Technical Barriers and Solutions

One of the central observations of this report is that AI adoption is a whole-of-society challenge. At the same time, the multidimensional nature of the endeavor—encompassing society, policy, and political economy—should not obscure the fact that technology remains core to whether and how AI is used across the Global South. 

This section examines four dimensions of AI technology: compute, models, data, and infrastructure (covering electricity and connectivity). While there are other relevant technological parameters—the software and orchestration layers that allow models to be deployed and chained into working applications would be a plausible fifth dimension—these four are the variables that surfaced most often in our research.

Almost half a decade into the generative AI revolution, with the technology having now reached more than half the world’s population, access to compute remains a stubborn and foundational bottleneck. The numbers across the Global South are stark. Africa, home to 18 percent of the world’s people, accounts for less than 1 percent of global data center capacity and an even smaller share of GPU infrastructure. India generates roughly a fifth of the world’s data but holds only around 3 percent of global data center capacity. The United States and China together contain 90 percent of the world’s GPU cluster performance.

On the ground, these disparities translate directly into deployment constraints. Crane AI Labs, a Ugandan company, was forced to operate with just two donated GPUs for training, leaving none available for inference (the stage at which the model actually reaches users). In Uruguay, a university research lab reported being limited to fine-tuning models with a maximum of 7 billion parameters (some frontier models now exceed a trillion) due to limits on the national computing cluster. Rodrigo Durán of CENIA (Chile’s National Center for Artificial Intelligence) estimated that compute access is the primary bottleneck for the vast majority of AI companies in Latin America and that the constraint is even worse for universities and civil society organizations, which lack the philanthropic cultures that cushion nonprofit organizations in wealthier countries.

AI sovereignty policies play a role, too, if inadvertently. Data localization and similar laws are often reasonable in intent but can create compliance traps that undermine the very goals they are meant to serve. Several organizations told us they find themselves caught between sovereignty aspirations and infrastructure realities—legally prohibited from using foreign cloud providers but lacking the local compute to run models themselves. DataGénero’s AymurAI, a tool for processing judicial documents in Latin America, illustrates the bind: The system is designed to run locally so that sensitive judicial data stays protected, but judicial offices lack the necessary hardware, forcing the organization to fall back on less capable models calibrated to limited compute. 

Digital Umuganda, a Rwandan company that builds capacity for low-resource languages, faces a different version of the problem: Its team collects community language data across Africa but encounters legal obstacles transferring it back to Rwanda to train the models that would arguably benefit the originating countries most. The deeper irony is that because larger—and often foreign—companies are better positioned to operate local compute facilities, data localization laws can disadvantage the very domestic start-ups and communities they were meant to nurture. They also foreclose the kinds of cross-border compute-sharing arrangements—regional clusters, pooled GPU access, federated training across jurisdictions—that could offer smaller and middle-power countries a way around their individual infrastructure limits. 

Solutions: Despite these stark disparities, compute did not in fact rank among the chief obstacles identified in our case studies (Table 1). This may reflect the fact that solutions are emerging—though they require rethinking the traditional paradigm of compute. AI compute in the Global South may come to look fundamentally different from the large-scale, centralized, and capital-intensive infrastructure that dominates mainstream AI discourse. 

One insight recurred across our interviews: Organizations deploying on the ground are finding they may not in fact need nearly as much compute as the traditional paradigm assumes. Part of this reflects an important but often elided distinction between two very different compute needs. Training (the process of building a model from scratch) is enormously compute-intensive; by contrast, inference (the stage at which a trained model actually serves users) requires far less compute. Most organizations we spoke with are not training frontier models: They are fine-tuning open-source models for specific use cases or for running inference on models built elsewhere. Their compute needs, accordingly, look very different from the headline numbers that dominate AI infrastructure debates.

This reality is reshaping how many Global South organizations think about compute. Kate Kallot of Amini, a Kenya-based company, said that governments are often told that “the only way to do AI is to build gigawatt factory data centers” costing hundreds of millions of dollars. Her company takes a different approach, scaling compute incrementally for smaller models targeted at particular use cases and problems. Building modularly—“literally shipping containers” of GPU capacity, stacked as necessary—Amini has been able to help governments deploy robust AI capacity with initial investments of under $5 million. Crane AI Labs arrived at a similar insight, reducing its model size from 4 billion to 1 billion parameters and dramatically cutting compute requirements. Both examples illustrate the kind of frugal, need-based innovation emerging across the Global South—a recurring theme in our research. As mDoc Healthcare, a Nigerian digital health platform, put it in a survey response: “Frugal engineering is not a choice for us. It is the condition under which we operate.”

