I recently read Open Weights and American AI Leadership, an open letter from a coalition of American technology companies arguing that open weight AI models are central to the country’s future prosperity and technological leadership. What struck me was not the case it made for the United States, but the mechanism it proposed for achieving it.

Microsoft signed the letter, and Satya Nadella endorsed it publicly on the day it was published, describing open weight models as essential to a healthy AI ecosystem and framing them as a route to strengthening American competitiveness while protecting national security. The proposition is not cheaper subscriptions or wider API access. It is that AI leadership will be determined by whether the technology diffuses into every sector rather than by which company holds the single most capable frontier model, and that downloadable models are what makes that diffusion possible.

The letter’s final paragraph reads:

“The age of AI can be one of prosperity. With the right choices, open weight AI can expand opportunity, strengthen competition, extend American technological leadership, mitigate risk, and ensure that the benefits of this extraordinary technology are shared broadly across our economy. That future is worth building, and the United States should lead in building it.”

The paragraph was written as the conclusion to an American policy argument. It is the starting point for this one. What follows examines why the mechanism it describes may have implications well beyond the audience it was written for, and what that could mean for Africa.

What an open weight model actually is

Most people use AI through a website or an app. You type into a chat interface, the answer comes back, and the model itself stays on the provider’s servers. You are renting access. An open weight model works differently. The trained model is published as a file that you can download, run on your own hardware and adapt for your own use, without needing the provider’s permission.

The distinction is roughly that between a restaurant and a recipe. In the first case, the cooking happens elsewhere and you pay each time you eat. In the second, the recipe is handed over and the cooking happens in your own kitchen, for whoever you choose, as often as you like.

For readers familiar with the category, the letter’s own definition is worth taking literally, because each verb is doing deliberate work. It describes open weight models as AI models that anyone can “download, inspect, modify, and run on their own infrastructure.”

Inspect and modify underpin the transparency and safety arguments that dominate the policy discussion. Run on their own infrastructure is the phrase that matters most for institutions outside the United States.

Consider one institution - a Nigerian bank has until 1 January 2027 to ensure that payment transaction data generated in the country is stored and managed in Nigeria, under a circular the Central Bank of Nigeria issued on 15 June 2026. That bank will still want to apply machine learning to fraud detection, credit decisioning and reconciliation. If the only models available to it are reachable through a foreign provider’s interface, its compliance position depends on whether that provider chooses to build a supervised local deployment, and on a timeline the bank does not set. If it can download a model and run it on hardware it already owns, the question largely resolves itself. That distinction, between renting capability and holding it, is what the letter is arguing about.

Four ways this could matter for Africa

If the coalition is right, then open weights matter not because they make AI free, but because they change where advanced AI can be built, adapted and deployed. The opportunities are broad, but four stand out because they strengthen capabilities that will shape Africa’s long-term AI ecosystem.

Cost is the most immediate consequence. The letter’s economic case is that organisations can build on advanced models without training one from scratch or paying frontier-model prices for every task, reserving the most expensive capability for genuinely difficult problems and running smaller specialised models elsewhere. That argument reads differently in economies where AI services are priced in dollars but revenues are earned in naira, shillings or rand. In that context, per-token pricing is not simply a cost to be optimised. It places a limit on how widely AI can be deployed, since the obligation grows with usage in a currency the institution does not earn. Open weight models change that equation. Where the upfront investment in compute can be made, they replace an open-ended foreign-currency operating expense with infrastructure that institutions can own, budget for and build upon.

Language may be the more important long term implication. Commercial frontier models are trained overwhelmingly on English and a small set of European and Asian languages, and waiting for a foreign laboratory to conclude that Hausa is commercially interesting is not a strategy. Lelapa AI, a South African research lab, built InkubaLM, a 400 million parameter model trained from scratch on 1.9 billion tokens across five African languages plus English and French, covering isiZulu, Yoruba, Swahili, isiXhosa and Hausa, languages with roughly 360 million speakers. The weights are published on Hugging Face. Work of that kind is not possible where the only available relationship with a model is rental.

Local control is where the open weight argument becomes most practical. With a closed model, an institution’s data must be sent to a provider’s environment for processing. With an open weight model, the model can instead be deployed within the institution’s own infrastructure, allowing data to remain where it is governed. This distinction matters most in regulated sectors, where financial services and government organisations have historically moved more cautiously to cloud environments. Financial services showed the lowest cloud adoption of any sector in a 2024 survey of major African businesses, with regulatory restriction named among the reasons. For these institutions, adopting a closed model is not only a technical decision. It may require placing supervised data within a foreign provider’s environment and satisfying regulators that the arrangement meets local obligations. The Central Bank of Nigeria’s payment data localisation directive makes this point concrete. From 1 January 2027, payment transaction data generated in Nigeria must be stored and managed within the country, turning local hosting from a procurement preference into a regulatory requirement for affected institutions. In that context, the ability to deploy models within controlled environments becomes more than a workaround. It allows institutions to build AI capability while maintaining governance over their data and reducing dependence on a small number of foreign cloud providers.

