A Lagos startup founder opens her laptop, types a prompt into an AI model and gets an answer in seconds. The experience feels instant and local. It is neither. The prompt travels to servers that are, in all likelihood, thousands of kilometres away. The chips inside them were designed and bought abroad. The bill arrives in US dollars, and the company that sends it is headquartered somewhere else.
As generative AI spreads across African offices, classrooms and clinics, a question that once belonged to economists is becoming a business problem: when Africans use AI products built elsewhere, where does the economic value go, and what capability stays behind in African institutions?
This investigation looks at four pieces of that question: who owns the infrastructure, what African businesses pay for outside services, whether local universities and startups can get affordable computing power, and what happens to a small firm when a foreign provider changes its price or its terms. The honest summary is that the evidence on some of these points is strong, and on others it barely exists.
A warning from the UN, and a missing number
The UN Economic Commission for Africa (ECA) released a position paper on responsible AI governance this month, centred on putting the African Union’s Continental AI Strategy into practice. According to summaries of the paper, it proposes readiness evaluations, data-sharing systems and ethical frameworks, and it calls for investment in infrastructure and education to build institutional capacity.
That emphasis matches the AU strategy itself, which says AI development rests on reliable electricity, broadband, data centres, cloud and computing power, along with large quantities of quality data and skilled people. None of these arrives automatically with a downloaded app.
The ECA is also building the measuring stick that the debate lacks. In a public commitment registered with the PARIS21 data community, ECA says Africa is entering the AI era with uneven readiness, citing weak infrastructure, fragmented data systems, thin public-sector skills and dependence on imported platforms. It plans to map compute resources, language resources and centres of excellence, and to publish a first African AI Readiness and Maturity Index between January and June 2027.
That timeline matters for this story. Today, no public source tells us how much money leaves Africa each year for foreign AI services. Cloud providers do not publish country-level AI revenue, and national statistics rarely separate AI subscriptions from other software imports. What follows is therefore built from surveys, industry reports and company disclosures that point in one direction, not from a single authoritative total.
Who owns the infrastructure?
The starting fact is scarcity. A February 2026 study by the Africa Data Centres Association and Rising Advisory put Africa’s active data centre load at 360 megawatts, with 238 megawatts under construction and 656 megawatts planned. If every project is completed, capacity reaches about 1.2 gigawatts by 2030. Even then, the continent’s share of global capacity would stay at roughly 0.6 percent, because investment is accelerating just as fast in the United States, Europe and Asia. Africa is home to roughly 18 percent of the world’s population.
Ownership is mixed, and it is shifting. Another study counted 211 active data centres in Africa, with 46 percent of them in South Africa, Kenya, Nigeria and Egypt. Hyperscale sites in South Africa are run by Amazon Web Services, Microsoft Azure and operators such as Teraco, alongside Vodacom Business, Equinix and NTT. Google opened a Johannesburg cloud region in early 2024. Microsoft and the UAE-based AI company G42 announced a $1 billion digital investment initiative in 2024 centred on a geothermal-powered campus at Olkaria in Kenya.
The picture is not simply American. In April 2026 Huawei released a reference design for AI data centres in Northern Africa. This August, MTN announced a partnership with a UAE-backed data-centre investor to build AI-ready infrastructure across the continent. Analysts at Further Africa describe a contest that is no longer just US versus China, with Gulf investors supplying capital and expertise and African operators trying to secure a place in the deal flow.
African-owned capacity exists and is growing. Nigeria’s Galaxy Backbone, Ghana’s National Data Centre, Rwanda’s national data centre and state-backed projects in Ethiopia and Togo aim to keep sensitive public data at home. Cassava Technologies, working with NVIDIA, has built what it calls Africa’s first multi-site AI factory. Yet a telling detail sits in the ADCA report: outside South Africa, only about a third of built capacity is in use. Buildings are going up ahead of demand from hyperscale and AI customers, and the main brake on new sites is now electricity, not fibre.
The practical conclusion is uncomfortable. Even when a data centre stands on African soil, the AI model running inside it, the chips and often the contract belong to a company outside the continent. A building in Nairobi or Johannesburg is not the same thing as owning the AI stack.
Where the money goes
No one can yet say how many dollars African firms spend on external AI services. The best available evidence comes from the people paying the bills.
A September 2026 study by GIZ, the German development agency, surveyed 85 African AI businesses between April and July and mapped 311 in total. It found that 65 percent name cloud and compute as their single highest cost, ahead of buying data. It also found that 56 percent price their own products below $2,500 a year per customer. The report describes a structural asymmetry: these companies carry the cost of internationally priced infrastructure, including APIs, GPUs and cloud, while selling into lower-income markets. It concludes that the vast majority of AI computation for these firms happens on foreign servers, paid for in foreign currency.
