Article
Africa’s AI strategy doesn’t have to be borrowed
AI series · Part 1 of 11
Africa is being sold a model of sovereign AI built for Europe, Britain and North America. The more useful question is whether its strategy has to look the same.
That model assumes Western demand: large clusters and frontier models serving software and office work, run on hyperscaler clouds. Copy it, and the outcome is predictable. Host countries supply the site and the electricity, foreign firms supply the processors and the cloud platform that customers pay for, and most of the margin leaves.
Kenya shows the cost. Microsoft and G42 announced a US$1 billion geothermal-powered data centre at Olkaria in May 2024; by May 2026 it was on hold, after the companies sought a state guarantee on part of its capacity. President William Ruto said running it would mean shutting down half the country. Analysts countered that its first phase needed about 100 MW, not the 1,000 MW he cited. Either way, the risk was being pushed onto the host.
A strategy built around Africa’s own uses, data and climate would start somewhere else: smaller models in local languages, compute sited where power is surplus and the air is cool, and ownership of the layers local users pay for. Value capture then follows from the design instead of being haggled over afterwards, and the trade-offs become choices governments make on purpose.
The story African governments are being asked to buy
Sovereign AI arrives in Africa packaged as a strategy. Accenture’s November 2025 report, typical of the genre, tells chief executives and governments to treat sovereignty as a source of growth rather than a compliance cost. Its own survey shows how thin that growth is: 46% of organisations cite compliance as their reason for seeking sovereign AI, and only 13% cite making money or tailoring AI to local value.
Read from Addis Ababa, Lagos or Nairobi, the problem is the author: the firms that sell chips, cloud and consulting, writing for a government they treat as a customer rather than a sovereign. It promises control while leaving the value, and the decision about what AI is for, with the seller.
The stakes are high because Africa starts from so little: about 0.6% of the world’s data-centre capacity against roughly 18% of its people. Almost all of its AI infrastructure is still to be built, on terms that will last decades, and national AI strategies and COP32 in Addis Ababa will keep the offers coming.
The players: who wins and who is at risk
The winners so far are the chip-makers and cloud providers, who keep the highest-margin layers wherever the building sits, and the Gulf investors financing a growing share of the build. A few African operators with their own power and sites, led by Strive Masiyiwa’s Cassava, are winning too. The risk falls on governments trading cheap power and land for a ribbon-cutting, and on grids asked to serve a single giant customer.
A government that supplies cheap power and land but takes no equity, no data rights and no share of cloud revenue is a landlord, not an owner.
Africa’s AI strategy does not have to copy the West
The building blocks are the same everywhere. What Africa needs them for is not, and that changes what is worth building and owning.
Africa does not use AI the way the US does
The AI being sold to Africa was built for Western demand, much of it coding and office work. Anthropic’s January 2026 Economic Index finds that coursework makes up a larger share of AI use in lower-income countries, while work and personal use dominate in richer ones. Africa’s needs are different again: services in local languages, farm advice, health triage and public administration.
Those needs do not require frontier-scale compute. Lelapa AI’s InkubaLM, a 0.4-billion-parameter model for five African languages, matches much larger models on some tasks. A sovereign agenda could put scarce GPUs into small models trained on African data, kept on African soil.
Africa’s data rules are a head start
Africa has agreed continental data rules before most of its AI infrastructure exists. The African Union adopted a Data Policy Framework in 2022, and in December 2025 the AU Commission validated follow-on frameworks for data sharing, cross-border flows and open data, aimed at a digital single market by 2030. Most large economies, including Nigeria and Kenya, now have data-protection laws of their own.
Rules agreed early can shape infrastructure instead of being retrofitted onto systems built for someone else. A shared, African-governed pool of training data would let countries train models at a scale none could reach alone. The gap is execution: enforcement is national, and no shared facility exists yet.
Africa’s climate is an advantage
North American data centres often sit in hot, dry places and use large amounts of water for cooling. Highland cities such as Addis Ababa, at about 2,350 metres, are cool year-round. The East African Rift offers geothermal baseload, Ethiopia and the DRC have large hydro resources, and solar is abundant across the Sahel and the south.
Coming second helps too. Western operators are retrofitting halls built for older chips; African facilities can be designed for today’s hardware, renewable power and water-light cooling from day one. The chips cost the same; the mistakes need not.
An advantage can also be abused. Cheap hydropower has already been exported as bitcoin mining with little left behind, and a data centre that takes a community’s water or land without consent is extraction under a new name. The test is whether a facility leaves the grid, the watershed and its neighbours better off. Regenerative Africa is developing an AI infrastructure guide built on that test.
