On building and funding public AI, governments must be tactical, but remain flexible

Questions about the role of the state are central. In principle, DPI is not only about government: It emphasises public interest rather than public ownership, and its goal is to create more competitive markets. In practice, the state plays a decisive role, whether through policy or funding.

The apparent tension in a framework that emphasises markets yet relies on the state is particularly acute when it comes to public investments. AI is capital intensive, and it emerges at a moment of global fiscal constraint. How much should taxpayers underwrite the development of public AI systems, and how should policymakers weigh the risks of crowding out private investment against the benefits of building shared capacity? As governments around the world consider alternatives to proprietary AI, does DPI offer guidance on how to finance these systems?

Our answer, in brief, is that DPI does offer a useful framework for thinking about public investment in AI — but only to a point. We encourage governments, including India’s, to draw seriously on the lessons of DPI, but we caution against treating DPI as a comprehensive blueprint or checklist. Another pitfall is for investment to focus exclusively on fostering AI innovation; often, it is the rate of diffusion that will ultimately determine AI’s impact.

Much of the conversation around DPI emphasises its public dimension. But when considering its lessons for AI, the infrastructural component is equally important. At its core, the DPI argument is that some technologies are foundational and cross-cutting: They support a wide range of downstream uses and function as neutral, open platforms for innovation. Analogies are drawn to roads and bridges; the World Bank describes DPI as “digital plumbing”.

Many elements of AI do lend themselves to this infrastructural interpretation — but that does not necessarily make them all suitable candidates for public investment. We are particularly sceptical of two categories often favoured by governments: Large frontier language models and data centres.

Governments typically justify investments in domestic compute on grounds of strategic autonomy or digital sovereignty. The impulse is understandable, especially in an unstable geopolitical environment, but the actual autonomy granted is illusory. The AI stack is complex and multi-layered; control over one element — compute — does not translate automatically into control over downstream components like models, data, or applications. Moreover, demand for inference is variable and unpredictable, making the quest to secure autonomy through domestic compute at best partial, prone to both underutilisation and shortages.

The quest for sovereignty through public funding of domestic models is similarly weak — and increasingly unnecessary. Today’s AI landscape is characterised by a paradox in which frontier models are both prohibitively expensive and ever-more commoditised. Governments now have access to a growing ecosystem of open-source alternatives whose performance often trails the frontier by mere months. The capabilities offered by such models are more than sufficient for most public-interest applications. Instead of building models, governments would be better served by adapting open-source models to local conditions.

If foundation models and data centres don’t make sense, then where should governments allocate their resources? We categorise our responses into aspects of AI that are infrastructural (i.e., that fit within the DPI paradigm), and those that are not.

When it comes to AI-as-infrastructure, one particularly promising avenue is high-quality datasets. India’s recently released AI governance guidelines already emphasise the importance of curating “high-quality and representative datasets,” some of which could be drawn from DPI systems (e.g., by repurposing payment data). We agree with this orientation and argue that governments are uniquely positioned to support such initiatives, with one important caveat: Data reuse must be accompanied by robust governance to protect privacy and other rights (a point acknowledged by the AI guidelines).

Governments should also consider altogether new forms of infrastructure. Just as the Global South pioneered lighter, lower-cost forms of physical infrastructure — for instance, bus rapid transit systems — there are opportunities to rethink foundational AI. General-purpose translation modules offer an example. Initiatives like Bhashini in India and Masakhane in Africa are not end-user applications but shared layers that provide reusable linguistic capabilities. Other candidates include open speech-to-text systems for low-resource languages, or evaluation benchmarks that facilitate more context-sensitive, locally-aligned models.

Infrastructural AI investments have the advantage of being high-leverage; they support a range of downstream innovation. But there are also cases where more targeted public funding of specific applications makes sense, typically because of various forms of market failure. Examples include India’s Kisan e-Mitra chatbot for agriculture and Health Sentinel for disease surveillance. Even if the private sector does build such tools, they may be optimised for cost or profitability rather than public-interest goals like equity and accessibility.

A similar logic applies to AI used within government. Internal tools — copilots for officials, AI-based training simulations — can yield large public returns but rarely drive private investment. These tools depend more on institutional and administrative fit than technical sophistication; public funding must extend beyond mere procurement to training and organisational integration.

The spread of DPI has been rapid and impressive. Delegates from around the world will be looking to glean lessons from this. Our message to them is clear. Infrastructure-based investments offer the highest-leverage, especially when they embrace lighter, more frugal forms. But governments must also step outside the DPI framework to make targeted interventions addressing public-interest goals private actors won’t pursue.

In short, we are infrastructure first, but not infrastructure only. When it comes to building and funding public AI, governments must be tactical, but also flexible.

Narayanan is professor of computer science at Princeton University and director of its Centre for Information Technology Policy. Kapur is a Visiting Fellow, Princeton’s Chadha Centre for Global India and Senior Fellow, New America