AI Data Centers vs Power Limits in East Africa

Africa’s AI compute build-out is real, and so is the wall it just hit. In March 2026, Cassava Technologies switched on the continent’s first Nvidia-powered “AI factory” in South Africa — 3,000 GPUs scaling toward 12,000 across facilities planned for Kenya, Nigeria, Egypt, and Morocco — explicitly framed as sovereign African compute (1, 2). Two months later, Kenya suspended the $1 billion Microsoft–G42 geothermal data center at Olkaria because its requested 1 gigawatt of power would have consumed roughly a third of the nation’s installed generation capacity (3, 4). Those two events are the whole story in miniature: African compute is coming, but on infrastructure timelines measured in five-to-ten-year increments, gated by electricity and transmission rather than by chips or ambition. For an East African SME or builder, the strategic conclusion is direct — adopt AI now with the global APIs and small open models already available, keep your data and workflows portable, and be ready to migrate the day African compute becomes cheap and local.

Key Takeaways

• Cassava Technologies deployed Africa’s first Nvidia-powered AI factory in South Africa in March 2026 with 3,000 GPUs, expanding toward 12,000 GPUs across Kenya, Nigeria, Egypt, and Morocco — offering GPU-as-a-service designed to keep African data within African borders (1, 2).
• Kenya suspended the $1 billion Microsoft–G42 geothermal data center at Olkaria in May 2026: the project’s 1 GW power demand collided with Kenya’s roughly 3 GW of installed capacity, and government balked at underwriting the load through long-term power purchase agreements (3, 4).
• The gap is structural: Africa hosts under 2% of global data center capacity — roughly 1.2 GW of IT load in 2025, projected to nearly triple by 2030 — on a continent where about 600 million people still lack electricity (5, 6, 7).
• Compute sovereignty runs through foreign supply chains anyway: Africa’s roughly $60 billion in AI-infrastructure ambitions depend on Nvidia chips and hyperscaler partnerships, so “sovereign” means jurisdiction over data and pricing power over inference — not chip independence (8).
• For SMEs the practical stakes are latency, data-residency compliance under national AI and data-protection laws, and eventually cheaper inference — none of which justify delaying adoption today (9).
• The operating doctrine: run on the business clock, not the infrastructure clock — adopt via cheap APIs and small open models now, keep the minimum AI dataset portable, and treat local compute as a migration option, not a prerequisite.

What Is Actually Being Built — and What Just Stalled?

Two flagship projects define the moment, and they point in opposite directions.

The one that shipped: Cassava’s AI factory. Strive Masiyiwa’s Cassava Technologies — Africa’s first Nvidia Cloud Partner — brought its South African facility live in March 2026 with 3,000 Nvidia GPUs, the first tranche of a planned 12,000 distributed across Egypt, Nigeria, Kenya, and Morocco over the next few years (1, 2). The offer is GPU-as-a-service and AI-as-a-service: African developers, enterprises, and governments rent serious accelerated computing without building their own facilities, and — the explicit selling point — the infrastructure and the data stay under African jurisdiction. Alongside the silicon, Cassava launched CAIMEx, a multi-model exchange giving African developers unified access to leading models (2). This is the sovereignty thesis in commercial form: not “Africa builds its own chips,” but “African workloads run on African soil, priced by an African provider.”

The one that stalled: Olkaria. The Microsoft–G42 plan for a $1 billion geothermal-powered data center in Kenya was the continent’s marquee hyperscale project — and in May 2026 the government suspended it. The arithmetic was unforgiving: the consortium wanted assurance of up to 1 GW of power for full build-out, against Kenya’s total installed capacity of roughly 3 GW; President Ruto’s framing was that switching on the data center would mean switching off half the country (3, 4). The deeper failure points matter more than the headline: the consortium wanted the government to underwrite the entire load through state geothermal plants and long-term purchase agreements — financial exposure officials priced in the hundreds of millions annually — and Kenya’s transmission grid could not reliably deliver that much power to the site even if generation existed (4). Negotiations broke down on exactly the question every African mega-project must now answer: who carries the energy risk?