Skilling is the point most often left out of this argument. An industry cannot be built on a product that can only be consumed. Open weights create something that engineers, researchers and students can study, adapt and improve, turning AI from a service to be used into a capability to be developed. Lelapa’s Buzuzu-Mavi Challenge, run with Zindi, invited participants to compress InkubaLM and drew 490 participants from 61 countries, with all three top places taken by African entrants. Using techniques such as distillation, where a smaller model is trained to reproduce the behaviour of a larger one, they reduced the model to 40 million parameters while holding translation quality for Swahili and Hausa. That is engineering capability being built in public, on a model that was open enough to be taken apart.

Democratisation, and what it does not cover

The word carrying the most weight in the letter is not access but diffusion. The argument is that AI creates the most value when it spreads into ordinary workflows rather than remaining concentrated behind a small number of expensive endpoints.

Open weights remove the permission barrier, but do not remove the harder constraints around building AI capability: GPUs remain expensive and difficult to import; power infrastructure is uneven across much of the continent; and specialists are scarce and mobile. For many African languages, there is still not enough digitised text to train on, which is a data problem that no licence can solve. Where internet access remains limited and around 70 per cent of smartphone users are on entry-level devices, the constraint is seldom the terms of use alone.

There are also limits to what “open” means in practice. InkubaLM is released under a CC BY-NC 4.0 licence, which is non-commercial, and Lelapa has described it as a base model, a building block rather than a finished production system. Lelapa has also credited Microsoft’s AI for Good Lab with the compute used to train it. Across the industry, open weight releases often provide access to model parameters without releasing training data or training code, and future access remains dependent on the choices of the organisations that publish them.

Democratisation, therefore, is a starting point rather than an outcome. Open weights make capability possible, but turning that possibility into lasting advantage will require investment in compute, data and institutions as much as openness itself.

The precedent Africa already has

One of the clearest examples on this continent of open infrastructure creating long-term institutional capability is not an AI system. DHIS2, developed by the Health Information Systems Programme at the University of Oslo with roots in South Africa, is now used by ministries of health in more than 75 low and middle income countries as the government-wide system of record for health data, covering a population of approximately 3.2 billion people, and in more than 130 countries once deployments beyond health are counted. The web-based platform was released in 2006, building on a programme that began in 1994. No ministry negotiated a proprietary licence, no vendor retained the ability to withdraw the system, the data stayed under national control, and a regional network of implementers grew up around it. The instructive detail is that openness alone did not produce that outcome. Three decades of sustained Norwegian public funding across HISP and DHIS2, alongside a deliberately constructed implementation network, converted an open licence into working national systems. That is the part of the model that still has no equivalent in the current AI conversation.

The letter argues that the age of AI can be one of prosperity and that the United States should lead in building it. What it defends is downloadable by anyone with connectivity and storage, and the permission it grants does not stop at a border.

Open weights solve the problem of permission. They do not solve the problem of capacity. Prosperity will depend less on whether models can cross borders than on whether investment, compute and implementation ecosystems can be built around them.

That is why this debate matters here. For perhaps the first time in the foundation model era, the technological artefact is available before the industrial ecosystem around it has settled. Countries that invest early in compute, local datasets, engineering capability and public institutions may find themselves participating in the AI economy as builders rather than customers. Open weights do not guarantee that outcome, but they make it possible to pursue.

Sources

• Open Weights and American AI Leadership, 24 July 2026: microsoft.com/en-us/corporate-responsibility/topics/open-weight

• Satya Nadella on open weight models, X, 24 July 2026: x.com/satyanadella

• Central Bank of Nigeria payment data localisation directive, Circular PSS/DIR/PUB/CIR/001/004, 15 June 2026: punchng.com and privalexadvisory.com

• Africa’s leap ahead into cloud: opportunities and barriers, McKinsey, 2024: mckinsey.com

• InkubaLM, Lelapa AI: lelapa.ai and huggingface.co/lelapa

• InkubaLM language coverage and compute acknowledgement: itweb.co.za

• Buzuzu-Mavi Challenge results, Lelapa AI and Zindi: lelapa.ai

• Model compression techniques and results: exchange.design

• DHIS2 key facts and deployment figures: dhis2.org/press

• HISP history and origins: dhis2.org/hisp-history

• DHIS2 platform release history: openhealthnews.com

The Central Bank of Nigeria has issued a payments circular that addresses market concentration, data sovereignty, ownership transparency, and systemic oversight in a single regulatory instrument. The strategic direction is sound and consistent with a broader global pattern, but the sequencing raises questions about implementation. This piece, the first of two, examines what the circular says and where its strategy holds. Part II turns to where the implementation does not, and what a sequenced rollout might have looked like.

