advertisement
Built In Africa, Trained Abroad: The Cost Of AI Computing
African AI developers are increasingly building solutions around local languages, industries and business challenges. However, while the talent and ideas may originate on the continent, much of the computing power required to develop and train these models still comes from outside Africa. The result is an AI industry that is locally built but heavily dependent on foreign infrastructure.
When Johannesburg-based Vambo AI set out to develop MORENA, a 1.5-billion-parameter language model supporting 12 African languages, it had to look beyond the continent for the computing resources needed to train it.
The training process used approximately 22,000 hours of computing on Nvidia A100 chips in Italy, at an estimated value of between $25,000 and $40,000, according to TechCabal. Vambo did not pay for the computing time directly, as access to the CINECA supercomputer was provided through the UNDP AI Hub for Sustainable Development.
advertisement
The arrangement highlights a challenge facing African AI developers: access to affordable computing infrastructure remains limited, often forcing companies to depend on foreign cloud providers, international research partnerships and donated computing credits.
For companies without such support, computing costs can become a significant barrier to growth. A survey published in October by GIZ’s AI Made in Africa initiative and Briter Bridges, covering 85 AI companies across the continent, found that 65 per cent identified cloud computing, GPUs and other AI infrastructure as their largest development expense. Another 61 per cent cited high cloud costs as a major bottleneck.
Shina Arogundade, chief executive of Nigerian health technology company MyItura, which has relied on cloud credits to train and operate its models, described the financial pressure. “Developing AI products, testing them, and scaling them cost a lot on a monthly basis,” he said.
advertisement
The challenge has become increasingly relevant as discussions at Africa Tech Festival in Cape Town focus on expanding local computing capacity and developing a more self-sufficient AI ecosystem.
One obstacle is that much of Africa’s existing data centre infrastructure was not designed for the demands of modern AI workloads. According to analysis published by TechAfrica News, conventional facilities typically support around 5 kilowatts of power per rack, while AI applications can require between 20 and 50 kilowatts, with some approaching 100 kilowatts.
Investment in new capacity is also complicated by relatively slow market uptake. The Africa Data Centres Association’s 2026 economic report found that data centres in African markets can take six to eight years to reach 85 per cent occupancy, compared with two to three years in Europe. This creates a cycle in which investors are cautious about expanding capacity, while developers struggle to scale because suitable infrastructure is unavailable.
advertisement
Consequently, an estimated 80 to 90 per cent of Africa’s data is still processed and stored outside the continent, in established technology hubs such as Dublin, Frankfurt and Amsterdam.
Another factor keeping African AI developers on international platforms is access to free cloud credits. Amazon Web Services, Microsoft Azure and Google Cloud offer start-ups computing credits directly or through accelerator programmes, allowing young companies to develop products without immediately absorbing the full infrastructure costs.
Vambo AI’s access to Italy’s supercomputer and Lelapa AI’s use of computing credits from Microsoft’s AI for Good Lab illustrate how such arrangements support local innovation. However, they can also create long-term dependencies, as companies build their applications, systems and technical expertise around platforms they may eventually struggle to afford.
Currency volatility makes that transition even more difficult. Most international cloud services are priced in US dollars, leaving African businesses exposed to exchange-rate movements even when their revenues are generated in local currencies.
Nigeria provides a clear example. The naira depreciated from approximately ₦471 against the dollar in 2023 to ₦1,534 by August 2025. BusinessDay calculated that a $1,000 cloud service that previously cost ₦471,000 had risen to approximately ₦1.57 million, representing an increase of more than 230 per cent for the same service.
AWS began accepting naira payments in January 2025, but the underlying pricing remains dollar-based and is converted at prevailing exchange rates. Stephen Okoye, chief operating officer of Lagos-based Nobus Cloud, said his company’s prices remain “about 30-35% lower than what AWS charges.”
Beyond cost, offshore computing raises questions about data governance, regulatory compliance and operational control. Organisations hosting sensitive information abroad must consider the jurisdictions governing that data, while start-ups remain exposed to pricing changes and the commercial decisions of international cloud providers.
A growing number of African technology companies are attempting to address this gap by bringing GPU capacity closer to local developers.
Udu Technologies, founded by former Nvidia cloud executive Alexander Tsado, launched its Africa GPU Hub in September, offering access to Nvidia H100 and Blackwell computing clusters in markets including Kenya, Rwanda, South Africa, Togo and Zambia. The company advertises access at prices starting below $1 per hour and says it has served approximately 2,000 developers.
In Nigeria, Chassis, a cloud GPU start-up founded in 2025 by 18-year-old programmer Okechukwu Nwaozor, is providing locally hosted computing resources to AI developers.
Other initiatives are also emerging. Data Science Nigeria announced free GPU access for researchers and start-ups in June, while the country’s National Information Technology Development Agency (NITDA) and state-owned cloud provider Galaxy Backbone opened government cloud infrastructure to start-ups in May.
For CIOs and enterprise technology leaders, these developments introduce another consideration when selecting infrastructure providers. Comparing cloud services based solely on hourly GPU rates may overlook the wider costs associated with currency fluctuations, data transfers, latency and regulatory requirements.
Local computing capacity could help reduce these dependencies, but its long-term viability will depend on whether African enterprises, governments and research institutions are willing to commit workloads and sustained spending to locally operated infrastructure.
Tsado has described GPUs as the refineries of the AI age, an analogy that resonates with Africa’s longstanding experience of exporting raw resources while importing higher-value finished products.
The same pattern risks emerging in artificial intelligence. Africa has the talent, languages, data and growing pool of developers needed to build relevant AI applications, but much of the computing infrastructure that turns those resources into products remains elsewhere. Bringing more of that processing capacity home will be essential if the continent hopes to retain a greater share of the value generated by its AI industry.