This month in brief
In September, the African Union consulted on draft AI guidelines and a model AI law, while IGAD and COMESA advanced their own regional AI frameworks. At the national level, countries moved towards implementing their strategies, with Algeria setting up implementation committees, Ghana planning AI regulation and Egypt's data protection regulator confirming that AI systems fall under its data protection law. AI was a central theme at the UN General Assembly, where Kenya signed cooperation agreements with Anthropic and Stanford's Hoover Institution. The US International Development Finance Corporation approved up to $155 million in equity for WIOCC, its largest equity commitment to date, and this edition also looks at whether demand will keep pace with the continent's new data centres.
Policy news in Africa
This month's news from across the continent spans two main areas namely the development of regional AI frameworks and the implementation of national AI strategies.
At the continental level, the African Union held public consultations for its proposed AU Guidelines on AI Governance and Regulation and AU Model AI Law. Both instruments form part of the AU's implementation of the Continental AI Strategy adopted in 2024. Regional Economic Communities are also developing their own frameworks. IGAD held two-day consultations on its regional AI strategy, and COMESA held a validation meeting on a study that will inform its Regional AI Strategy, Model Policy Guidelines and Model Regulatory Framework. The IGAD process is funded by the Food Systems Resilience Programme for Eastern and Southern Africa, while the COMESA process is funded by the World Bank through the Inclusive Digitalization in Eastern and Southern Africa (IDEA) Programme. The East African Community's regional AI strategy was validated earlier this year.
The formulation of RECs specific frameworks resembles the way regional data protection instruments were developed. The impact of those instruments remains limited, particularly when measured against the maturity of national data protection frameworks, and experience suggests that regional frameworks do not necessarily shape the national legislation that follows.
At the national level, countries are intensifying their policy and regulatory efforts. While much of the activity to date has centred on adopting strategies, some countries are now establishing the institutional arrangements needed to implement them. Algeria's creation of five implementation committees, including one on large language models, is an example of this shift. Ghana announced plans to implement its national AI strategy through regulation that would promote responsible experimentation while focusing on risks and harms, with an emphasis on African-led regulatory solutions. In Egypt, the government signalled a new AI strategy, and commentators are unsure how it will relate to the existing one. Egypt's data protection regulator, the Personal Data Protection Centre (PDPC), also made clear that data used to train or run AI systems falls fully under the Personal Data Protection Law, and that privacy by design is required from the start. This places AI governance within an existing, enforceable legal regime ahead of any dedicated AI law. In Kenya, the Central Bank of Kenya released draft guidelines on the use of AI that would require banks to obtain written regulatory approval before deploying or updating AI systems.
Global Policy Watch
AI was a central theme at the UN General Assembly. Secretary-General António Guterres named AI as one of four "power tests" facing the world and called for responsibly paced AI development under a multilateral risk framework with independent oversight. President Trump rejected global control schemes for AI, while President Macron warned against leaving AI governance to two superpowers and a few technology companies. African leaders, including those of South Africa and Angola, called for Africa to help set the rules, and Kenya joined France, Spain and Australia in presenting a 2026–2027 roadmap for a coalition on children's rights in the age of AI, backed by 26 states and eight UN entities.
Prospects for wider coordination remain limited. In July, more than 1,100 AI company employees urged the US government to support international tools for managing the pace of AI development, but commentators note that China has dismissed the idea, and the Trump–Xi summit produced a bilateral channel for AI incidents and little else on AI. With the major powers divided, UN processes and issue-based coalitions such as the children's rights initiative remain among the main avenues through which African states can shape global AI rules.
In the United States, senators introduced a bill setting enforceable safety standards for frontier AI and proposed a federal commission to oversee frontier models and major platforms, while 26 state attorneys general urged Congress to adopt a comprehensive framework. As most frontier models used across Africa are built by US companies, rules adopted in Washington could shape the safeguards and disclosures that reach the continent.
In China, the Supreme People's Court issued 24 provisions in the “Opinions on Adjudicating Disputes Involving Artificial Intelligence in Accordance with the Law” on AI disputes covering face swapping, voice cloning, AI-generated content, training data and AI-generated evidence, with ordinary fault liability as the default for AI-related harms. Copyright questions over AI outputs and training data remain open. The guidance shows how courts can address AI disputes under existing law, an approach open to African judiciaries while dedicated AI legislation is still being developed.
In Europe, the Dutch data protection authority fined Uber more than €800 million over fully automated deactivations of drivers' accounts, and the European Commission designated ChatGPT a very large online search engine under the Digital Services Act, requiring OpenAI to assess and mitigate systemic risks, including fundamental rights. The Uber decision is directly relevant on the continent, where several data protection laws, including Kenya's and South Africa's, give people rights in relation to solely automated decisions and where platform work is growing.
State of technology: The demand problem beneath the compute rush
Across Africa, arguments about digital sovereignty increasingly come back to compute. The idea is that local data centres will allow African data to be stored, and African models to run, within African countries. September brought a wave of investment in response, from the DFC's investment in WIOCC to a series of new data centre announcements. Power supply through the grid remains a constraint, but there is another question that receives less attention, which is whether there will be enough customers for this capacity once it is built.