The pressure is visible beyond startups. A Backbase and African Banker survey of 277 senior banking executives across 37 countries, with fieldwork in the second quarter of 2026, says foreign-exchange pressure, rising dollar-denominated cloud costs and tightening data-localisation rules are forcing boards to demand proof of return on AI spending.
The demand side is still modest, which means the outflow could grow sharply. Microsoft’s AI Economy Institute found that 23.1 percent of working-age people in South Africa used generative AI in the first quarter of 2026. Nigeria and Ghana stood at 10.1 percent and Kenya at 8.7 percent, against a global average of 17.8 percent. Most of that usage runs on tools hosted abroad. If Nigeria reached South Africa’s level of use on the same imported terms, the recurring dollar bill would rise with it.
There is also a payments layer. Because subscriptions are priced in dollars and many African cardholders face international spending limits, local intermediaries have appeared. Techpoint Africa reported on Mercurie, a platform that lets businesses pay for Google Cloud, Workspace and other global software in local currency. It is a useful fix, and also a sign of how much friction sits between an African buyer and a foreign provider.
The compute queue
For researchers and young companies, the sharpest form of dependence is access to processing power.
In an analysis published through UNDP, the founders of Alliance4AI and Zindi, two African AI talent networks, estimated that only five percent of the continent’s AI talent has access to the compute needed for serious research. Another four percent can pay for GPUs in the cloud. They also described African users waiting for newer chips until American demand eases, and they calculated that the same $1 million in cloud credits supports far fewer startups than an on-premise subsidy would. Alliance4AI founder Tsado also runs Udutech, a GPU rental platform, so readers should weigh the proposed fixes with that interest in mind. The underlying scarcity, though, is echoed by the GIZ survey, in which 64 percent of founders ranked R&D and compute subsidies among their top needs.
There are real signs of change. In March 2026 the University of Cape Town announced the African Compute Initiative, which it describes as Africa’s largest GPU-intensive cluster dedicated to AI in higher education. It will be paired with cloud access through the Mozilla Foundation, powered partly by solar, and designed so researchers can log in with their home-institution credentials across partner universities. UCT says it will also produce a toolkit so other African universities can copy the model. UCT’s Jonathan Shock said African researchers have the ideas and the talent but have lacked the computing power.
In Lagos, Udutech launched an Africa GPU Hub in August 2025 that rents GPUs for less than $1 an hour, according to TechCabal, with subsidised access for students and innovation hubs. A new AISCA Foundation in Kigali is targeting subsidised compute and skills. And GIZ finds that startups’ choices of AI tools are driven mostly by cost, not by preference: early on, firms split among building their own models, using open-source models and calling third-party APIs. Only after raising roughly $1 million to $10 million do many move to hybrids that layer proprietary data and fine-tuning on open-source foundations.
The pattern is encouraging but thin. Most of these facilities are new, small relative to demand, or still at the announcement stage. The test will be whether a final-year student in Kano or Kumasi can book time on them as easily as a researcher in Cape Town.
When the vendor changes the deal
The fourth question is the least documented and probably the most urgent for small firms.
Commentators in Nigeria’s BusinessDay have warned that companies earning in naira, cedi or shillings while paying metered AI fees in dollars face a double exposure: a currency slide and a jump in usage can hit the same budget at once. One column pointed to reports that Uber used up its entire 2026 AI budget by April after giving thousands of engineers unrestricted access to coding assistants. We could not independently confirm that claim, but the columnist’s point stands. If a company with Uber’s resources can misjudge usage-based AI pricing, a Lagos logistics startup with no venture cushion has less room for error.
Model providers also retire older products and revise pricing on their own schedules, and industry writers note that standard vendor agreements rarely promise a minimum notice period or compensation for migration costs. For a team of five engineers, rewriting prompts and retesting a product after a model is withdrawn can consume weeks. We did not find a documented African case in which a small business was forced to close or re-price after a provider’s change, and that gap itself is worth reporting further.
A recent episode shows that terms can also change for reasons unrelated to price. On June 12, 2026, Anthropic suspended access to its Claude Fable 5 and Claude Mythos 5 models to comply with US Department of Commerce export controls. The department lifted the controls on June 30 and access was restored on July 1, according to Anthropic’s public statement. It was brief, and we found no data on how African customers were affected. But it showed that availability of a foreign AI service can depend on decisions made in another capital.
