�鶹������ý

This report has been deliberately descriptive. Its goal has been to map a terrain that is too often described in the abstract and to surface the textures of what is actually happening on the ground. Although we list emerging workarounds, we have generally resisted the move from description to prescription—in part because the evidence base is not yet wide or representative enough to bear the weight of confident policy recommendations, and in part because we believe the more useful contribution here is to open questions rather than close them.

What follows, then, is not a conventional list of policy recommendations but rather a set of research and inquiry directions. Returning to our four cross-cutting observations, we suggest some areas where further work is most needed and where the conversations sparked by this report might productively focus.

The Thailand diabetic retinopathy deployment with which we opened is a useful reminder: A model that achieved 90 percent accuracy in the lab still foundered when dropped into clinics with variable lighting and slow internet. The pattern recurs throughout our evidence—the legal AI platform in India that struggled to reach scale without integration into judicial workflows, the agricultural tools that worked best when embedded in WhatsApp. In case after case, the frontier that matters for adoption appears to be contextual fit rather than raw capability.

Yet the field lacks shared ways of measuring (or supporting) the notion of contextual fit. How should on-the-ground success be defined for AI deployments—by build, by launch, by use, by outcome? What would meaningful evaluation of contextual fit look like in practice, and who is positioned to conduct it? How might procurement and funding frameworks incorporate fit alongside performance? These are just some of the questions worth exploring as we seek to better understand real-world AI adoption in the Global South and beyond.

We repeatedly found that successful deployments depend on an ability to work across layers. IndiaAI Mission has made 38,000 GPUs available to nonprofits and start-ups, but one of our interviewees pointed out that hardware access is only a beginning: Smaller and more remote organizations often lack the awareness and technical assistance that would let them actually benefit. A similar pattern emerged around open-source models, where many grassroots organizations continue to rely on commercial U.S. providers, despite known risks, because those providers offer credits, salespeople, and onboarding support that open-source alternatives cannot yet match.

These examples suggest that the layers of an AI deployment—hardware, software, data, governance, capacity—depend on each other in ways that are not yet well understood. What does an effective public compute ecosystem look like in practice, beyond GPU access alone? What would it take to build deployment infrastructure for open-source models that can compete with proprietary onboarding? Importantly, this is not an argument for every country to seek to develop capacity across every layer of the AI stack. That would be neither feasible nor desirable. The more useful goals are to understand how the layers interact, to identify which dependencies matter most for which use cases, and then for each country or region to make deliberate choices about where to invest, where to partner, and where to accept reliance on others. The question is less about building the full stack than about identifying points of leverage at which targeted investment yields the greatest downstream effect.

Among the most striking of our findings is the observation that, in the Global South, constraints are often producing genuinely novel approaches. For example, Crane AI Labs reduced its model size from 4 billion to 1 billion parameters to fit local compute realities. Amini ships modular GPU capacity in literal shipping containers, allowing governments to start with investments under $5 million rather than hundreds of millions. Likewise, Digital Umuganda is rethinking what data collection means, asking contributors to narrate stories around images rather than simply label them.

Whether these approaches add up to something coherent—a distinct paradigm of frugal AI development—is a question the evidence cannot yet fully answer. But it is worth taking seriously and investigating further. How might “” patterns from earlier technologies repeat or differ with AI? Under what conditions do frugal approaches outperform scale-driven ones, and where do they fall short? We should also consider what the Global North might learn in this area from the Global South—particularly as energy and compute costs constrain even the richest countries—and what it would take to support Global South AI as a generative source of innovation rather than simply a passive transfer destination.

AI adoption gaps are not evenly distributed. The Lebanese survey of young women, in which 54 percent cited privacy as a reason for hesitating to use AI chatbots for intimate health concerns, points to one form of unevenness. The Indian agricultural platform on which isolated women users began sharing sensitive personal information, suggesting forms of risk and safety challenges that are not yet well understood, is another example. Voice and multimodal interfaces emerged repeatedly in our research as essential for reaching low-literacy and digitally excluded users—yet voice and video are bandwidth-hungry, and the communities that would benefit most often have the weakest connectivity.

These patterns raise both empirical and normative questions. What does disaggregated evidence on AI adoption—by gender, language, geography, literacy—actually show, and how can it be gathered more systematically? What forms of safety, trust, and consent matter most, and for which populations? There are also concerns related to design and architecture: for example, how can AI deployers default to the needs of populations least likely to benefit, and what commercial models could make such design choices viable? More generally, we perhaps need to rethink what meaningful inclusion looks like, beyond just adoption. There are glimpses of possibilities in this paper: benchmarks that are more linguistically and culturally representative, for instance, and data compensation mechanisms that treat community contributions as ongoing sources of income rather than one-time transactions.

Taken together, the four directions above gesture toward the contours of a distinct approach to AI development—one that is frugal rather than capital-intensive, embedded rather than imposed, and attentive to gaps and inequalities. Whether that approach coheres into something recognizable will depend on the choices made by policymakers, funders, and practitioners over the coming years. It will also depend on the kinds of conversations the field is willing to have—about what counts as success, about who is being served, about what the Global South has to teach as well as to learn, and about how to balance the legitimate pull of sovereignty against the practical constraints of feasibility.

The deployment work documented in this paper—often unrecognized and unsupported, far from the glare of frontier labs—suggests that the building has already begun. Now the policy frameworks, institutions, and funding need to catch up. Our hope is that this paper contributes to building a more rigorous evidence base that can help foster this enabling ecosystem.