Sovereignty Week — Cassava AI Factory,
NVIDIA GTC, and Korea-Africa

Every African national AI strategy has been undermined by the same structural absence: no sovereign, production-grade compute within African borders for regulated workloads. The Servernah launch in Nairobi the previous week addressed East Africa. Cassava's AI Factory addresses the continent at scale — with an NVIDIA partnership that means the Vera Rubin GPU generation announced at GTC this week will flow directly into African AI infrastructure. The Zindi partnership matters beyond the marketing: 70,000 African data scientists gaining access to local GPU resources for training models on African problems creates the feedback loop between infrastructure and talent that has been absent. This is not a data centre announcement. It is the enabling layer for African AI sovereignty across the full stack — compute, models, data, and deployment.

The shift from training to inference is structurally significant for Africa. Training large models requires compute concentrations that Africa cannot yet host competitively. Inference — running trained models in production for real applications — requires distributed, sovereign, low-latency compute that is exactly what Cassava, Servernah, and the forthcoming MTN data centres are building. Africa's AI Factory strategy is, without having been framed this way, perfectly timed for the Inference Era: the continent is building the inference infrastructure layer at the precise moment when that layer becomes the primary value-creation point in the global AI stack. The $1 trillion order backlog means GPU availability through 2027 will remain constrained — Cassava's NVIDIA partnership provides preferential access to the scarce resource that will shape every African AI deployment in the next 24 months.

MTN's Capital Markets Day will name the strategic co-investment partners — and the identities of those partners will reveal whether Africa's AI infrastructure strategy is being built through US capital (aligned with Microsoft's Elevate initiative and Azure's Africa expansion), European capital (aligned with the EU-Kenya Digital Dialogue and French development finance), or a combination. Given the geopolitical context — Microsoft vs DeepSeek, EU-Africa digital partnerships, and China's existing surveillance infrastructure — the partner identities are not merely commercial decisions. They are geopolitical signals. MTN's 300 million subscribers and its distribution relationship with Microsoft Copilot mean that the data centre partners it chooses will have preferential access to the most extensive AI distribution infrastructure on the continent.

Kenya is running a domestic legislation track and an international diplomacy track simultaneously — and the EU's choice of Kenya as Africa's first formal digital cooperation partner is directly linked to Kenya's regulatory credibility, including the AI Bill in Senate deliberation. This creates a reinforcing dynamic: stronger domestic legislation increases Kenya's attractiveness as an EU partner, which increases the development finance flowing into the infrastructure the legislation assumes. The Vellum Kenya analysis raises a real concern — three new regulatory bodies may fragment oversight rather than concentrate accountability — but the EU's €430 million commitment suggests that the international community is pricing Kenya's governance ambition, not just its current institutional capacity. The open-source compliance gap identified in the analysis is the same problem flagged for Nigeria's bill and South Africa's policy: African developers who build on global open-source models cannot provide the training-data provenance these frameworks require.

The Rest of World investigation resolves any ambiguity about whether Africa's AI governance emergency is hypothetical. Two billion dollars of AI infrastructure has been deployed, financed by foreign loans, across 11 countries, with no adequate rights frameworks in any of them. The operational outcome is documented: surveillance enabling political repression. The governance frameworks being written in Nigeria, Kenya, and South Africa are not regulatory tidiness exercises — they are the legal architecture for democratic accountability that should have preceded these deployments. The investigation also reveals the structural problem with the Chinese AI surveillance model specifically: loan conditionality on technology purchase creates dependency on Chinese vendors and Chinese data infrastructure for the AI systems governing citizens' public movements, which no AI governance framework can retroactively unwind. The procurement decisions made before legislation are the decisions that will be hardest to reverse.

For every South African organisation deploying AI in financial services, healthcare, HR, or the public sector: begin internal AI audits immediately. Map all AI-driven decision flows. Document model training data provenance where possible. Assess algorithmic explainability against POPIA's existing automated-decision provisions. Identify sector-specific oversight gaps — the FSCA, Prudential Authority, and ICASA will each develop sector instruments, and the comment submissions made during the 60-day window will shape those instruments more directly than any post-implementation lobbying. The window, once open, closes in 60 days. Those who engage early will have materially more influence than those who engage after reading the final regulations.

If Nigeria's bill misses the end-of-March deadline, it does not die — it slips to Q2 2026. But the context of the slip matters: GITEX Africa Morocco opens on 7 April with $350 billion in investor assets, and the first question international capital will ask is whether Africa has binding AI governance. A missed Nigerian deadline, combined with South Africa's gazette still pending and Kenya's bill still in Senate committee, means investors arrive at GITEX with no enacted African AI law to anchor their compliance analysis. The IDS surveillance investigation makes this concrete: $2 billion deployed without accountability is what the absence of governance looks like. The Nigeria bill's deadline is not a legislative formality — it is Africa's most important governance signal of the quarter.

If Africa's AI economy is already larger than the labelled-AI funding data suggests, then the investment thesis for African AI is not "will Africa catch up to global AI" — it is "how do we capitalise on the embedded AI advantage that already exists in Africa's most successful tech companies?" This framing has three strategic implications. First, African AI companies should not compete on foundation model metrics but on sector-specific AI depth in markets they understand from the inside. Second, investors should stop measuring Africa's AI economy by "AI startup" funding and start measuring the AI capability embedded in African tech's most successful companies. Third, governance frameworks should be designed for embedded-AI-in-existing-sectors, not just for identifiable "AI companies" — because the former represents the majority of Africa's actual AI economy.

Three percent of global AI talent, 1% of global AI infrastructure investment, and 14 countries with active AI policies. These numbers are not discouraging — they are specific. They tell us exactly where the interventions need to occur: talent development programmes scaled to demographic size, infrastructure investment at a ratio proportional to market scale, and governance frameworks deployed across all 54 AU member states rather than the four that currently command legislative attention. BAFAI's curriculum expansion this week — adding Agentic AI, AI Ethics, and Data Annotation without coding prerequisites — is the human-capital response to the 3% figure. Cassava's AI Factory is the infrastructure response. The governance frameworks in Nigeria, Kenya, and South Africa are the policy response. All three are required simultaneously.

BAFAI's decision to design all seven certificate programmes without coding prerequisites is not a compromise on quality — it is a structural decision about who Africa's AI talent base should include. Most global AI education assumes a programming foundation that represents a small fraction of Africa's working population. BAFAI's model reaches the decision-makers, entrepreneurs, healthcare professionals, policy analysts, and educators who will deploy, govern, and procure AI systems — not just the engineers who build them. For Africa's AI governance ambition to be realised, the people making procurement, regulatory, and deployment decisions need AI literacy. BAFAI's no-prerequisite design is the education policy equivalent of Cassava's sovereign compute: building for African realities rather than importing frameworks designed elsewhere.