A brief note before we begin: I write this from an implementation perspective. My work sits closer to data, AI, and digital transformation than to financial policy, so this analysis focuses on what the circular looks like when an institution actually has to absorb it.

Contents

A note on structure: Sections I and II set out what the circular contains, for readers who want the background. Readers already familiar with the circular can skip to Section III, where the analysis begins.

I. Understanding the Circular: what the CBN issued, and why

II. The Four Pillars: the provisions, in CBN’s own words

III. Strategic Assessment: where the strategy holds

IV. Implementation Risks: where the implementation does not

(Part II continues with: V. The Opportunity, VI. Sequencing and Implementation, VII. Conclusion)

I. Understanding the Circular

On 15 June 2026, the Central Bank of Nigeria issued circular PSS/DIR/PUB/CIR/001/004, titled Introduction of Market Structure Requirements, Data Localisation, Ultimate Beneficial Ownership Disclosure, and Systemic Oversight Measures in the Nigeria Payments System. The circular was signed by the Director of the Payments System Supervision Department, Dr. Rakiya Yusuf, and addressed to Deposit Money Banks, Microfinance Banks, Mobile Money Operators, Switching Companies, Payment Solution Service Providers, Payment Terminal Service Providers, Super Agents, and other licensed payment operators.

The title is doing important work, because it names four distinct interventions packaged into a single regulatory action. Most of the public coverage has framed the document as a “data localisation circular,” which understates what is actually being introduced. The circular addresses market structure, data sovereignty, ownership transparency, and systemic oversight as parts of one coherent intervention rather than as separate measures.

The CBN’s own rationale, taken directly from the document, is worth quoting in full because it establishes the framing the rest of the circular operates within:

“The Nigerian payments ecosystem has witnessed significant structural developments characterised by rapid growth in electronic payments, increasing adoption of digital financial services and the emergence of operators with substantial market presence across key payment activities. While these developments have supported innovation, efficiency, and financial inclusion, they have also raised concerns relating to market concentration, operational dependence, systemic importance, transparency of ownership structures, and the localisation of critical payment data.”

The regulator is observing that the Nigerian payments market has grown rapidly, that a small number of players have become very large within it, and that it considers this the moment to draw structural lines before those concerns develop into structural problems. The circular addresses this through four pillars, which are best understood individually before being assessed together.

II. The Four Pillars

Pillar 1: The Market Structure Rule

This provision has received less public attention than the data localisation requirement, but its structural significance is at least as large. The rule states:

“Any licensed financial institution engaged in card-issuing activities, whether individually or as part of a group of related entities, that holds more than 25 per cent market share in card issuing within any rolling 12-month period shall not hold more than 15 per cent market share in merchant-acquiring activities during the same period.”

The same cap applies in reverse to merchant-acquiring dominance.

Payments has two sides. Issuing refers to the institution that provides the consumer with the card or wallet. Acquiring refers to the institution that processes the merchant’s side of the transaction. A number of Nigerian institutions have built integrated operations spanning both, capturing margin at every stage. The CBN is establishing that an institution can be dominant in one side of the market, but not in both at once.

The critical phrase is “whether individually or as part of a group of related entities.” It closes the workaround of splitting issuing and acquiring activities into separate subsidiaries under a shared parent. By explicitly looking through corporate structure, the CBN is signalling that the rule is enforceable rather than cosmetic.

Pillar 2: Data Localisation

The circular provides:

“All financial institutions and participants facilitating payments within Nigeria shall ensure that payments transaction data generated within Nigeria are stored and managed in Nigeria in accordance with data protection laws and regulations applicable in Nigeria.”

Full compliance is mandated by 1 January 2027. The practical effect is that every payment transaction generated within Nigeria must reside on infrastructure located inside the country. This is also where the central technical ambiguity in the directive sits, which we return to in the implementation risks below.

Pillar 3: Ultimate Beneficial Ownership Disclosure

The provision states:

“All Deposit Money Banks, Payment Service Providers, and Other Financial Institutions with digital payment footprints shall disclose the Ultimate Beneficial Ownership (UBO) of significant shareholders in accordance with applicable extant laws and regulations.”

Every regulated institution must disclose not just the shareholders on its corporate register, but the natural persons who ultimately control those shareholders. Shell companies, trust structures, and layered holdings must resolve to identifiable human owners, with records kept current and made available on request.