The defences discussed most often are practical: build software so that one AI vendor can be swapped for another, track token spending as closely as payroll, and keep at least some critical functions on open or locally hosted models. The GIZ finding that African firms already lean on open-source foundations as they grow suggests many are heading that way, mostly because they must.
What stays, and what leaves
Putting the evidence together, the value flow looks like this. Subscription and usage fees leave in dollars. Most of the compute, and almost all of the frontier model development, sits outside the continent. What remains in Africa is real but narrower than the headlines imply: customer relationships, local datasets, domain knowledge, a fast-growing developer base (the Boston Consulting Group estimated 4.7 million African developers in 2024, growing about 21 percent a year from 2019), and a first generation of AI companies. GIZ counts about $1.3 billion raised by African AI businesses between 2014 and 2026, with a median round of $225,000, and notes that Africa holds only one to three percent of the global AI talent pool.
Capabilities that tend to stay in a country are those it builds on purpose: its own compute, trained engineers, high-quality local data and the contract leverage that comes from having alternatives. The ECA paper, the AU strategy and the new compute projects all point at those four things.
What to watch
Several markers will show whether the balance is shifting. Watch whether data centre occupancy outside South Africa rises from the current one-third as AI demand arrives. Watch whether ECA’s readiness index, due in the first half of 2027, finally publishes compute and spending data country by country. Watch whether public universities beyond Cape Town secure guaranteed GPU time, and whether governments use their own purchasing to anchor local providers. And watch the plain arithmetic of the exchange rate, because for most African AI users that is where dependence is felt first.
Methodology note
This article draws on published reports, company announcements and press coverage. We reviewed summaries of the ECA position paper but not its full text, so its specific recommendations should be checked against the original document. No source provided a continental total for African spending on foreign AI services, and we have not estimated one.
Sources
- UN Economic Commission for Africa position paper on responsible AI governance, reported by Africa.net, October 7, 2026: https://af.net/realtime/un-economic-commission-for-africa-releases-position-paper-on-responsible-ai-governance/
- ECA, African AI Readiness and Maturity Data Stories (PARIS21 commitment): https://ct.paris21.org/en/commitments/african-ai-readiness-and-maturity-data-storiessce96l80sadovlsvpn84uwmc0yplnmipi7mqmadm
- African Union Continental AI Strategy (July 2024): https://au.int/sites/default/files/documents/44004-doc-EN-_Continental_AI_Strategy_July_2024.pdf
- Africa Data Centres Association and Rising Advisory, Data Centres in Africa 2026, as reported by Kapitalis and Techeconomy: https://kapitalis.com/tunisie/?p=18404793 and https://techeconomy.ng/africa-data-centre-capacity-0-6-percent-ai-1-2gw-2030
- Ecofin Agency, Africa’s Big Four host 46% of the continent’s data centers: https://www.ecofinagency.com/news/0309-48382-africa-s-big-four-host-46-of-the-continent-s-data-centers-study
- Further Africa, US and China race for Africa’s next strategic infrastructure (August 31, 2026): https://furtherafrica.com/2026/08/31/us-and-china-race-for-africas-next-strategic-infrastructure-data-centres/
- GIZ, AI Made in Africa: Voices from AI businesses across Africa, executive summary (September 2026): https://www.giz.de/sites/default/files/media/els-document/2026-09/executive-summary-ai-made-africa.pdf
- Backbase and African Banker, The State of AI in African Banking 2026: https://research.backbase.com/research/state-of-ai-african-banking-2026
- Microsoft AI Economy Institute, Global AI Diffusion Q1 2026, as reported by Ecofin and Telecom Review Africa: https://www.ecofinagency.com/news/2205-55838-south-africa-tops-african-generative-ai-adoption-rankings-in-2026-microsoft-report and https://www.telecomreviewafrica.com/?p=28781
- BusinessDay, columns on AI costs and foreign-exchange risk: https://businessday.ng/?p=1138509 and https://businessday.ng/?p=1141163
- Techpoint Africa, Mercurie: https://techpoint.africa/news/former-google-executive-launches-mercurie/
- UNDP, Only five percent of Africa’s AI talent has the compute power it needs: https://www.undp.org/node/493186
- University of Cape Town, African Compute Initiative announcement (March 26, 2026): https://www.news.uct.ac.za/article/-2026-03-26-uct-to-lead-africas-first-higher-education-dedicated-ai-compute-initiative
- TechCabal, Udutech Africa GPU Hub: https://techcabal.com/?p=171314
- Anthropic, statement on Fable and Mythos access: https://www.anthropic.com/news/fable-mythos-access