Read together, the lesson is not “Africa can’t build AI infrastructure.” It is that the binding constraint is electricity and transmission, not capital or chips — and that the projects which proceed will be the ones sized to the grid that exists (Cassava’s thousands of GPUs) rather than the grid we wish existed (Olkaria’s gigawatt). The continent-level numbers confirm the gap: Africa hosts under 2% of global data center capacity, around 1.2 GW of IT load in 2025 against projections of roughly 3.5 GW by 2030 — growth, but from a base smaller than single American states — on a continent where some 600 million people still lack electricity at all (5, 6, 7). The same power constraint that frames the data-center race also frames the productive-use solar revolution in East Africa: energy, not enthusiasm, is the gating asset of the decade.

Does Compute Sovereignty Matter for an East African SME?

The phrase “compute sovereignty” gets used as if it were one thing. For a small firm it decomposes into four distinct stakes — two that matter soon, one that matters eventually, and one that is mostly political theater at SME scale.

Data residency matters soon. East African governments are writing AI and data rules in real time — Kenya’s National AI Strategy 2025–2030, Uganda’s forthcoming AI legislation, existing data-protection regimes — and sectors like banking, health, and government procurement increasingly require data to remain under national jurisdiction (9). Today, an SME using a US-hosted API is in a gray zone its lawyers tolerate; as rules harden, local hosting options like Cassava’s become the compliance path of least resistance. The firms that win government and bank contracts in 2028 will be the ones that can answer “where does the data live?” with an African address. This is one more reason to read national AI strategies as a price list of coming requirements — the argument I made in what national AI plans mean for a small firm.

Latency and resilience matter soon for some workloads. Voice agents, real-time translation, and point-of-sale AI feel the round-trip to Virginia or Frankfurt. Local inference shaves that latency and keeps services alive when international links degrade. If your product serves customers in real time in local languages, regional compute is a genuine product-quality upgrade, not a flag-waving exercise.

Price matters eventually. Inference priced in shillings by a provider whose costs are African, with competition from multiple regional facilities, should eventually undercut dollar-denominated hyperscaler pricing for many workloads — and removes FX exposure from your cost of goods. But “eventually” is doing heavy lifting: new infrastructure amortizing fresh capital expenditure rarely opens cheaper than global providers operating at hundred-fold scale. Expect local compute to win first on compliance and latency, and only later on price.

Chip independence is not your problem. Critics correctly note the contradiction in Africa’s roughly $60 billion sovereignty agenda: it runs through Nvidia GPUs, Microsoft partnerships, and Gulf capital end to end (8). True — and strategically irrelevant for a small firm. Sovereignty at the SME level is not about where the silicon was fabricated; it is about whether you can exit any provider without losing your business’s accumulated intelligence. Which brings us to the framework.

How Should a Small Firm Plan When Infrastructure Runs on Decade Timelines?

The planning error I see constantly — in both directions — comes from mixing up timescales. Some founders delay AI adoption “until the local data center opens.” Others sign multi-year commitments to a single global provider as if the landscape were finished. Both mistakes dissolve under what I call The Three Clocks — the recognition that African AI runs on three different clocks, and your strategy must read all three without confusing them.

The infrastructure clock ticks in five-to-ten-year increments. Power plants, transmission lines, and data centers move at the speed of concrete, turbines, and sovereign guarantees — Olkaria just demonstrated how that clock can stop entirely (3, 4). Nothing your business needs this year should depend on it. Its strategic use is directional: it tells you what will be possible in 2030 — regional GPU clouds, local-language model hosting, shilling-priced inference — so you can design today’s systems to be ready for it.

The policy clock ticks in one-to-three-year increments. AI strategies, data-protection enforcement, and sector rules will land before the big infrastructure does (9). This clock sets your compliance posture: classify your data now, know which workflows touch regulated categories, and prefer architectures you can re-host when residency rules tighten.

The business clock ticks in quarters. Your competitors adopt WhatsApp agents, AI bookkeeping, and automated follow-up this quarter, using tools that exist today. This is the only clock your P&L answers to — and it is the clock the playbook disciplines: pick one workflow, ground the AI in approved knowledge, measure within 90 days whether it pays rent.