Compute is often framed as a technical issue, but some of the most effective solutions are to be found in policy. India’s investment in public compute is among the clearest examples. Under the IndiaAI Mission, the country has made more than 38,000 GPUs available to nonprofits and start-ups at subsidized rates—part of a broader interest in public compute across the Global South, including in Morocco, Brazil, and Senegal. Several interviewees noted that this has substantially eased compute constraints in India’s social impact AI ecosystem, especially for resource-constrained organizations.

One interviewee offered an important caveat to this generally positive picture, with global implications. Access to GPUs is necessary, she said, but not sufficient. Despite technical availability, smaller and more remote organizations—those removed from the corridors of power—often lack the human connections or even awareness that would allow them to benefit from public compute. The real unlock, she said, is ecosystem awareness: knowing that public compute exists, knowing how to access it, and—importantly—knowing what has worked for other organizations (i.e., best practices). Case studies, she suggested, can offer road maps and frameworks for adoption and often matter as much as the hardware itself—an indication that the empirical work this report calls for is not only a precondition for good policy but even a precondition for use.

Challenges:

• GPU and compute scarcity
• Sovereignty policy traps

Solutions:

• Frugal engineering: modular and edge compute, smaller models
• Public compute
• Focus on inference and fine-tuning, not frontier training
• Regional collaborations and compute pooling

The situation with models is more nuanced than with compute—the problem is more uneven, less about a single bottleneck, and solutions are correspondingly varied. Global South organizations trying to deploy AI models face at least two core challenges: achieving contextual fit, and avoiding dependency or lock-in. Solving for one does not necessarily solve for the other, and both require a mix of technical, social, and policy responses.

The contextual fit problem is by now well documented: Proprietary models, trained predominantly on English-language and Global North data, struggle for relevance amid very different linguistic, cultural, and sectoral contexts. The dependency problem is somewhat different. It arises when organizations build critical workflows on top of models they do not own or control, creating exposure to decisions made elsewhere. The problem compounds over time as switching costs rise. The concern was most often raised in our interviews by officials and policymakers expressing concerns about national sovereignty, but it applies equally to organizational sovereignty and agency. Like countries, companies and NGOs seek to avoid being locked into a small number of dominant providers. This problem is global—online discussion boards are full of users complaining about deprecated models and unannounced changes—but the risks can be particularly acute in the Global South, where small organizations often lack the technical capacity and financial means to adapt to new APIs, pricing tiers, and model behaviors. 

Solutions: The dominant workaround to the first challenge (insufficient context) is to layer smaller, locally trained models on top of larger foundation models or, alternatively, to fine-tune foundation models on local data. The approach is particularly common in language, health, and agriculture, where local relevance is especially critical. As Rikin Gandhi, CEO of Digital Green, puts it, “Agriculture is hyperlocal: soil type, rainfall, altitude, pests, and markets vary village to village. Model learning must stay close to those realities.” Many organizations we interviewed are now pursuing this strategy, with meaningful results. 

But model stacking or fine-tuning leaves the second challenge (model dependency) largely unaddressed. Smaller contextual models are typically grafted onto imported foundation models—a strategy the AI Now Institute has, in a related context, described as “barnacles on the hull of Big Tech.” The structural dependency remains, along with the sovereignty and market concentration concerns that come with it.

Open source is frequently proposed as a solution to this problem, and our interviews suggest it has genuine promise. But open source is also more complicated than sometimes advertised. Many of the organizations we spoke with used proprietary models despite the availability of open-source solutions. This was for at least two reasons: a lack of trust in the security, performance, and sustainability of open-source solutions; and a shortage of technical capacity to implement and adapt open-source models. One interviewee from India emphasized that open-source models—including China’s DeepSeek—were primarily being used by larger enterprises with in-house technical skills. Smaller organizations, unable to adapt and deploy open source, were left with little alternative but to rely on commercial solutions.

As this interviewee pointed out, models operate in an “ecosystem.” She told us that, in her experience, a large number of grassroots Indian organizations used a commercial U.S. provider despite being aware of the risks and shortcomings, simply because this company offered credits for inference and a training and deployment apparatus—including salespeople and technical assistance—that made accessing and building on top of the model easy. Without this broader ecosystem, the promise of open source remains difficult to manifest. 