Although framed as an anti-money laundering measure, the provision’s more important function is enabling enforcement of the market structure rule in Pillar 1. The “group of related entities” clause is unenforceable without visibility into who ultimately controls each entity, and UBO disclosure provides that visibility. The two pillars work together.

Pillar 4: Monthly Market-Share Returns

The CBN requires all regulated entities to submit monthly market-share returns using prescribed templates and reporting timelines.

This creates a continuous, near-real-time view of market share in reporting formats the CBN itself defines, which prevents institutions from shaping the data through their own internal reporting. Without it, the 25%/15% cap cannot be verified, and the regulator would be intervening only after concentration problems became visible through other means.

III. Strategic Assessment

Read together, the four pillars constitute a coherent intervention. They address a structural problem in the Nigerian payments market that is genuinely present. The market is concentrated, several institutions operate on both sides of the value chain, a substantial volume of Nigerian payment data is held on infrastructure outside the country, and the ownership of some of the largest operators traces through corporate structures that are not transparent to the regulator. The strategic direction is appropriate to the problem it identifies.

The intervention is also part of a clear global pattern. The Reserve Bank of India issued a directly comparable payment data localisation directive in April 2018, through circular RBI/2017-18/153, requiring all payment system data to be stored within India. Indonesia introduced similar provisions through its OJK financial regulator. Saudi Arabia did the same through SAMA. China embedded payment data localisation in its Personal Information Protection Law and Data Security Law. Russia adopted Federal Law 242-FZ in 2014. Vietnam introduced Decree 53. Brazil adopted comparable measures through its broader data protection framework. According to the Center for Global Development, more than 100 countries now have some form of data localisation rule in place, with financial data localisation representing the most globally consistent sector pattern.

What is genuinely distinctive about the directive is not the data localisation provision itself, which is broadly conventional, but the way the CBN has packaged data localisation with anti-monopoly market structure rules, beneficial ownership transparency, and continuous reporting, all enforced through a single coordinated instrument. The Indian 2018 directive addressed data alone. The Saudi and Indonesian frameworks treat data and market structure as separate regulatory questions. By contrast, the CBN’s directive treats the payments system as an architecture with multiple layers that need to be regulated together rather than in isolation. That integrated approach is uncommon among the comparable regulations issued by other jurisdictions over the past decade.

In broader infrastructure terms, the CBN is asserting sovereignty across three layers of the payments system simultaneously: the data layer, the governance layer, and the systemic oversight layer. The instinct is the right one, and it is consistent with what digital sovereignty looks like when implemented coherently rather than piecemeal.

IV. Implementation Risks

The strategic direction of the circular is defensible. The implementation is where the analysis becomes more difficult, and four concerns deserve examination.

The first concern is whether Nigeria’s data centre infrastructure can scale rapidly enough to accommodate the localisation requirement by January 2027. Nigeria currently hosts approximately 15 per cent of Africa’s data centre capacity across 26 operational and planned facilities, with a total installed IT load of roughly 75 megawatts. Industry forecasts suggest that capacity could reach 150 megawatts by 2027 if all planned facilities are delivered on schedule, which would represent a doubling in approximately eighteen months. A large proportion of Nigerian financial institutions currently host significant workloads in South African cloud regions, primarily AWS Cape Town and Azure Johannesburg. The repatriation of those workloads to infrastructure inside Nigeria, on the timeline the CBN has set, would require the local data centre industry to absorb a step-change in demand it was not previously sized to handle. Whether the capacity build can keep pace with the regulatory deadline is, at present, an open question.

The second concern is that compliance costs are likely to reach consumers, and the evidence from comparable markets suggests the effect will be material. Localisation requires institutions to either contract premium local data centre capacity, build their own infrastructure, or migrate to international cloud providers’ Nigerian zones if and when those become available. Each option is more expensive than the existing cross-border arrangements that currently underpin much of the industry. Historically, in payments markets, cost differentials of this kind travel through to consumers via higher transaction fees or service charges.

The distributional implications of this passthrough deserve particular attention. Analysis from the Center for Global Development examining data localisation regimes across more than one hundred countries concludes that, while proponents argue these policies advance national interests, in practice they often fail to deliver on their stated goals while imposing what the analysis terms a “localization tax” on low-income consumers. By increasing the operating costs of providers serving lower-income segments and restricting access to globally competitive cloud services, the policies tend to fall hardest on those least able to absorb the additional cost. The CBN circular does not address whether institutions will be permitted to pass these costs through to end users, nor whether consumer pricing for affected services will be regulated alongside the new requirements. This is a meaningful gap, and one that sits in tension with the CBN’s own stated priorities, given that financial inclusion remains a central policy objective for the regulator and that a substantial portion of the Nigerian population is either unbanked or marginally banked. A regulation that strengthens data sovereignty at the cost of widening the affordability gap for the financially excluded is one that may achieve one policy goal by undermining another.