The Three Clocks yield one operating doctrine: adopt on the business clock, comply on the policy clock, stay portable for the infrastructure clock. Portability is the executable part, and it has three components. First, keep your minimum AI dataset — price lists, policies, product knowledge, golden conversations — in plain, structured formats you own, never locked inside one vendor’s platform; the dataset, not the model, is your firm’s accumulated intelligence. Second, prefer open-weight-compatible architectures where stakes are high: the decision logic I detailed in the open-weights versus API choice facing African builders — use cheap APIs to find what works, then graduate proven, sensitive workflows to small open models you control — is exactly what makes a future migration to African compute a weekend project instead of a rebuild. Models like the compressed, African-language-capable systems I profiled in the tiny-model revolution built for African constraints already run on hardware a Kampala firm can afford, which means a useful fraction of “sovereign compute” is available today at desk scale. Third, budget AI like airtime, not rent — per-workflow, capped, and reviewed — the discipline from what AI actually costs a Kampala SME, so that no provider’s pricing change can ambush you.

A firm running this doctrine gets the best of every scenario. If Cassava’s Kenyan facility opens on schedule and prices well, migration is cheap because nothing was locked in. If the infrastructure clock slips — and Olkaria says it will, repeatedly — the business lost nothing, because it never waited.

What Are Realistic Timelines — and What Should Builders Watch?

Forecasting African infrastructure rewards humility, but the signals are legible enough to bracket.

Now through 2027: the rental era. Global APIs and small local models carry essentially all East African SME workloads. Cassava’s South African capacity serves the continent’s early heavy users — broadcasters, banks, governments, AI startups — while its Kenyan, Nigerian, Egyptian, and Moroccan expansions work through power agreements and construction (1, 2). Expect announcements to outpace energized racks; the 350 MW of additional African data-center capacity expected in 2025 was continental, across all workloads, not AI-specific (5). SMEs should treat any “wait for local compute” advice in this window as a cost, not a strategy.

2027 through 2030: the hybrid era. Regional GPU capacity comes online in Nairobi and beyond at moderate scale — tens of megawatts, not gigawatts — priced first for compliance-sensitive and latency-sensitive workloads. National AI rules begin to bite, making local hosting a procurement requirement in regulated sectors (9). The market projection of African IT load roughly tripling to about 3.5 GW by 2030 implies real but bounded headroom (6). This is when portable firms migrate their proven workflows and capture the latency and compliance dividends.

Beyond 2030: the power-determined era. Whether East Africa gets hyperscale AI compute is decided in the energy sector, not the tech sector: geothermal expansion in Kenya, hydro in Uganda and Ethiopia, grid interconnection, and — possibly the sleeper — data centers acting as anchor tenants that finance new generation, the model African energy analysts argue could make compute demand a grid-building force rather than a grid-eating one (5, 7). The hopeful reading of Olkaria is exactly this: Kenya did not reject the data center; it refused to subsidize it, and in doing so set the precedent that compute must pay for new power rather than cannibalize existing power. Projects structured that way — generation bundled with computation — are the ones to watch.

For builders, each era is a product map. Now: AI services that run lean on APIs and edge models — the entire SME application layer is open. 2027–2030: migration tooling, compliance-as-a-service, and African-language model hosting on regional GPUs. Beyond: the energy-compute interface itself — power purchase structuring, demand aggregation, cooling, and the operational software of African data centers. The factories are coming. The firms that win meanwhile are the ones that never waited for them.

Frequently Asked Questions

What is Cassava’s AI factory?
Cassava Technologies deployed Africa’s first Nvidia-powered AI factory in South Africa in March 2026 — 3,000 GPUs scaling toward 12,000 across planned facilities in Kenya, Nigeria, Egypt, and Morocco. It sells GPU-as-a-service and AI-as-a-service so African firms can rent accelerated computing while data stays under African jurisdiction (1, 2).

Why did Kenya’s $1 billion Microsoft–G42 data center stall?
Power. The Olkaria project sought assurance of up to 1 GW — roughly a third of Kenya’s ~3 GW installed capacity — and asked the government to underwrite the load through long-term power agreements. Officials refused the financial exposure, and transmission to the site was inadequate regardless (3, 4).

Should my SME wait for local African compute before adopting AI?
No. Local compute arrives on infrastructure timelines of five to ten years; your competition moves quarterly. Adopt now with global APIs and small open models, keep your business data in portable formats you own, and migrate to regional providers when they win on compliance, latency, or price.

What does compute sovereignty actually mean for small firms?
Three practical things: data residency (compliance with national data and AI rules), latency (faster real-time services), and eventually cheaper, shilling-priced inference. It does not mean chip independence — Africa’s AI infrastructure runs on Nvidia hardware and global partnerships throughout (8, 9).