The figures suggest reason for caution. Africa holds roughly 0.6% of global data centre capacity, and outside South Africa only about a third of the capacity already built is live or fully fitted. African facilities typically take six to eight years to reach 85% occupancy, compared with two to three years in Europe, while 80 to 90% of African data is still processed abroad. As unused capacity is expensive to maintain, operators eventually pass the cost on to their customers, who are the same businesses and institutions that localisation is meant to serve.
Requiring data to stay in the country does little to make local hosting more competitive. Businesses place their workloads wherever price, reliability and performance make the most sense, and at present offshore providers often offer a better deal. The most effective policy for compute sovereignty may therefore look quite different from a localisation law. It would focus on the slower economic work of raising incomes, expanding digital adoption and lowering hosting costs until choosing an African provider makes commercial sense. If demand grows first, sovereignty will rest on a stable foundation. Without that demand, localisation mandates risk leaving data centres half empty, meeting the formal requirements of sovereignty while workloads continue to move offshore.
Explainer: What Is Carrier-Neutral, Open-Access Infrastructure?
A carrier-neutral, open-access network is digital infrastructure, such as fibre, subsea cables or a data centre, whose operator is independent of any single telecoms company and sells capacity to anyone who wants it on the same terms.
To see why this matters, compare it with a data centre owned by one telecoms company. There, customers are usually tied to that company's network, prices and service conditions. A neutral operator does not own, resell or favour any telecoms provider. This means that competing telecoms operators, cloud companies and internet firms can all connect in the same building on equal terms.
Having many providers under one roof increases competition, which tends to lower prices and improve reliability. It also makes it easier for customers to switch providers, because they only need to change their contract and do not have to move their servers to another building.
For these reasons, carrier-neutral infrastructure features prominently in debates on digital sovereignty. This month, WIOCC's carrier-neutral platform drew foreign investment. Neutral facilities are open to all providers on equal terms, which is intended to limit any single company's control over a country's digital backbone. Ownership and financing are separate questions, and they increasingly feature in African discussions of compute sovereignty because they affect who controls the infrastructure and who benefits from it.
Research & community voices
In this edition we look at two recent papers that look at how African countries can benefit from AI without competing at the frontier.
A GIGA Focus paper notes that the global AI infrastructure boom has largely bypassed the continent. The largest US technology firms plan almost USD 800 billion in capital spending in 2026, while Africa hosts few AI-focused data centers, and concerns over the scale of power required to support large-scale AI infrastructure. AI use among Africa's working-age population stands at 8.6 per cent, compared with 17.8 per cent globally, and African languages make up less than 1 per cent of the internet content on which frontier models are trained. The paper argues that Africa's best route is therefore to adapt open-weight models to local needs, invest in power, connectivity and distributed computing, and cooperate regionally, while managing dependence on foreign providers by diversifying partners and insisting on open standards.
A paper presented at the AfriLang AI conference in Kampala shows what this adaptation looks like in practice. Researchers from Dedan Kimathi University of Technology and Maseno University took Meta's openly available Llama 3.2 model and trained it further to translate English into Kikuyu, using 30,000 pairs of English and Kikuyu sentences. Over six rounds of training, the model's translation score rose from 2.41 to 25.21 out of 100. The authors describe this as partial learning and note that some of the human translations in their dataset contained errors in sentences with complex grammar and idioms, particularly in health texts. Tools at this stage of development would need further testing, including evaluation by Kikuyu speakers, before being used in sensitive areas such as health or public services.
Together, the two papers suggest that African countries can adapt existing AI models to local languages with modest resources. Doing this well, however, requires sustained investment in local language data and in the technical skills needed to use it.
Partnerships & collaborations
September's partnerships focused largely on the financing of digital infrastructure and on governance capacity. On 16 September, the US International Development Finance Corporation (DFC) approved up to $155 million in equity for WIOCC, its largest equity commitment to date. WIOCC, based in Johannesburg, operates subsea cables, fibre networks and more than 40 data centres in 30 countries, and runs an open-access, carrier-neutral platform. The deal is one of several recent foreign investments in Africa's digital infrastructure, which has drawn attention to who finances and owns the continent's networks and data centres.
In the same breadth, on 21 September, during the UN General Assembly, Kenya signed a five-year partnership with Stanford University's Hoover Institution covering AI, data governance and institutional capacity. The following day, Kenya signed a Joint Declaration with Anthropic to deepen cooperation on AI skills development, research, safety and public-sector applications, with a framework for AI applications in priority sectors including education and health.
AI in the public sector
South Africa's Communications Minister Solly Malatsi said at GovTech 2026 in Durban that a new draft national AI policy should reach Cabinet by March 2027. The previous draft was withdrawn in April after it was found to contain fictitious, AI-generated citations.
On 25 September, the African Development Bank Group, the UNDP AI Hub for Sustainable Development, Google and Apolitical launched a free training programme to help public officials across Africa understand and use AI in their work. Courses cover AI fundamentals for public servants, AI leadership in government, and smarter procurement with AI.