The third concern is structural rather than transitional. Nigeria’s national grid produces approximately 6 gigawatts of generation capacity for a population of 230 million people, whereas South Africa’s grid produces approximately 48 gigawatts for a population of 63 million. Nigerian data centres operate with a baseline assumption of diesel backup generation, which adds roughly 40 per cent to operating costs, with a typical 1.5-megawatt facility storing approximately 90,000 litres of diesel on-site. This is not a problem that resolves on an eighteen-month timeline, because it reflects a structural constraint on how much compliance-grade data centre infrastructure can exist in Nigeria at any given price point.

The Kenyan experience over the past year is instructive on precisely this point. In May 2024, Kenya announced a $1 billion hyperscale data centre partnership with Microsoft and G42 at the Olkaria geothermal site, with a 100-megawatt first phase and a long-term ambition of 1 gigawatt. In May 2026, the project was suspended after the partners and the Kenyan government were unable to agree on power and capacity guarantees. President William Ruto explained the underlying problem directly: “To switch on that one data centre, we would need to shut off power for half the country. That’s when I knew there was a problem.” Kenya, with roughly 3.3 gigawatts of installed capacity and nearly 90 per cent of its generation from renewable sources, with geothermal alone accounting for almost half, could not commit to powering a single hyperscale facility without disrupting service to the rest of the economy. The country is now pursuing expansion to 10 gigawatts of installed capacity by 2030, seeking an estimated $38 billion in funding. The lesson is consequential: even a country with stronger renewable credentials than Nigeria was unable to absorb a single large facility without a multi-year grid expansion programme funded in parallel. A localisation policy that requires an entire payments sector to repatriate workloads to domestic infrastructure, without an equivalent grid investment programme running alongside, leaves a substantial gap between the regulatory ambition and the physical capacity to support it.

The fourth concern is the cloud provider ambiguity, which is the most significant unresolved question in the circular. The document does not specify whether international cloud providers operating local availability zones inside Nigeria would count as compliant infrastructure under the localisation rule. If AWS, Microsoft Azure, or Google Cloud were to build Nigerian data centres, the question of whether institutions could host on those zones and remain compliant is not addressed. The two readings produce materially different outcomes for the industry.

Under a liberal reading of the rule, foreign hyperscalers would build local availability zones to comply, institutions would migrate workloads to those zones, and the data localisation rule would solve data residency without addressing the underlying dependency on foreign cloud providers. This would constitute sovereignty in geographic terms only. Under a strict reading, only Nigerian-owned or Nigerian-operated data centres would qualify, and the market would shift decisively toward domestic operators including MainOne, Rack Centre, Africa Data Centres, and Nxtra by Airtel, with significant implications for pricing across the industry. This is not a marginal question. It changes the cost basis for the entire payments sector and determines whether the rule achieves genuine sovereignty or merely relocates the geographic address of dependency. The CBN will need to clarify its position before institutions can plan compliance with any precision, and the fact that this clarification has not been issued alongside a circular setting an eighteen-month deadline is itself indicative of the gap between regulatory ambition and operational readiness.

Part II picks up here: the commercial and architectural opportunity the circular creates, why sequencing matters as much as strategy in determining whether a regulation like this succeeds, and what a phased version of the CBN circular might have looked like.

Sources and further reading:

• Central Bank of Nigeria circular PSS/DIR/PUB/CIR/001/004 (15 June 2026)

• Premium Times: CBN orders banks, fintechs to disclose owners, store payment data in Nigeria

• Reserve Bank of India: Storage of Payment System Data (Circular RBI/2017-18/153, April 2018)

• TechCabal: Can Nigeria’s data centres power Africa’s AI future?

• Introl: Africa’s AI Data Center Boom, Nigeria, Kenya, South Africa

• African Business: Inside the race to fire up Africa’s power-hungry data centres

• Center for Global Development: Data Localization, a Hidden Tax on the Poor

• Data Center Dynamics: Microsoft and G42 data center in Kenya stalled due to lack of power capacity

• International Energy Agency: Kenya 2024 Energy Policy Review

This is the second piece in a two-part series on the Central Bank of Nigeria’s June 2026 payments circular. Part I set out the circular’s four pillars (market structure, data localisation, ultimate beneficial ownership disclosure, and systemic oversight) and argued that the strategic logic behind them is sound, even as it flagged four unresolved implementation risks: data centre capacity, consumer cost passthrough, grid constraints, and cloud provider ambiguity.