How big is Africa’s data center capacity today?
Small but growing fast: under 2% of global capacity, with roughly 1.2 GW of IT load in 2025 projected to near 3.5 GW by 2030. The constraint is electricity — about 600 million Africans still lack power, and grids strain to host gigawatt-scale computing (5, 6, 7).

Related Reading

• DeepSeek is quietly winning Africa: the open-weights vs API decision
• What national AI plans mean for a small firm
• The productive-use solar revolution in East Africa
• What AI actually costs a Kampala SME

Sources and Evidence

1. Cassava Technologies, 2026. “Cassava scales African AI Infrastructure with NVIDIA-Powered AI Factories to accelerate sovereign data capabilities.” https://www.cassavatechnologies.com/cassava-scales-african-ai-infrastructure-with-nvidia-powered-ai-factories-to-accelerate-sovereign-data-capabilities/ — Primary company announcement; source for the AI factory program, GPU counts, and sovereignty framing. Company-issued, so deployment claims are verified against independent trade press (2).
2. Connecting Africa / TechMoran, March 2026. “Cassava deploys ‘AI Factory’ in SA, more countries in pipeline.” https://www.connectingafrica.com/ai/cassava-deploys-ai-factory-in-sa-more-countries-in-the-pipeline and https://techmoran.com/2026/03/19/cassava-technologies-deploys-nvidia-powered-ai-factory-in-south-africa-plans-kenya-nigeria-egypt-launch/ — Independent telecom and tech trade coverage; source for the March 2026 go-live, 3,000 initial GPUs, 12,000-GPU expansion plan, the Kenya/Nigeria/Egypt/Morocco pipeline, and the CAIMEx model exchange.
3. Data Center Dynamics, May 2026. “Microsoft and G42 data center in Kenya stalled due to lack of power capacity.” https://www.datacenterdynamics.com/en/news/microsoft-and-g42-data-center-in-kenya-stalled-due-to-lack-of-power-capacity/ — The data-center industry’s journal of record; source for the suspension of the $1B Olkaria project and its power-capacity cause.
4. Semafor, May 2026. “Energy shortfall ‘problem’ scuppers Kenya’s $1B Microsoft data center.” https://www.semafor.com/article/05/06/2026/energy-shortfall-problem-scuppers-kenyas-1b-microsoft-data-center — International news outlet with sourced reporting on the negotiation breakdown: the 1 GW underwriting demand, the government’s refusal of the PPA exposure, and the transmission constraints at Olkaria.
5. African Business, December 2025. “Inside the race to fire up Africa’s power-hungry data centres.” https://african.business/2025/12/long-reads/inside-the-race-to-fire-up-africas-power-hungry-data-centres — Established pan-African business publication; source for Africa’s sub-2% share of global data-center capacity and the energy-bottleneck analysis, including the anchor-tenant financing model.
6. Mordor Intelligence, 2025. “Africa Data Center Market Size & Share Outlook to 2031.” https://www.mordorintelligence.com/industry-reports/africa-data-center-market — Market research firm; source for the ~1.17 GW (2025) to ~3.46 GW (2030) IT-load projection. Analyst forecasts carry standard uncertainty; used here for order of magnitude.
7. IEA / World Bank, 2025. “Financing Electricity Access in Africa” and “Mission 300.” https://www.iea.org/reports/financing-electricity-access-in-africa and https://www.worldbank.org/en/news/feature/2025/01/13/power-for-progress-a-call-from-african-leaders-and-partners-to-electrify-africa — Institutional sources for the ~600 million Africans without electricity access and the continental electrification agenda.
8. Silicon Canals, 2026. “Africa’s $60B AI sovereignty plan runs directly through 12,000 Nvidia GPUs and Microsoft data centres.” https://siliconcanals.com/sc-d-africas-60b-ai-sovereignty-plan-runs-directly-through-12-000-nvidia-gpus-and-microsoft-data-centres-and-the-contradiction-is-exactly-the-point-everyone-keeps-missing/ — Tech commentary; source for the dependency critique of the sovereignty agenda. Opinion-leaning; used for the argument, with figures cross-checked against (1)(2).
9. Bowmans Law, 2025. “Kenya: Unveiling of the National AI Strategy 2025–2030.” https://bowmanslaw.com/insights/kenya-unveiling-of-the-national-ai-strategy-2025-2030-a-bold-step-into-the-future/ — Leading African law firm’s regulatory analysis; source for the national AI strategy and the direction of data-governance requirements affecting firms.