Artificial Intelligence has dominated many major development conferences this year, and it will likely be a key topic at the upcoming EU-AU Summit. As the Global North pours billions into AI models and data centers—big tech alone is set to spend over $320 billion in 2025—African policymakers are asking whether to follow suit. This blog argues that while large-scale infrastructure will take time to build, there are high-impact, lower-hanging opportunities they can act on now.
The case for large-scale AI infrastructure in Africa is far from clear-cut. Some argue that countries can simply rely on cloud services hosted in the Global North, but leaders like South Africa’s Blade Nzimande and Kenya’s Eliud Owalo warn that this deepens dependence on foreign technology. The term “digital sovereignty” is growing in popularity, as many call for greater control over data, compute, and infrastructure. The concern is understandable: Africa hosts just 1 percent of global data center capacity despite nearing a quarter of the world’s population. Recent events have also underscored these risks—earlier this year, Elon Musk’s musing on X that he could cut Starlink access to Ukraine’s armed forces highlighted the dangers of reliance on foreign technology providers. No nation wants its critical infrastructure(whether that’s internet or AI systems) at the mercy of external powers.
However, building the infrastructure for sovereign AI is a monumental task. Beyond the large capital expenditures, these facilities require a constant, reliable power supply. While there is early exploration of whether data centers could drive electrification by offering long-term contracts that incentivize utilities to expand capacity, the continent remains years away from meeting the necessary demands. Beyond the issue of unreliable power, the continent faces low-internet usage. These challenges don’t mean African countries shouldn’t pursue sovereign AI— that choice requires more careful analysis. It does however suggest that if governments are concerned about being left behind, they should consider complementary, high-impact investments that can deliver results more quickly.
Here are two actionable ideas that leverage existing infrastructure:
1. Deliver generative AI over SMS
Most consumer-facing generative AI applications today are some sort of chatbot. While users in high-income countries typically access them through apps like ChatGPT or web browsers, that need not be the case. Chatbots can also run over SMS, reaching feature phones already used by low-income households across Africa. Such systems require no new internet connections or smartphones—an important consideration given that, according to ITU data from 2023, 63 percent of Africans own a mobile phone, but only 37 percent use the internet.Delivering chatbot interfaces to millions of households over SMS is technically simple, but prohibitively expensive due to steep SMS price markups. As Dan Björkegren notes in a recent CGD blog, sending a single SMS can cost as much as sending 58,775 similarly sized messages via mobile data. He estimates that the median global markup on SMS for non-internet phones may reach 9,000 percent, meaning operators could theoretically deliver over 90 texts for the price of one. Stripped of these markups, running the AI itself is inexpensive: Björkegren calculates that 100 interactions cost just $0.09 in compute, compared to $5.32 to transmit them over SMS.
The bottom line is that AI can reach scale—extending to the 1.7 billion phones without internet access—without any new infrastructure investment. What it requires is for Mobile Network Operators (MNOs) to lower the cost of SMS. This is where initiatives like the Africa–Europe Digital Regulators Partnership, and especially its ICT Policy & Regulation Institutional Strengthening (iPRIS) program, could help. iPRIS could explore policy tools and align MNO incentives to make SMS more affordable. The program already offers a model: through the West Africa Telecommunications Regulators’ Assembly, iPRIS supported the ECOWAS Regional Free-Roaming Regulation, which reduced communication costs across nine countries. Similar regulatory cooperation could do the same for SMS—unlocking AI’s potential to turn existing mobile networks into engines of innovation and inclusive development.
2. Deliver generative AI over voice calls
While lowering the price of SMS can increase AI access, roughly one in three adults in Africa lack basic reading and writing skills. AI over regular voice calls could be even more inclusive and allow for richer engagements with users. Like SMS, delivering AI services over a voice call does not require users to use a smartphone or access the internet. A recent study led by Chicago Booth's Brian Jabarian conducted in the Philippines highlights the immense potential of AI over voice to perform complex, human-like interactions that increase firm productivity. In his field experiment on AI-led job interviews, voice-based AI not only replicated but even outperformed human interviewers. Applicants interviewed by AI had a 17 percent higher one month retention rate.
These voice based AI services however work best in high-resource languages. Word-error rates, a common metric to measure the performance of automated speech recognition systems, are below five percent for languages such as English and Spanish but rise to 30–50 percent for under-resourced languages such as Swahili and Tamil. Although specialized fine-tuned models exist on the market for some under-represented languages, developing and maintaining them typically requires substantial labeled data and significant computational investment. As the most linguistically diverse region in the world, Africa is home to thousands of minority languages. Insights from our work through the AI for Global Development Accelerator, in collaboration with the Agency Fund and OpenAI, suggest that current major speech-recognition models still underperform on minority and low-resource languages, requiring additional effort to collect local training data and fine-tune models before they can be deployed effectively.
Jabarian’s paper however does suggest that one can get productivity gains in low- and middle- income countries today, through voice based AI products that operate in high-resource languages. While running a voice based AI intervention in English or French won’t reach the poorest across the African continent, stimulating the development and deployment of voice based tools in these languages is within reach. There is also exciting early work being trialed in Rwanda, where MTN and Viamo are partnering on a new product called Miss Baza, which aims to enable users to access services in English, French, and Kinyarwanda.
Initiatives like the EU’s Africa Connected Programme could accelerate this progress by investing in African companies developing datasets, automated speech recognition models, and other voice based models, in addition to their investments on hard infrastructure. The fund could adopt a two prong approach—starting first with investing in voiced based solutions in High-Resource Languages, while simultaneously funding the development of technology for underrepresented languages.This could help stimulate more partnerships such as the MTN and Viamo collaboration in Rwanda.
So much of the discussion about what to do about the growing AI divide has centered on increasing physical infrastructure on the African continent. While important, such investments take a long time to yield benefits, and we are probably paying insufficient attention to how we can leverage infrastructure like SMS and voice based phone lines that are available today.
While both channels—SMS and voice—focus on reaching the most vulnerable in modalities they can afford, the policy levers for each one differ. Making generative AI affordable over SMS may require regulatory action, while delivering AI over voice would benefit from funding and private sector investment.
Policymakers at the upcoming AU-EU summit should consider these approaches as part of their toolkit to close the widening AI divide.