This piece picks up from there. It looks at the commercial and architectural opportunity the circular creates, then makes the case, using India’s 2018 experience and Rwanda’s phased approach as contrasting precedents, that a sequenced rollout would have served the CBN’s own goals better than the deadline it chose.

Contents

(Continued from Part I: I. Understanding the Circular, II. The Four Pillars, III. Strategic Assessment, IV. Implementation Risks)

V. The Opportunity: what the circular makes possible

VI. Sequencing and Implementation: the case for a phased approach

VII. Conclusion

V. The Opportunity

A balanced assessment of the circular must acknowledge that the same provisions which create implementation pressure also create commercial and architectural opportunity. Viewed from an implementation perspective, three opportunities stand out, each tied directly to a specific pillar of the circular and each creating demand for a category of work that Nigerian institutions had collectively been deferring.

The first opportunity flows from the data localisation provision in Pillar 2, which forces an architectural decision that institutions deploying AI had previously been able to defer. The requirement that payment data be stored and managed in Nigeria by January 2027 means that AI workloads in fraud detection, credit scoring, risk modelling, and customer intelligence can no longer be designed against cross-border cloud infrastructure as a matter of convenience. They must be designed against domestic infrastructure as a matter of regulation. For institutions that treat this as a compliance task, the outcome is AI infrastructure that is geographically compliant but operationally constrained. For institutions that treat it as an architectural opportunity, the outcome is AI-ready data foundations that compound across every workload deployed afterwards. The most visible supply-side response to this specific question is MTN’s $235 million Sifiso Dabengwa Data Centre, launched in July 2025, which paired a 4.5-megawatt Tier III facility with a $20 million domestic cloud service explicitly positioned for compute-intensive workloads. MTN’s chief executive described the second phase as readiness for the processing power that artificial intelligence requires. Pricing in naira rather than US dollars removes the foreign-currency exposure that has historically made GPU-intensive AI workloads prohibitively expensive for Nigerian institutions. The opportunity extends beyond additional hosting capacity. It includes the emergence of a domestic compute environment capable of supporting more advanced AI workloads at scale, in a market that previously had to rent that capability from abroad.

The second opportunity flows from the market structure provision in Pillar 1, and most public commentary has missed its technical dimension. Restructuring an institution to comply with the 25%/15% cap is not primarily a legal exercise in corporate reorganisation. It is data engineering work. Affected institutions must disentangle shared platforms across card issuing and merchant acquiring, separate previously-integrated data flows, build cleaner APIs between newly-separated services, and rebuild identity federation across entities that used to operate as one. This is the kind of architectural work for which Nigeria has a thin specialist base, particularly in regulated-industry data architecture at scale. The institutions and individuals positioned to do this work, especially indigenous data and integration specialists, gain a structural advantage during the 2026 implementation window. The work itself does not appear in colocation revenue numbers or in cloud market sizing, but it determines whether the affected institutions emerge from restructuring with cleaner architectures or with technical debt that constrains them for the next decade.

The third opportunity flows from the combined effect of the UBO disclosure requirement in Pillar 3 and the monthly market-share returns required by Pillar 4. Together, these create a continuous data governance obligation that Nigerian regulated institutions have not historically maintained at the required standard. Compliance requires accurate, audited, structured data about ownership and market position, current enough to support monthly submission and defensible enough to survive regulatory scrutiny. The work this generates is master data management, ownership-tracking systems that resolve through corporate structure to natural persons, real-time market-share calculation, and audit trails designed for regulatory inspection. This is a category of work that is likely to expand significantly in scale as a result of the circular. The firms that build credibility in regulated-industry data governance during the implementation window position themselves as the trusted operators of that capability for years afterwards. If the CBN extends similar reporting and transparency obligations to adjacent sectors, which is the consistent pattern observed in other jurisdictions that began with payments, the same capability becomes increasingly valuable.

None of this resolves the implementation concerns raised in Part I. Aggressive timelines, cost passthrough to consumers, structural grid constraints, and the cloud provider ambiguity all remain real. The opportunity side of the ledger matters because it reframes the strategic question. The CBN circular is generating demand for specific kinds of architectural, AI, and governance work that Nigerian institutions had collectively been postponing. The firms and specialists positioned to do that work are beginning to move. Whether the regulation succeeds in its stated goals depends partly on whether the architectural capacity to absorb it can scale at the pace the deadlines require.

VI. Sequencing and Implementation

This is the question that matters most for a practitioner audience, and the available evidence suggests the answer is yes: this should have been a plan rather than a deadline. Regulations of this magnitude are not new, and the countries that have implemented them successfully, alongside those that have stumbled, leave a useful evidence base.

