AI for Sustainable Development: When Bigger Is Not Better

AI is already being tested to improve efficiency in tax administration, health care, legal services, agriculture, and public programs. A recent World Bank survey of governments in 60 economies found that 44 percent of reported internal government use consisted of individual officials using AI for discrete tasks such as searching for information or summarizing documents. These uses may save time and improve services. They also illustrate that many useful applications do not require the largest or most resource-intensive AI systems.

Consider electricity forecasting. AI can help grid operators anticipate demand and variable renewable generation, but better forecasts matter only if utilities have the transmission, storage, operational, and institutional capacity to act on them. The same principle applies elsewhere. An agricultural advisory tool must reflect local crops and conditions and reach farmers in a form they can use, while an early-warning system needs communications networks and public agencies capable of reaching people before a disaster. 

AI is one part of each system. It cannot substitute for the infrastructure, data, skills, and institutions needed to deliver results.

The World Development Report 2026 distinguishes among three pathways: adopting existing AI tools, adapting them to local conditions, or advancing frontier models. Each requires a different level of investment. Adoption and adaptation can often draw on existing models and more limited computing capacity, while developing frontier models requires much larger investments in energy, data, specialized skills, and computing infrastructure. Countries therefore do not need to pursue the same infrastructure strategy to benefit from AI. 

Investment patterns make these choices increasingly consequential. The 2026 World Investment Report from the United Nations Conference on Trade and Development finds that announced investment in new AI infrastructure projects grew by an average of 47 percent a year between 2020 and 2025, reaching nearly $350 billion in 2025. Yet in many developing regions, investment in data centers is rising while investment in basic connectivity has stagnated or fallen. The challenge is therefore not simply to attract more computing capacity but to ensure that digital investment strengthens the energy, connectivity, and institutional systems needed to deliver broader development benefits. 

Progress in AI is too often equated with larger models and more computing capacity. But neither is, by itself, a measure of development. Many useful applications do not require the largest general-purpose models. Translation, document processing, agricultural advice, and some forms of forecasting can be handled by smaller systems designed for a specific task. Research published by UNESCO and University College London examined the energy used when already-trained models respond to requests. In the applications tested, smaller task-specific models used up to 90 percent less energy while maintaining strong performance. Training and adapting models also require energy and computing resources and should be assessed separately. 

However, smaller will not always be better. Some scientific, research, and climate applications will need considerable computing power. The principle is proportionality: Choose an approach that can do the job reliably without demanding more computing capacity, energy, or supporting infrastructure than the outcome warrants. Sometimes that will mean adapting an existing model, using a smaller model on a local device, or sharing computing capacity. Sometimes an ordinary digital tool will be enough.

Countries also have legitimate concerns about control over sensitive data, continuity of essential services, and dependence on external providers. Domestic computing capacity may therefore carry strategic value even when it is not the cheapest option. But a data center does not remove every dependency; it may still rely on imported chips, foreign platforms, and outside expertise.

Data centers can support research, services, and local innovation. They can also require public incentives and major investments in electricity, grids, land, and water. The World Investment Report notes that the surge in digital infrastructure investment is occurring as investment in renewable energy and other basic infrastructure has weakened, raising concerns about the competition for capital in constrained markets. From an energy transition perspective, the test is whether a facility brings additional clean power, strengthens rather than strains the grid, pays its fair share of infrastructure costs, and produces benefits that extend beyond the facility itself.

Before governments or development institutions commit public resources, four questions should travel together.

The test is not which option is smallest but which one is proportional to deliver the required result with the least pressure on energy, water, and public infrastructure over its lifetime. That means looking beyond a single efficiency rating to the resources needed to train or adapt the model, operate it at the expected scale, and provide the necessary hardware and supporting infrastructure.

AI can improve public services, widen access to expertise, and help manage more complex energy systems. Those possibilities deserve serious investment. But development value does not flow automatically from more computing capacity, and digital infrastructure does not become sustainable simply because it supports an AI application.

This is where an environmental and energy transition perspective is essential. The question is not only whether AI can help address climate change. It is also whether the infrastructure behind it supports clean, affordable, and resilient energy systems.

The international conversation is beginning to connect these issues. In June 2026, U.N. Secretary-General António Guterres proposed an AI Environmental Transparency Initiative, calling on major companies to disclose the carbon, water, and land footprints of their AI systems. At COP31, the proposed Antalya Artificial Intelligence Pledge can carry the discussion into how AI is designed, procured, powered, deployed, and measured. Meanwhile, the UNFCCC Technology Mechanism’s AI for Climate Action initiative is supporting practical applications for mitigation and adaptation, particularly in developing countries.

These initiatives should reinforce one another. Transparency can make resource demands visible. Climate commitments can shape how systems are powered and procured. Development policy can help determine which applications and infrastructure proposals merit investment or other forms of public support.

The decisions being made now will shape electricity systems, public budgets, and development pathways for years. We should therefore ask what kind of AI is appropriate, what infrastructure it requires, and whether the expected benefits justify the resources committed.

If AI is to serve sustainable development, the standard should be straightforward: useful outcomes, proportionate infrastructure, clean and affordable energy, and public benefits that justify public costs. Bigger may sometimes be necessary. But the measure of better is the outcome delivered, not the computing capacity deployed.