Africa has its own players, and most are already live
The strategy handed to African governments assumes the hyperscalers are the only serious suppliers. They are not. African firms now operate at most layers of the stack. Most are live, but not yet at national scale.
They are ready now for regulated public data, local-language tools and pilot-scale compute, which is where sovereignty matters most and the hyperscalers’ advantage is smallest. To carry national-scale AI they need governments as anchor customers. That is a policy choice, not a market accident.
Four countries are ready to start now. Kenya, Nigeria, South Africa and Egypt all have data-protection laws and African-owned AI capacity running: Nairobi’s sovereign public cloud, the sovereign fund-backed Kasi campus in Lagos, Cassava’s AI Factory in South Africa and Elsewedy’s stake in Egypt’s first sovereign AI data centre. Each has a catch: Kenya’s grid suits phased builds, Nigeria’s unreliable power raises costs, and South Africa’s local players compete with all three hyperscalers.
Ethiopia has the power surplus and the cool highland climate, but not yet a local cloud layer to run on them. That is the gap to close first.
A sovereign can only own an agenda that pays it back
When the returns from AI are captured abroad, a government’s influence stops at the permit: it can set rules, but not priorities. Owning the agenda means owning at least one paying layer, whether equity in the facility, a share of the cloud, or the models and data local users depend on.
A sovereign AI strategy that serves the sovereign
A strategy that serves the country starts from what the country needs AI to do, not from what a vendor needs to sell. It sizes compute to those uses, rather than to a hyperscaler’s minimum order.
It treats data as a national asset, with clear terms for who may use it. It requires a data centre to add power to the grid rather than draw it away. It decides deliberately which layers to own, which to share with neighbours and which to rent. And it weighs every concession, from a revenue guarantee to a tax holiday, against what stays in the country over the life of the deal.
What the shift is worth: a 100 MW facility in Kenya
A sovereign design would lift Kenya’s share of a 100 MW AI data centre, the first phase planned at Olkaria, from about 12% of its revenue to about 20%: roughly US$80 million more a year. At 2026 contract prices for rented GPUs, such a facility would earn about US$1 billion a year.
Show the data as a table
| Revenue layer | Total | Kept, copied | Kept, sovereign |
|---|---|---|---|
| Electricity (state utility) | 126 | 126 | 126 |
| Facility rent | 180 | 0 | 54 |
| Cloud operator’s margin | 123 | 0 | 25 |
| GPU hardware | 600 | 0 | 0 |
| Total | 1,029 | 126 (12%) | 205 (20%) |
The biggest leak is one no African country can plug yet: almost 60% of the revenue pays for hardware made abroad, which is why chip design belongs in a long-term strategy. The result is sensitive to price. At US$3 per GPU-hour, Kenya’s sovereign take rises to about US$262 million, double the copied model. Local providers would need capital: a fifth of the compute means owning about 14,000 GPUs, roughly US$600 million of hardware.
We used public data wherever it exists and say where we have had to assume.
| Input | Value | Basis | Public data or assumption |
|---|---|---|---|
| Size | 100 MW of IT load | Planned first phase at Olkaria | Public |
| Power price | US$0.11 per kWh | KSh 10 tariff for special economic zones (EPRA, 2023), plus 2026 fuel and forex charges, at KSh 130 to the dollar (RA intelligence model, 31 October 2026) | Public |
| GPU count | About 71,000 H100-class | 10.2 kW per eight-GPU server (Nvidia), plus about 10% for networking and storage | Public, with an allowance |
| Rental price | US$2.35 per GPU-hour | One-year H100 contract rate, March 2026 (SemiAnalysis) | Public |
| Hardware cost | US$42,000 per GPU, over five years | Value of H100 servers (AM Compute, October 2026) | Public |
| Cooling efficiency | Power usage effectiveness of 1.3 | Achievable in new liquid-cooled halls; the global average has stalled well above this (Uptime Institute) | Assumption |
| Utilisation | 70% | GPU capacity sold out in 2026, lifting contract rates; 70% allows for ramp-up and churn | Assumption |
| Facility rent | US$150 per kW per month | No public wholesale price for Kenya; the bottom of the US$150–400 range for GPU-ready space, since one tenant takes 100 MW | Assumption |
| Local equity | 30% | KenGen, which is 70% state-owned, owns the Olkaria site, so land and grid access could be traded for equity; below the 40% Ethiopian Investment Holdings holds in Dangote’s Gode plant | Assumption |
| Local share of compute | 20% | Kenya’s 2021 data regulations require health, basic education, public finance, civil registration and election data to be processed in Kenya; no public figure shows their share of AI demand | Assumption |
The figures are revenues, not profits, and leave out wages and taxes, which a facility pays under either model.
