public balance sheet through at least five channels:

  • Direct capital spending. Brazil is putting R$2.3 billion (about $440 million) from its national science and technology fund into two AI supercomputing projects, one built with Huawei and iFlytek, the other through an open tender. This is the most transparent channel.
  • Guarantees and offtake commitments. What Microsoft and G42 asked of Kenya: a promise to buy capacity whether or not it is needed. It looks like procurement; economically it is debt, and it would typically sit outside the debt sustainability analysis until called.
  • Tax incentives. Malaysia’s data center boom in Johor rests partly on tax incentives. The revenue forgone is real; the domestic payoff less so. According to AMRO, a typical facility employs 30 to 50 people and imports most of its equipment.
  • Quasi-fiscal costs. Grid upgrades and discounted power tariffs are carried by state-owned utilities, and they become fiscal when the utility needs a bailout. In 2025, Pakistan, under an IMF program, proposed allocating 2,000 megawatts of surplus power to bitcoin mining and AI data centers at discounted rates; the IMF raised concerns about the cost and the strain on an already fragile grid.
  • Public equity. Stakes in national AI champions or state-backed funds are assets on paper, and they may turn into liabilities when failure triggers a bailout.

The deeper irony is that even China keeps these costs off the budget. Chen and Rithmire are explicit that China’s investor-state “is enabled by financial repression.” Household savings run at 25 to 35 percent of GDP, deposits sit in state-controlled banks at low rates, and capital controls keep them onshore. State investment therefore runs “through the financial rather than fiscal channel… and with no public scrutiny over the allocation of resources and therefore the costs.”

This financing structure matters. China can channel domestic savings into strategic investments through state-controlled financial institutions and capital controls, thereby shifting much of the cost outside the central government budget. Countries with open capital accounts, fragile currencies, and market-priced borrowing do not generally have access to comparable financing arrangements. Their AI ambitions are more likely to translate directly into fiscal expenditures, sovereign borrowing, or contingent liabilities.

Nor is the Chinese model as clean as the headline suggests. The same paper documents that China’s broader state-led industrial policy has “generated inefficiency, even tremendous waste”: hundreds of local electric vehicle firms and chip fabrication plants in dozens of cities. China can absorb that. A country with a tax take of 15 percent of GDP and an IMF-supported program acting as a fiscal straitjacket cannot place bets on that scale and write off the losses.

The much-less-expensive alternative

Public financing nevertheless has a potentially useful role when it mobilizes additional investment rather than replacing it. Chen and Rithmire report that Shanghai’s AI “mother fund” mobilized close to RMB 50 billion of “social capital” (much of it from other state-linked investors) against RMB 7.8 billion in subscribed commitments within eight months, implying a leverage multiple of up to 5.8. The relevant lesson for fiscally constrained governments is the potential to crowd in financing while limiting direct public exposure.

India provides another approach that may be better suited to countries facing binding budget constraints. Under the IndiaAI Mission, with a budget of about $1.25 billion over five years, private providers selected through competitive tenders own and operate more than 38,000 GPUs. The government instead subsidizes the hourly cost paid by startups and researchers. This leaves utilization risk with operators and avoids placing rapidly depreciating equipment on the sovereign balance sheet. Although disbursement has lagged announcements, the underlying financing model merits attention.

Industrial applications run on governance, and governance is cheap

A quarter century ago, Albert Martinez and I hypothesized that industrial growth is institutional-quality-intensive: the more a production process depends on suppliers, credit, and new technology, the more it depends on contract enforcement, low corruption, and a predictable bureaucracy. We found a strong link between the quality of institutions and industrial growth, working through both the level of investment and the efficiency with which it is used. Putting AI to work is the extreme case of that dependence. Where these institutions are weak, investors do not simply invest less; they ask the state to underwrite the risk, turning a governance gap into a contingent liability.

The policy agenda need not be expensive. Four institutional and regulatory priorities could materially improve the environment for AI adoption, without requiring large public capital expenditures.

  • Data governance. Firms cannot deploy what they cannot lawfully use. In most developing economies, unresolved rules on data access, sharing, and cross-border transfer bind harder on adoption than compute does.
  • Procurement. The state is typically the largest early adopter in the economy. It can create a first market for domestic firms by buying differently rather than by buying more.
  • Regulatory clarity. A rulebook that permits AI use in health, finance, logistics, and customs, rather than freezing it pending a framework that never arrives.
  • Skills and retention. Including treating the diaspora as a channel rather than as a leak.

Where public money is used at all, use it as leverage, on the Indian or Shanghai model, and cap the contingent exposure before signing rather than after.

Adopt before you build

There is also a potentially temporary opportunity for countries that choose adoption over domestic frontier production. The excess capacity described by Chen and Rithmire has put downward pressure on prices for industrial robots and AI-enabled capital goods. From the perspective of importing economies, these lower prices can facilitate technology diffusion. The authors nevertheless caution that prices “could push higher again, only after much of the world has lost the ability to compete.”

Developing countries should therefore consider accelerating AI adoption in sectors where they already possess a comparative advantage, including by integrating domestic firms into international AI-related value chains. This approach has precedents in other technology-intensive industries. Malaysia has no leading-edge chip fabrication plant, yet it accounts for about 13 percent of global chip assembly, testing, and packaging, activities essential to AI accelerators. Vietnam became a major electronics exporter largely by assembling imported components, which represent around 60 percent of the sector’s export value. Both cases demonstrate that participation in advanced technology industries does not require domestic ownership of the technological frontier.

A test for finance ministers

Finance ministries should apply a consistent fiscal test to proposals for national AI infrastructure. Does the project involve a sovereign guarantee or take-or-pay commitment? Is the exposure captured in the debt sustainability analysis, including any foreign-exchange or maturity mismatch? Could the same development objective be achieved by supporting access to computing services rather than owning the underlying infrastructure? Where these questions are not satisfactorily answered, an industrial policy initiative may create debt vulnerabilities disproportionate to its expected benefits.

For most developing countries, competing directly at the frontier of AI development would require resources that are difficult to justify against pressing fiscal and development needs. The alternative is not to forgo AI, but to focus public policy on the institutions and incentives that make its adoption possible.

The principal fiscal danger is committing scarce borrowed resources to rapidly depreciating infrastructure before establishing whether the investment will generate commensurate economic benefits. By the time those costs become fully apparent, the technology may already be obsolete.

Claude Opus 5.5 by Anthropic was used in drafting this piece. The author reviewed, revised, and takes full responsibility for the content.

DISCLAIMER & PERMISSIONS

CGD's publications reflect the views of the authors, drawing on prior research and experience in their areas of expertise. CGD is a nonpartisan, independent organization and does not take institutional positions. You may use and disseminate CGD's publications under these conditions.


Thumbnail image by: Tom Perry / World Bank