rarely translates into bargaining power without downstream processing capacity, which most developing countries lack. Discovered mineral reserves cannot always be extracted in a commercially viable way. Data can be another potential counterbalancing force. Many developing countries generate large volumes of data that remain relatively untapped. For the major powers, access to such data is attractive for model development and market access alike. However, data in many developing countries are highly fragmented13—scattered across informal systems and government agencies and divided among dozens or hundreds of language communities, often too small individually to justify building specialized AI models. The pursuit of AI sovereignty risks becoming a development trap Faced with concerns about overdependence, many governments have started to adopt strategies to reduce dependence by securing ownership or control of the key physical resources that power AI. Although these strategies may also serve economic objectives, they are often framed as efforts to achieve AI sovereignty. Measures to achieve this include developing local models and capabilities to manufacture semiconductors, subsidizing domestic computing infrastructure,14 and localizing data. From a national standpoint, AI sovereignty measures are rational if the objective is reduced dependence; they may also help countries move up the AI value chain. However, these measures often fail to deliver the promised sovereignty or a sustainable economic path for developing economies.15 First, AI sovereignty measures frequently fail to achieve their intended goals, especially if the goals involve complete self-sufficiency. Data localiza­ tion can be complicated by the extraterritorial reach of a cloud provider’s home-country laws, which may be triggered in situations such as 206 serious criminal investigations.16 Domestic semiconductor initiatives often struggle to match the scale economies of established incumbents. Even countries with a local data center may not be able to avoid upstream dependencies on hardware and software from major foreign providers.17 Second, AI sovereignty strategies commit governments to building at home, and this can generate domestic political costs. AI infrastructure is resource-intensive. As a result, AI sovereignty strategies bring to the forefront of political discourse concerns about the siting of such infrastructure, especially data centers.18 In Malaysia, the state of Johor, which is the fastest-growing data center hub in Southeast Asia, is facing community backlash over the rapid expansion of data centers.19 Similar opposition has emerged in other developing countries such as Brazil and Mexico.20 Unlike a commercial developer that can relocate to another country if blocked in one, a government that has staked a sovereignty claim on building data centers cannot. It faces the political liabilities of public backlash as well as absorbing the fiscal cost of the abandoned or delayed project—often for an unmet goal. Third, AI sovereignty imposes an aggregate cost through fragmentation. AI benefits from scale. Integrated global markets lower costs in many layers of the stack, such as those for training ­models, manufacturing chips, and building applications. Excessive fragmentation into duplicative ­country-level stacks reverses these scale e­ conomies. What emerges is a form of a self-reinforcing “fragmentation doom loop,”21 which can create a development trap. Efforts to duplicate resources are already prohibitively expensive. Yet the more countries fragment, the more expensive it is for them to build their own AI stacks, adding strain to alreadyweak fiscal balances and diverting resources from pressing needs in health, education, or other ­infrastructure. Smaller developing countries bear a greater burden because they are unable to replicate at the national level the scale economies that they World Development Report 2026