What to watch over the next 12 months
- Whether the Microsoft and G42 Olkaria project returns in a smaller, phased form, and on what payment terms.
- How fast Cassava’s 12,000-GPU plan reaches Nigeria, Kenya, Egypt and Morocco, and who buys the capacity.
- The terms of MTN’s 150 MW joint venture: who holds the equity, and where the power comes from.
- Whether Ethiopia shifts surplus power from bitcoin mining to AI compute, and whether it takes equity this time.
- How COP32 in Addis Ababa (2027) frames AI power demand against the energy-access agenda.
The subscriber edition
The Regenerative Africa Intelligence subscriber edition complements this article with detailed deal data: every announced AI and data-centre deal in Africa, with its owners, financing, power terms, capacity and status, scored for how much value stays in the host country. This is the first of eleven pieces in RA’s AI series; the next maps who is building what, where. Elim Shanko will be discussing these questions at AfroVest in November.
There is no single template: the right answer depends on each country’s grid, data laws, budget and neighbours. Regenerative Africa works with governments and investment promotion agencies to build that answer and to negotiate the deals that follow. If you are shaping an AI strategy or weighing a data-centre offer, talk to us.
Question for readers: if your government could secure one term in an AI data-centre deal, would it be equity, power that builds the grid, or reserved capacity for local users?
Sources
Accessed 26 October to 2 November 2026. Primary sources (company, government and research publications) are listed first; trade and news reports are used where no primary source is public.
Primary sources
- Accenture, Sovereign AI: Own your AI future (Nov 2025)
- Cassava Technologies, Cassava scales African AI infrastructure with NVIDIA-powered AI Factories
- Nigeria Sovereign Investment Authority, Kasi Cloud commissions West Africa’s first hyperscale-ready, AI-capable campus (May 2026)
- Lelapa AI, InkubaLM: a small language model for African languages
- African Union Commission, validation of continental data governance frameworks (Dec 2025)
- African Union, Data Policy Framework implementation forum
- Regenerative Africa Intelligence Model, USD/KES weekly close, 31 October 2026
- Anthropic, Economic Index report: economic primitives (Jan 2026)
- Microsoft, Microsoft and G42 announce US$1 billion digital ecosystem initiative for Kenya (May 2024)
- Office of the Data Protection Commissioner, Kenya, Data Protection (General) Regulations 2021, regulation 26
- Nvidia, DGX H100 system specifications
- Uptime Institute, Global Data Center Survey 2025
- Regenerative Africa, Ethiopia opens up: three ways in (Oct 2026), citing Ethiopian Investment Holdings, for the Gode joint venture
Trade, research and news reports
- Baker McKenzie, The megawatt and the token: Africa’s compute equation (Jun 2026)
- ThinkGeoEnergy, Kenya suspends plans for Microsoft’s geothermal-powered data centre (May 2026)
- TechCentral, Setback for Microsoft’s Africa cloud ambitions (reporting Bloomberg)
- Tech-ish, Kenya’s Microsoft-G42 data centre stalled, on a possible 100 MW first phase (May 2026)
- Techweez, on geothermal’s share of Kenya’s power (May 2026)
- Connecting Africa, Cassava deploys AI Factory in South Africa (Mar 2026)
- TechAfrica News, iXAfrica and Baobab Cloud launch sovereign public cloud in Kenya (May 2026)
- BusinessDay, UniCloud and OADC sovereign cloud deal
- TechTrends Kenya, Inside AfriQloud’s first sovereign deployments (May 2026)
- TechEconomy, ChipMango raises $1.9m (Sep 2026)
- TechAfrica News, Rwanda signs MoU with ChipMango (Sep 2026)
- DataCenterDynamics, MTN forms data centre partnership with UAE-backed investor (Aug 2026)
- TechAfrica News, Egypt partners with Vodafone, Elsewedy and Cassava on $1bn data centre investment (Sep 2026)
- Morocco World News, Egypt’s second national AI strategy (Jan 2026)
- Addis Fortune, bitcoin mining’s power use in Ethiopia
- The Standard, Ruto’s claim on data-centre power consumption untrue, say analysts (Nov 2025)
Kenya sizing inputs
- Business Daily, EPRA approves KSh 10 tariff for special economic zones (Mar 2023)
- Eastleigh Voice, EPRA fuel cost charge for June 2026
- BusinessFront, EPRA April 2026 tariff review, forex charge
- IntuitionLabs, cloud GPU rental prices, citing SemiAnalysis one-year H100 contract rates (Sep 2026)
- AM Compute, H100 HGX server values (Oct 2026)
- AI Data Centers Directory, GPU colocation pricing (Feb 2026)