The most instructive cautionary precedent is India in 2018. The Reserve Bank of India issued its payment data localisation directive in April of that year, giving payment system providers six months to comply. The deadline was missed. Industry concerns about implementation feasibility forced the RBI to issue an extensive FAQ document clarifying the rule. In 2019, the RBI relaxed the original provision to permit data processing abroad, provided the data was deleted from foreign systems within twenty-four hours and stored only in India. Further clarifications followed in September 2021. As recently as the present period, payment system operators in India are still working through implementation guidance, with major banks publishing phased migration plans across different channels and products as they continue to align with the rule.

The Indian experience is not a failure. It is, more accurately, a regulation that landed in strategically correct territory but operationally premature condition, and which required years of phased rollback, clarification, and industry consultation to become workable. The Nigerian circular is structurally similar to the 2018 RBI directive, with one significant difference: India had considerably larger domestic data centre capacity and a substantially more stable electrical grid when its directive landed than Nigeria does today. If the Indian rollout required FAQs, formal relaxations, and seven years of phased implementation guidance to function reliably, the foreseeable trajectory for Nigeria’s circular is likely to follow a similar path.

The most instructive positive precedent is Rwanda. Three features of the Rwandan approach distinguish it from the CBN circular. First, regulation is tied to a multi-year national strategy rather than announced through standalone circulars with short compliance windows, giving industry visibility into the regulatory trajectory years in advance. Second, the infrastructure was built before the regulation required its use, rather than in parallel, so operators could plan capacity expansion against a known and stable roadmap. Third, a regulatory sandbox established in 2022 allowed regulators and industry to test compliance in operational conditions before it became binding, surfacing implementation problems early enough to be addressed through rule design rather than retroactive clarification.

The underlying principle is straightforward. Regulations that depend on infrastructure tend to work when the infrastructure already exists, and tend to fail when the infrastructure is being built in parallel with the rule. India in 2018 demonstrated the second pattern. Rwanda has demonstrated the first.

A more strategic version of the Nigerian circular would have anchored its provisions to a multi-year national milestone rather than to an eighteen-month deadline. One plausible sequencing would have introduced the components in stages, beginning with the provisions that require no new infrastructure and ending with the provisions that depend most heavily on capacity that does not yet exist at the required scale.

In such a sequenced approach, UBO disclosure and the monthly reporting framework could have been made operational immediately, since both require regulatory and procedural changes rather than physical infrastructure. The market structure caps could have been phased in across 2027 and 2028, with formal industry consultation on transition plans and clear runways for affected institutions to restructure or divest where required. Data localisation could have been phased in by data category, beginning with the most sensitive categories such as cardholder data and authentication credentials, and expanding outward in subsequent phases. The migration timeline could have been tied explicitly to documented growth in domestic data centre capacity and grid reliability, with the CBN publishing benchmarks against which the regulatory timeline is measured. Throughout the rollout, a regulatory sandbox could have allowed institutions to test compliance approaches in controlled conditions and the CBN to refine its guidance before final enforcement. The entire programme could have been anchored to a published national digital infrastructure strategy with measurable milestones for power generation, data centre capacity, and skills development.

This approach is not less sovereign than the circular as currently drafted. It is arguably more sovereign in practice, because it produces a rule that holds reliably from the date of enforcement rather than a rule that softens through quiet supervisory discretion in response to implementation pressure. The Indian 2018 directive looks more aggressive on paper than the operational outcome that eventually emerged through years of relaxation. A phased Nigerian version would look less aggressive on paper, but it would bind from the first day of enforcement, which is a stronger position for the regulator to hold than the alternative.

VII. Conclusion

The CBN circular addresses four real problems through a single coordinated instrument, and its strategic logic is coherent and consistent with what serious sovereignty looks like across the data, governance, and oversight layers of a modern payments system. On the strategic dimension, there is little to disagree with.

The implementation dimension is where the difficulty lies. The deadlines are aggressive relative to the underlying capacity. The infrastructure required to comply is being built in parallel with the regulation rather than in advance of it. The cloud provider ambiguity is likely to produce a period of regulatory uncertainty before the CBN is in a position to clarify its interpretation. Compliance costs will be significant, particularly in the early phase, and the question of who absorbs those costs remains unaddressed in the published document.

The most plausible outcomes over the next eighteen months fall into three broad scenarios. In the first, the CBN issues clarifying guidance that treats international cloud provider local zones as compliant infrastructure, foreign hyperscalers accelerate their Nigerian builds in response, institutions migrate workloads to those zones, and data sovereignty is achieved in geographic terms but not in terms of underlying dependency. This is the most likely short-term outcome, given the existing posture of major cloud providers in adjacent African markets. In the second, deadlines are quietly extended through supervisory discretion, as has happened in India and elsewhere when implementation realities have outpaced regulatory ambition. The circular would remain technically in force, but the binding deadline would drift, which is the path of least resistance and is consistent with the CBN’s historical approach to technical compliance challenges in the payments sector. In the third, Nigerian operators including MainOne, Rack Centre, Africa Data Centres, and Nxtra accelerate construction, supported by major energy investments alongside. This is the outcome that would constitute genuine sovereignty in infrastructure terms, but it cannot be achieved on an eighteen-month timeline regardless of policy ambition, which limits its viability as a near-term path.

