Demand for AI is the bottleneck, so we’ll tackle it by starting with Africa

The AI 10 Billion Initiative is based on a clear insight: the main constraint on the adoption of artificial intelligence is not the volume of investment. It is sequencing – that is, ensuring that the right tool meets the right set of enabling conditions at the right time.

The urgency is demographic in nature. By 2040, Africa’s working-age population will exceed one billion people; by 2050, one in four people in the world will be African. Meanwhile, cutting-edge AI models are already replacing many entry-level cognitive jobs in advanced economies: the very jobs that have historically served as the main gateway for young people from emerging economies into the global labour market. If the adoption of AI in Africa were to be delayed by the lack of the necessary enabling conditions, whilst its spread in advanced economies continued to accelerate, the result would not be a mere delay that could be made up over time. Instead, it would lead to a structural divergence, in which the continent with the youngest population on the planet would also be the one with the least capacity to participate in the AI economy. The issue of sequencing, therefore, is not merely a matter of efficiency. It is, above all, a matter of inequality, and the time factor is crucial.

The Initiative organises its work around five strategic areas, each of which addresses a specific critical issue relating to the conditions necessary for a genuine demand for AI to develop.

Infrastructure: Data centres, GPU-based computing power, fibre-optic networks and energy: the physical foundation without which no other element can function. Private capital and public-private partnerships.

1-Data & Models: Centralised, governed repositories and language models for African languages: the layer of resources that makes artificial intelligence relevant to local contexts. This area requires public-sector-led investment, as market forces alone do not incentivise the development of local language models for markets characterised by low profit margins.
2-Economy: AI applications in agriculture, healthcare and finance: the level at which demand is transformed into productivity and economic value. Private-sector leadership, but only once the infrastructure and data have created the necessary conditions.
3-Human capital: Training of developers and dissemination of AI skills within businesses: the level of skills that determines whether the technology can effectively be transformed into institutional capacity. Funded through public investment.
4-Policies & Governance: Regulatory frameworks for data protection, rules for public procurement, regulatory sandboxes: the structural conditions that enable all other interventions to reinforce one another rather than proceeding in a fragmented manner. In the absence of this element, all other areas of intervention end up functioning as temporary and isolated solutions.

These five areas do not compete with one another for the same capital and cannot function independently of one another. Capital follows its own logic: it naturally tends to flow towards opportunities where risk is mitigated and the expected return is high. The ‘Economy’ pillar can attract private investment only once ‘Infrastructure’ and ‘Data’ have already reduced the level of risk. Conversely, ‘Human Capital’ and ‘Governance’ require public or sovereign funding and non-repayable grants, as they do not generate a direct economic return. Yet it is precisely these conditions that are indispensable; without them, private capital alone is unable to deliver results on a large scale. Development finance that builds data centres without simultaneously developing the necessary ecosystem of policies, skills and data around them is akin to installing an electrical socket without building the network that powers it.

The ‘100 Diffusion Pathways’ framework introduces a distinction that most analyses overlook: the diffusion of artificial intelligence is not a measure of the adoption of generative AI. It is a measure of a system’s capacity – that is, the extent to which a society has created the necessary conditions for artificial intelligence, in whatever form, to be adopted safely and productively in economic and civic life.

This distinction is crucial because the most widely used method of measurement — counting users of generative AI as a proportion of the working-age population — systematically underestimates what is actually happening, particularly in contexts where innovation has moved beyond the dominant paradigm of AI products. If adoption is measured solely by product usage, we end up designing programmes aimed simply at expanding access to those products.

If, on the other hand, uptake is measured in terms of capacity, then initiatives focus on investment in infrastructure, governance and skills that enable communities to shape and manage their own future in artificial intelligence independently. In other words, the measurement tool determines the type of initiative.

The case of Luweero is not an isolated incident. The most significant signs that demand for AI is beginning to grow come from innovators working in contexts characterised by limited resources, languages under-represented in AI models, environments where connectivity is poor, and farming communities that the mainstream AI market has not yet prioritised.

Voice-based artificial intelligence in East Africa provides the most obvious parallel. Voice assistants built on models that have been streamlined and optimised for local contexts are now being used to enrol farmers in support programmes, to provide agricultural advice and to collect data in rural areas, where low literacy rates and poor connectivity make text-based AI solutions inaccessible. A single agricultural extension worker tasked with supporting 5,000 farmers is a problem that is difficult to solve using traditional tools. A voice assistant capable of providing advice round the clock in the farmer’s own language and of operating directly on the device even without an internet connection, on the other hand, represents a completely different kind of solution.

However, smartphones account for only around half of all mobile connections in Africa. For this reason, the demand architecture must integrate with existing telecommunications infrastructure: USSD menus, SMS advisory services, IVR (Interactive Voice Response) systems and WhatsApp. African developers are distributing AI-generated agricultural advice via Africa’s Talking’s SMS infrastructure and Hello Tractor’s USSD-based network of farmers: channels that require neither an internet connection, a smartphone nor digital skills. This is not a stopgap solution, but a deliberate design choice that meets demand where it actually exists.

What makes all this possible is a structural innovation in the way artificial intelligence infrastructure is organised. Innovators are experimenting with trilateral partnership models: coalitions between national governments, regional development institutions and local AI laboratories that pool computing capacity, share governance frameworks and coordinate implementation in ways that no single actor would be able to sustain on its own. These represent prototypes of a different architecture for the roll-out of AI, in which the enabling conditions are built by the very communities that need them.

