The wrong bottleneck
A decade of policy documents about artificial intelligence in Africa has concentrated on regulation and on skills. Both matter. Neither is currently what limits what can be built.
What limits it is that training and serving models at scale requires reliable electricity, that reliable electricity is unevenly available, and that where it is available the transmission capacity to deliver it to a data centre frequently is not. A jurisdiction can pass an excellent AI statute and train excellent engineers, and if those engineers rent their compute from a provider on another continent then the jurisdiction has limited practical influence over the systems its economy depends on.
Connectivity: the middle mile
Subsea cable capacity reaching the continent has increased substantially. This is real and it is frequently the only figure cited. It is also the least binding part of the chain.
Bandwidth arriving at a landing station has to be carried inland, and terrestrial backhaul has not expanded at the same rate. The consequence is a pattern visible in pricing data across several markets: capacity is abundant and comparatively cheap at the coast and scarce and expensive a few hundred kilometres inland. Reporting that treats landing capacity as national capacity misses the constraint entirely.
Compute, and what renting it means
Access to accelerated compute across the continent is overwhelmingly rented from providers headquartered elsewhere. This has three consequences worth separating.
The first is cost, in foreign currency, which is a material constraint for a research group funded in local currency. The second is data residency, which interacts with domestic data protection law in ways that are frequently unresolved. The third is leverage: an organisation whose systems run on infrastructure it does not control has limited ability to insist on anything about how those systems behave, and a regulator supervising it has less still.
Regulatory capacity is the variable that matters
The distinction that predicts whether AI governance takes hold in a jurisdiction is not whether a strategy has been published. It is whether a supervisory authority exists, whether it has technical staff, and whether data protection law is in place for AI-specific requirements to attach to.
Where those conditions hold, comparatively modest instruments have effect. Where they do not, comprehensive statutes are complied with by organisations that choose to and ignored by those that do not. Capacity-building is therefore a more consequential intervention than further policy development, and it is funded at a fraction of the rate.
Language and the cost of being under-represented
Model performance across most African languages lags well behind performance on languages with far fewer speakers. The causes are structural: little digitised text, benchmarks translated from English rather than built natively, and tokenisers fitted on high-resource languages that fragment others into many more tokens per word.
That last point has a direct economic consequence that is rarely stated: speakers of under-represented languages pay more per query for worse output. It is a regressive property of how these systems are currently built, and it is among the more tractable problems on this list.
Investment concentration
Capital into African technology companies has remained concentrated in a small number of markets and a small number of sectors, with financial services absorbing a disproportionate share. AI-specific investment follows the same pattern, which means ecosystems outside those markets are building without the capital that would let them retain the people they train.
What this report will contain in full
The full edition carries a country annex covering regulatory status, data protection framework, national strategy status, grid reliability, and connectivity position for each of the fifty-four states, together with the underlying dataset and its sources. That annex is in preparation.