There is a persistent assumption that has quietly shaped AI policy across the Global South for the better part of a decade: that connectivity and infrastructure must be solved before AI adoption can begin. It sounds logical. It is also, the evidence now suggests, dangerously incomplete.
A systematic review of peer-reviewed literature across manufacturing, smart cities, healthcare, and education reveals a more nuanced picture. Infrastructure matters — but skills, governance, culture, and organizational readiness are often equally, if not more, decisive in determining whether AI truly takes hold.
How Big a Barrier Is Infrastructure, Really?
Across sectors — SMEs, smart cities, Industrial IoT, and low-resource healthcare — lack of connectivity consistently appears in the top five barriers to AI adoption. But it almost never appears alone, and rarely ranks as the single dominant constraint.
In manufacturing and IIoT contexts, organizational factors — skills gaps, leadership buy-in, company size, and R&D intensity — routinely outweigh infrastructure in predictive models of adoption (Kinkel et al., 2021; Magara & Phahlane, 2026). In MENA higher education and smart city deployments, financial constraints, policy vacuums, and cultural distrust sit alongside infrastructure as co-equal blockers (Wang et al., 2021; Rjab et al., 2023; Alzahrani & Alasmari, 2025).
"Even where connectivity exists — in parts of Nigeria, Nepal, Ecuador, and across MENA universities — adoption lags due to workforce unreadiness, leadership ambiguity, ethical uncertainty, and unclear return on investment." — Synthesized from Oyeyemi et al., 2025; Marey et al., 2025; Sone & Ebune, 2025
A Map of the Real Barriers
The infrastructure gap looks different depending on context. The table below synthesizes what the literature identifies as the dominant infrastructure-related friction across four major domains:
- SMEs & Business: Legacy IT systems, cloud fragmentation, and limited local computing power (Sánchez et al., 2025; Zavodna et al., 2024; Magara & Phahlane, 2026)
- Smart Cities / Urban AI: Telecom quality gaps, broadband dead zones, and funding shortfalls for network upgrades (Wang et al., 2021; Rjab et al., 2023; Das et al., 2025)
- LMIC Health & Medical Imaging: Unstable power grids, scarce diagnostic devices, and near-zero broadband (Wong et al., 2025; Mollura et al., 2020; Marey et al., 2025; Wahl et al., 2018)
- Developing Countries Broadly: Foundational gaps in internet access, electricity, data centers, and storage (Oyeyemi et al., 2025; Aderibigbe et al., 2023)
The Inconvenient Finding: Skills and Governance Often Come First
Multiple high-quality studies independently converge on a counterintuitive finding: knowledge and skill gaps, governance failures, and cultural distrust block AI adoption even in infrastructure-rich environments.
Survey-based research demonstrates that organizational readiness and strategic project management often explain more variance in adoption outcomes than connectivity status alone (Merhi & Harfouche, 2023; Kinkel et al., 2021). Gulamali et al. (2025) argue explicitly for "eliminating the AI digital divide by building local capacity" — positioning human capital as the primary lever, not cables.
This finding fundamentally reframes the policy conversation. Waiting for broadband towers before investing in AI education and governance capacity is not a careful sequencing strategy. It is a compounding delay that makes the overall adoption gap wider over time.
Three Practical Pathways Around the Infrastructure Gap
The most significant contribution of this literature is not its diagnosis, but its catalogue of working solutions. Three dominant strategies emerge from the evidence:
1. Cloud-Based Platforms and Shared AI Centers. Regional shared infrastructure — cloud platforms and hub-based AI centers — allows organizations to access sophisticated AI capability without maintaining it locally. This model has demonstrated success across radiology in Sub-Saharan Africa (Mollura et al., 2020), supply chain optimization in Southeast Asia (Aderibigbe et al., 2023), and SME digitalization in Latin America (Sánchez et al., 2025).
2. TinyML and Edge AI. Perhaps the most compelling technical finding: TinyML architectures enable genuinely useful machine intelligence to run on severely resource-constrained hardware — microcontrollers, low-power sensors, and sub-$50 devices — with no continuous internet dependency required (Ortiz et al., 2025; Magara & Phahlane, 2026). This is, for all practical purposes, the scientific basis for AFIRMASI's own Offline-First deployment methodology.
3. Hub-and-Spoke Capacity Networks. Public-private partnerships that position well-resourced institutions as "hubs" — providing AI infrastructure, training, and model development — while lower-resource "spokes" consume and localize these outputs have shown consistent effectiveness (Mollura et al., 2020; Gulamali et al., 2025; Aderibigbe et al., 2023).
"Connectivity and raw infrastructure are real constraints — especially for data-intensive, real-time systems. But research consistently shows that skills, governance, leadership, and trust are just as binding. And crucially, they can be addressed without waiting for the infrastructure gap to close." — Synthesized conclusion
What This Means for Frontier Education
For AFIRMASI's operational context — 3T schools in Eastern Indonesia with intermittent or nonexistent internet — this body of research is both validating and directional. The infrastructure gap is real. It is not, however, an excuse for waiting.
Edge AI and TinyML are mature enough to deliver meaningful, measurable learning support in offline environments. The more urgent investments are in teacher capacity, local governance of AI tools, community trust, and data sovereignty protocols. These are human problems, not engineering problems — and they are solvable now, with resources available now.
The road to AI-enabled education in Indonesia's frontier regions does not start with a 5G tower. It starts with a trained educator and a 4GB model running on hardware that fits in a backpack.