Kilwa Atlas No. 1: AI Infrastructure & Sovereign Compute

The electron era

For two decades the question about African digital infrastructure was whether the data could get there. That question is largely answered: the 2Africa system alone carries a design capacity of 180 terabits per second, more than every previous cable serving the continent combined (Meta, November 2025).

The question that now decides projects, returns and national strategies is whether the electrons can get there, at data-centre quality, at a contractable price. Every material event of the past eighteen months — Kenya's suspension of a $1 billion campus, Nigeria's gas-fired build plans, Morocco's renewable-backed AI factory — is a variation on that single constraint.

“Electrons now bind before bandwidth.”

What changed since mid-2025

Three signals. The 2Africa cable system was completed, broadly doubling the continent's international capacity. Kenya suspended the $1 billion Microsoft and G42 campus over power — the clearest evidence yet of the new binding constraint (May 2026). And the first sovereign-scale AI factories reached financial reality: Cassava's GPU deployment in South Africa and the $1.2 billion Nexus facility in Morocco (March and April 2026).

The gap between announced and operating capacity is at a record. Roughly 360 MW operates continent-wide while multi-gigawatt pipelines are announced in single countries. Investors who can price the difference hold an information edge for the 2026 to 2028 allocation window.

Three paths and a stress test

Likelihood bands follow the standard estimative scale and are Kilwa's subjective judgments as of the publication date, not measured frequencies and not forecasts of returns.

1. 01Base · grid-led, hub-concentrated buildout — 55–65%Operating capacity roughly doubles by 2030 against the projected 2 GW of demand, but 70% or more of new AI-grade capacity lands in South Africa, Morocco, Egypt and Lagos. Gigawatt visions elsewhere deflate quietly.
2. 02Upside · export breakout — 15–20%European power scarcity persists and North African renewable compute wins real training offtake. Morocco's pipeline converts at several hundred megawatts; Egypt follows.
3. 03Downside · announcement recession — 20–25%Global AI capex cools or financing costs stay punitive. Flagships slip, sovereign funds pledge but do not draw, and the announced pipeline visibly recedes.
4. 04Stress · access shock — 5–10%A restrictive successor to the Diffusion Rule or a GAIN-style allocation priority effectively caps advanced-chip flows to middle-tier countries.

What would change our view

Evidence that grid reform is adding firm, contractable capacity faster than we assume, above all in Kenya and Nigeria. A US chip-access regime that hard-caps African deployments. Or two or more flagship projects reaching operations on schedule, which would argue the announcement discount is too harsh.

The ten-indicator monitoring board — Eskom load-shedding days, the Olkaria restructuring outcome, Nexus financial close, Cassava GPU count, Nigerian gas-to-compute closings, the US chip rule, IFC benchmark commitments, ADCA construction-to-active conversion, funded national AI programmes, Red Sea cable incidents — is graded in public. Atlas No. 2 will publish every dated signpost above against what happened, with misses stated as plainly as hits.