E-Paper
AI will not scale globally without new capacity infrastructure
This article is authored by Amandeep Singh Gill and Vilas Dhar.
Artificial Intelligence (AI) is advancing at extraordinary speed, but its economic benefits are not spreading in the same way. The limiting factor is no longer talent, ideas, or even access to models. It is infrastructure: The systems that allow AI to move from research labs into economies and institutions across borders.
A machine learning engineer in Lagos and her counterpart in London may be working on the same agricultural forecasting problem for West Africa. The London engineer can iterate every thirty minutes. The Lagos engineer may wait days between runs. The gap is not skill or ambition but the absence of shared compute access, data pipelines, and institutional capacity that would let AI work diffuse beyond a handful of countries.
This will determine whether AI becomes an engine of broad growth or a force that concentrates economic power further. AI is projected to add roughly $15 trillion to global GDP by the end of this decade, but that value will not distribute itself. Whether it reaches the four billion people living in emerging markets depends on choices made now about how AI capacity is shared, governed, and scaled.
In three weeks, India will host the first global AI summit convened in a developing country, an opportunity to move beyond abstract commitments toward the practical systems that make inclusion possible.
In our work across AI policy and investment, we see this pattern repeatedly. The binding constraint is rarely the technology itself but the ability to deploy it at scale, across institutions, and across borders. The UN launches its Global Dialogue on AI Governance this July, giving every country a formal seat at the table. An international independent scientific panel is now forming to anchor policy in evidence. What remains largely unbuilt is the connective layer that turns participation into practical capability.
Much of the debate about AI's global divide still focuses on access: who has the tools and who does not. That framing misses what actually determines outcomes. The more decisive challenge is diffusion, the capacity to move AI resources into local economies and institutions in ways that generate durable value.
India's recent commitment to make 38,000 GPUs available at subsidised rates is a meaningful step, but hardware alone does not solve the problem. Compute sitting in Hyderabad does little for a researcher in Accra unless there are standards for access, protocols for scheduling, and agreements that govern how data and results are shared. Other scientific communities have faced this challenge before. Particle physics solved it decades ago through the Worldwide LHC Computing Grid, which links 170 computing centres across 42 countries and gives more than 12,000 scientists near real-time access to data from CERN. The system runs millions of tasks each day. It worked because of sustained investment in protocols, authentication systems, and governance arrangements, not hardware alone.
AI now requires a comparable approach, and Delhi is well-positioned to begin that work.
Three outcomes from the summit would signal a shift from aspiration to execution: agreement on baseline standards for compute access, allowing distributed resources to function as a shared system rather than isolated national assets; practical data partnerships that enable countries to contribute to and draw from common training sets without surrendering sovereignty, starting with health, agriculture, and multilingual applications; and investment in regulatory capacity, including technical training and institutional support, so that governments can govern AI systems with confidence. Companies seeking regulatory predictability should be prepared to help build the expertise that enables it.
The economic logic supports this. Without coordination, the alternative is fragmentation: national AI systems that cannot interoperate, regulatory regimes that conflict across borders, and a digital economy in which friction absorbs value. For investors, the markets with the greatest long-term growth potential are precisely those where diffusion infrastructure does not yet exist. Building it is what enables deployment at scale. For companies, the risk is just as clear. Models trained on narrow data will underperform in diverse markets and face increasing regulatory resistance as governments respond to populations excluded from the training process.
Earlier summits laid the groundwork: Bletchley Park focused attention on frontier risks, Seoul secured voluntary commitments from leading developers, and Paris broadened participation while raising ambition. Delhi inherits that progress but brings a different form of legitimacy. A summit hosted in the Global South can speak to the majority of the world's population in ways earlier gatherings could not. Seven working groups are already negotiating concrete outcomes.
What remains is the infrastructure layer that translates political alignment into working systems. Railways preceded the spread of industrial production beyond Britain. Undersea cables enabled global finance. The standards, partnerships, and training programmes that could emerge from Delhi would serve a similar function for AI.
Without them, the technology will transform life for a minority while leaving the rest watching from the outside. The engineer in Lagos should not have to wait days to participate in the future being built.
This article is authored by Amandeep Singh Gill, United Nations Under-Secretary-General and Special Envoy for Digital and Emerging Technologies and Vilas Dhar, president, Patrick J McGovern Foundation.