Challenges:

• Model relevance
• Model dependency 

Solutions:

• Build and layer small local models onto proprietary models (“barnacles on the hull of Big Tech”?)
• Open-source models
• Model ecosystems

Data emerged as the most common challenge in our structured case study analysis, cited by 79 percent of the 53 examples in our case study repository. It was also mentioned as an obstacle by nearly all our interviewees. Broadly, Global South organizations face two types of data challenges: those related to representativeness, and those related to accessibility. 

Data representativeness is closely linked to the contextual fit challenges discussed above. Although model relevance cannot be reduced entirely to training data—architecture, fine-tuning, and deployment context all play a role—the problem often starts with data. Many organizations we interviewed emphasized both the importance and the difficulty of sourcing locally relevant data. One respondent working on a low-resource language model in Rwanda told us the largest parallel corpus available—documents containing original and translated text—came from old Jehovah’s Witness texts, written in stilted, archaic language that is unsuitable for a modern chatbot. DataGénero noted that it is very hard to find high-quality data that is not gender biased.

The problem of data representativeness typically stems from an absence or paucity of data. Often, however, data exists but remains inaccessible. This can happen for a variety of reasons, but many of our interviewees spoke of technical barriers to AI-ready data. According to the World Bank’s Statistical Performance Indicators index—which includes several proxies for AI-ready data—low-income nations rank significantly behind high-income nations, and the gap is widening. 

A lack of AI-ready data can take many forms. Some of our respondents spoke of non-digitized or unstructured data, but the challenge can manifest in less intuitive ways. Rodrigo Durán of CENIA told us that when his team asked Latin American institutions for data to train LatamGPT, a regionally focused large language model (LLM), they received structured SQL databases—precisely the wrong format for training a language model, which requires unstructured natural language text. Durán also drew attention to the governance dimensions of data accessibility, telling us that his work was limited by the absence of standardized data access protocols across governmental organizations, meaning that data sharing arrangements had to be individually and painfully negotiated on a bilateral basis. We heard similar complaints from several others, who described navigating a thicket of emerging and often unclear data handling and privacy laws.

Finally, the data challenge applies not only to training data but to the benchmarks used to evaluate models. Happy Buzaaba, a researcher working on multilingual AI with a focus on low-resource languages, described co-creating a benchmark for 17 African languages that he ultimately concluded could inadvertently perpetuate cultural assumptions from their source datasets: Although the questions were translated into African languages, they were adapted from English-origin benchmarks and therefore did not fully reflect African cultural contexts. Overlooking the benchmark layer risks perpetuating biases beyond the model itself: even LLMs built on well-collected, contextually rich data are subject to distorted evaluation if the benchmarks used to assess them reflect Global North assumptions.

Solutions: Organizations are taking a range of approaches to solving data supply problems—some primarily technical, others more oriented around policy or society. On the technical side, synthetic data has emerged as one workaround when real-world data is scarce. Crane AI Labs, for instance, uses synthetic data for their low-resource language speech models; in particular, they use voice cloning to generate synthetic voices with specific accents, for which there is a paucity of data. iTanong, a Filipino-language government chatbot, is another organization that was forced to make up for a lack of voice data by relying on synthetic solutions.

Synthetic data, however, carries well-documented risks (recursive contamination, compounding errors), and several organizations we interviewed said they avoid it for these reasons. A second technical approach is building APIs and technical pipelines to pull data for inference. Rather than relying on the models themselves to contain the most up-to-date and relevant information, organizations are able to link the models to local, real-time external data sources. Digital Green, for example, uses live weather data via APIs to inform responses in its multimodal agricultural chatbot platform; others have built similar pipelines to access live market prices. 

One emerging approach is to repurpose data already collected through digital public infrastructure (DPI). One Indian government official told us that DPI investment had been transformative for AI deployment: Standardized electronic health records and unique citizen IDs mean that AI-ready, structured data already exists at scale. The bottleneck in this approach is not data supply per se but data governance (see Durán’s point, above): establishing the reuse rights, interoperability standards, and access protocols that allow DPI data to flow into AI systems. India’s recently released “AI Governance Guidelines” take a step in this direction, recommending extensions to existing consent-based data-sharing protocols, new data portability rights, and tighter integration between AI development and India’s DPI stack. This reflects a broader, cross-national push to develop the legal and technical protocols that govern how data can be reused, shared, and accessed for AI development. 

