Global AI Diffusion: What the Data Shows
Adoption gaps, infrastructure bottlenecks, and the widening Global North–South divide — a first look at the numbers behind AI access.
If the inaugural framing asked who has access, this piece asks what the evidence currently allows us to say. The answer is incomplete — and that incompleteness is itself a finding.
The headline gap
Microsoft’s AI Diffusion Index (end-2025) remains one of the few public cross-country snapshots of AI adoption. It is imperfect — survey and enterprise methodology, uneven sample depth — but directionally clear:
AI adoption rates by country
Approximate end-2025 adoption. African countries highlighted.
Source: Microsoft AI Diffusion Index, end-2025
No African country had crossed a durable 20% threshold by the end of 2025 except South Africa at the margin. The continent’s median sits far below mid-tier digital economies in Europe and East Asia.
The gap is not “Africa is late to a fad.” It is that the conditions for sustained adoption — connectivity, compute proximity, skilled labour, procurement capacity — are concentrated elsewhere.
Three bottlenecks that show up in every dataset
Connectivity
Reliable, affordable bandwidth remains uneven. Latency to frontier APIs is a product tax on every inference call.
Compute proximity
GPU clusters and hyperscale regions cluster in a handful of jurisdictions. Distance is cost, dependency, and political risk.
Institutional demand
Adoption tracks organisations that can buy, audit, and integrate models — not just populations that own smartphones.
Open-weight models narrow the capability gap faster than the infrastructure gap. A DeepSeek-class model available at a fraction of frontier API cost does not create data centres, fibre routes, or procurement teams. It can reduce unit costs for those who already have the rest of the stack.
What “diffusion” actually means here
Policy debate often treats diffusion as a single curve. For LMICs, at least three curves matter:
- Consumer / enterprise use — chatbots, copilots, office automation
- Sectoral deployment — clinical decision support, agritech advisories, credit scoring
- Sovereign capacity — local fine-tuning, evaluation, hosting, and governance
A country can rise on (1) while stalling on (2) and (3). That pattern — imported interfaces without domestic leverage — is the dependency risk this newsletter will keep returning to.
Methods note
For this inaugural issue we rely on secondary sources (Microsoft AI Diffusion Index; public infrastructure announcements; AU and national strategy documents). Future issues will publish our own indicator set: compute announcements, university programme counts, procurement notices, and language-coverage proxies — with full methods in an appendix.
The North–South divide in AI is measurable. Closing it is not a measurement problem. It is a politics of investment, ownership, and standards — which is why the next piece turns to LMIC-specific constraints.