The Italy–India–Kenya AI Voice Pathway, led by the UNDP AI Hub for Sustainable Development and mentioned in the Modi–Meloni Joint Declaration of April 2026, is the first operational prototype of this trilateral architecture at national level. The Italian supercomputer Leonardo, managed by CINECA, provided the infrastructure for training the models. India’s EkStep Foundation contributed the VoicERA orchestration platform. African developers — Crane AI Labs in Uganda and MsingiAI Kenya, whose speech recognition (ASR) and text-to-speech (TTS) models are considered the best published models worldwide for the Swahili language — provided the language models, cultural context expertise and on-the-ground implementation. On 20 May 2026, the programme held its first live demonstration, attended by ambassadors and supporters, at the Italia Pavilion during GITEX Kenya. This is the kind of approach that the AI 10 Billion Initiative was designed to replicate and scale up.

Three elements must unfold in a specific sequence, not simply in parallel.

Infrastructure integrated with energy. It is not possible to create demand for artificial intelligence in contexts where there is a structural shortage of electricity. The debate on data centres and that on energy infrastructure must proceed hand in hand: imposing data localisation requirements without investing in energy production and distribution means adopting policies that lack the necessary conditions for implementation. For this reason, the Infrastructure pillar of the AI 10 Billion Initiative considers this integration to be an essential requirement.

A measurement system that describes what is actually happening. The Diffusion Pathways coalition is developing an alternative assessment framework that measures the capacity to adopt and govern artificial intelligence, rather than simply the use of AI applications. This is the tool needed to demonstrate to governments, development finance institutions and private investors that the indicators currently in use underestimate actual progress and incorrectly identify the areas where action is needed.

Legislation as a key enabler. All initiatives aimed at stimulating demand for AI are, ultimately, temporary solutions until a solid regulatory framework is in place. Regulations on data localisation, requirements for investment in computing capacity linked to market access, and procurement procedures for artificial intelligence for public institutions are the conditions that enable infrastructure investments to produce lasting and cumulative effects. The lesson from M-Pesa is illuminating: its success did not require campaigns to promote its uptake. Above all, it required regulatory clarity. Once the appropriate conditions were in place, adoption followed almost inevitably.

Crane AI Labs — an artificial intelligence laboratory with operational headquarters in Kampala and its head office in Abu Dhabi, whose infrastructure choices and commitment to digital sovereignty were referred to on several occasions during the working sessions of the Diffusion Pathways programme — was highlighted on stage at Google I/O 2026 by Google DeepMind’s Gemma product team. It was mentioned, alongside MedGemma, Cell2Sentence and the Ukrainian E-Permit system, as a best practice example for developers who are customising Gemma models to make them accessible to traditionally underserved language communities. Crane AI Labs is also featured in the Gemmaverse, Google’s public showcase dedicated to the most significant implementations of Gemma models worldwide.

The financial aspect is just as important as the recognition received. In markets where the average revenue per user may never exceed a few dollars a year, the dominant cloud-based AI distribution model — which involves a charge for every API call — is structurally unsustainable. Running models directly on the device completely eliminates this variable cost. The cost of each inference approaches zero, and, for the first time, the economic structure of a consumer-facing AI product becomes sustainable. This is not merely a matter of improved efficiency. It is the difference between a business model that can work in Africa and one that cannot.

"Stop waiting for the internet infrastructure to become perfect. The efficiency of models such as Gemma means that it is already possible today to build world-class artificial intelligence, designed to work offline, serving your own community." — Bakunga Bronson, co-founder of Crane AI Labs

Crane AI Labs’ institutional impact demonstrates what can be achieved by an approach that first establishes the enabling conditions and only subsequently rolls out the technology: the provision of agricultural advisory services through Hello Tractor’s network of farmers; the installation of East Africa’s first NVIDIA DGX Spark system; one of the first versions in African languages of the NVIDIA Nemotron family of open-source models. In April 2026, during the World Bank’s Spring Meetings in Washington, Crane AI Labs’ work on offline artificial intelligence was presented as part of the session dedicated to Small AI, in the presence of finance ministers, central bank governors and executives from leading international development finance institutions.

A presence in the Gemmaverse highlights a key point: when innovators manage to solve the most complex problems on the demand side — language barriers, connectivity limitations and the economic viability of AI running directly on devices — the global supply chain begins to follow suit. Google’s strategy of making model weights available has created the conditions that have made Crane AI Labs’ innovation possible. In turn, Crane’s implementation work is demonstrating both the existence of a market and the social value of this approach. It is precisely this sequence that the AI 10 Billion Initiative aims to replicate: not a series of isolated cases, but an entire architecture for the development of artificial intelligence based on building demand and intended to be extended across the entire African continent.

African innovators are not waiting for infrastructure to be perfect, for major global cloud providers to extend their services to less profitable market segments, or for all regulatory frameworks to be finalised before they start innovating. They are simultaneously building both the necessary conditions and the technological tools. The midwife in Luweero, the voice assistant advising 5,000 farmers, and the trilateral partnership between Italia, India and Kenya are not exceptional cases. They are concrete proof that, when the enabling conditions are developed in the right sequence, demand does not need to be created artificially: it has always existed.

* Keyzom Ngodup Massally, Director of the AI Hub for Sustainable Development and Head of the Digital & AI Programme, United Nations Development Programme (UNDP)

* Ugyo Zine Edosio, Head of Innovation and Digital, African Development Bank