Technical and policy approaches of these kinds are promising for data that already exists but is inaccessible or poorly formatted. However, much of the data that Global South AI systems need does not exist in any written, much less digitized, form—cultural knowledge, oral traditions, Indigenous languages—and here the solutions emerging on the ground are fundamentally human and social. One approach that came up repeatedly was the use of participatory or community data collection. This method treats data creation as a social process: going into communities, building trust, and co-creating datasets with people whose lives the AI is meant to serve. The approach is being used across the Global South in a variety of ways. Masakhane, a pan-African collective, runs workshops stretching over days in which native speakers co-create corpora for under-resourced languages; Aakhor AI collects voice samples in low-resource languages, often through short WhatsApp recordings.

Digital Umuganda has rethought what data collection (and data itself) means. Rather than asking contributors simply to label or describe what they see—the standard annotation approach—Digital Umuganda asks them to narrate a story. Contributors shown an image of a local football stadium do not just say the local word for that stadium, they create a narrative around it, drawing on the significance of football in their community. The result is data that carries not just labels but context (local idioms and cultural associations) embedded directly into the dataset. This richer, more contextually grounded data helps address not just the supply problem but also the deeper challenge of relevance and cultural fit. 

Challenges: 

• Data representativeness (contextual, cultural, and linguistic fit)
• Data accessibility (machine readable, technical and policy silos)
• Distorted benchmarks and evaluation

Solutions:

• Synthetic data
• Pulling data from live APIs
• Repurposing DPI data (with governance safeguards)
• Community and participatory data collection

While much of the AI conversation focuses on compute, models, and data—our own focus so far—deployments in the Global South are often shaped by more fundamental infrastructural challenges. Many of the organizations we spoke with cited two in particular: electricity and connectivity. Infrastructure was also the third most-frequent constraint that arose in our case study analysis (Table 1). 

AirQo, a Ugandan company that uses AI to monitor air quality across African cities using IoT sensors, illustrates the electricity constraint. Deo Okure, AirQo’s head of research and global partnerships, described unreliable power supply as a primary bottleneck for deploying their hardware. His team’s response has been to integrate solar modules directly into hardware design. 

Connectivity is an even bigger infrastructural constraint, and it comes in several dimensions. The first is basic access: although many of the organizations we spoke with operate in regions with at least some internet connectivity, this is by no means universal, especially in rural or mountainous areas. Even when connectivity exists, pricing can be prohibitive, requiring innovative workarounds. Digital Green found that high cellular (internet) data costs in both Kenya and Ethiopia were a significant barrier to adoption of FarmerChat, its multimodal AI chatbot for agricultural support, and responded by negotiating zero-rating partnerships directly with telcos, making the app free for users to access.

Bandwidth is a third dimension of the connectivity challenge—one that poses particular dilemmas for the low-literacy and low-digital-fluency communities that rely on voice, image, and video interfaces. Arguably among our most important findings is the fact that voice has an increasingly important role to play in bridging the AI access gap, allowing citizens to interact more naturally with AI systems and playing a central role in community and participatory data efforts. But voice is bandwidth-hungry, and low or unstable connectivity limits the feasibility of precisely the modalities that matter most for underserved communities. Digital Umuganda’s team pointed out that broader language coverage means reaching deeper into rural communities and, in the process, pushing larger audio files through less stable connections.

Solutions: Organizations are responding with a range of workarounds to these infrastructural constraints. In Kenya, Microsoft and local ISP Mawingu Networks have teamed up to deploy satellite-based connectivity (in this case, via Starlink) for AI-powered agricultural services across 450 community hubs. AirQo’s solar integration (discussed above) addresses power intermittency at the device level, a strategy adopted by several other organizations that are essentially sidestepping infrastructural constraints by building their own portable infrastructure. This pattern has a longer history than AI: Projects like Internet-in-a-Box and RACHEL have for years bundled offline content servers with local Wi-Fi into single devices for deployment in schools, clinics, and libraries across the Global South. Many of the efforts we encountered represent the AI-era extension of this lineage.

The most striking example we encountered of this “AI in a box” approach was 0-LA, built by Crisis Cognition, a for-profit start-up focused on disaster response and humanitarian aid. 0-LA is a briefcase-sized solar-powered unit that bundles an LLM, database, and web server, connected by its own Wi-Fi network. It is, in effect, a self-contained AI deployment system, the entire technical and infrastructural stack compressed into something a person can carry. 0-LA represents an extreme—and particularly ingenious—illustration of edge AI, and a good example of the constraint-generated creativity that surfaced throughout our research.

Challenges:

• Unreliable electricity
• Limited or expensive connectivity
• Bandwidth too low for voice, image, and video 

Solutions:

• Partnerships with telcos for zero-rating
• Solar integrated hardware
• “Portable infrastructure”: self-contained AI solutions