What the circular reveals, more than anything, is that regulation and infrastructure need to be sequenced together rather than announced separately. Strategic direction without sequenced execution tends to produce aggressive deadlines that quietly soften under operational pressure. Sequenced execution without strategic direction tends to produce capacity that no one uses. The combination of the two is the harder regulatory work, and it is the work that the countries that have built durable digital sovereignty have, in different ways, all done. The next eighteen months will test whether the country can build the implementation timeline to match it.

Sources and further reading:

• Central Bank of Nigeria circular PSS/DIR/PUB/CIR/001/004 (15 June 2026)

• Reserve Bank of India: Storage of Payment System Data (Circular RBI/2017-18/153, April 2018)

• GovInsider: How Rwanda built a credible, execution-ready DPI

• National Bank of Rwanda: Regulatory Sandbox

• TechCabal: Can Nigeria’s data centres power Africa’s AI future?

• Introl: Africa’s AI Data Center Boom, Nigeria, Kenya, South Africa

• African Business: Inside the race to fire up Africa’s power-hungry data centres

• Center for Global Development: Data Localization, a Hidden Tax on the Poor

This is a little different from my usual content, but after spending six weeks as a trainer on an AI accelerator, I wanted to share some of what I covered. The frameworks I covered in my sessions, and that form the basis of this article, were created and shared with me by my colleague Kwadwo Benko, AI Transformation Change Lead in the Microsoft AI Transformation Office. Kwadwo has designed nationally adopted frameworks including A-LEVEL, ALERT, PROTECT, SPARK, REVIEW and INSIGHT, now used across policing, central government and regulatory sectors. His work shaped how I approached teaching this material.

Earlier this year I worked as a trainer on a six-week AI accelerator run by Girls in Marketing. Each week a different trainer covered a different area of AI, including ethics, search, data, automation, and the future of the industry. I delivered two of those sessions. My first session, in week one, focused on mindset and helping marketers reframe their relationship with AI before they started using it more seriously. My second session, in week four, focused on prompting, specifically how to communicate with AI in a way that produces strategic and genuinely useful output.

Teaching both sessions gave me a useful perspective. The barrier most people face with AI is rarely technical. It begins with how they think about the tool, and it becomes most visible in how they instruct it. In this article I want to focus on the prompting frameworks I covered, because they are the most practical thing anyone can take away and apply immediately, regardless of the industry they work in.

Why frameworks matter when prompting

Before getting into the specific frameworks, it helps to understand the difference between prompting and prompt engineering, because they are not the same thing.

The reason frameworks matter is that without one, most people prompt the same way they would type a search into Google. They keep it short, vague, and hope the tool understands what they meant. AI does not work like a search engine. It works more like a collaborator, and a collaborator produces better work when it is given proper context, clear constraints, and a defined objective. A framework gives you a reliable structure for providing those things every time, so the quality of your output does not depend on chance or on who happened to write the prompt that day. This is also what allows teams to build shared prompt libraries and maintain consistency across people and projects.

Below are the three frameworks I covered in the session, in the order I recommend learning them.

Framework one: GCSE, the foundation

GCSE is the best place to start. It is a four-part structure that addresses the most common reasons AI outputs fall short, and it is simple enough to apply straight away.

The G stands for Goal. This is what you actually want produced. Be specific about the output rather than the general topic. “Write something about our new product” describes a topic. “Write a 150-word email announcing our new product to mid-level operations managers” describes a goal.

The C stands for Context. This covers who you are, who the output is for, and why it matters. AI has no understanding of your situation unless you provide it, so the more relevant context you give, the more tailored the result.

The S stands for Source. This is the material the AI should draw from, such as uploaded documents, specific data, or existing content. Pointing it to the right source produces a stronger and more accurate output than letting it generate from nothing.

The E stands for Expectation. This covers the format, length, tone, and structure you need. If you want three options, state that. If it needs to be under 200 words, state that. If the tone should be warm rather than corporate, state that too.

The difference between a weak prompt and a GCSE prompt is completeness rather than complexity. You are giving the AI what it needs to complete the task properly on the first attempt. GCSE works well for single, clearly defined tasks, and it is the fastest way to see an immediate improvement in what you receive back.

Framework two: A-LEVEL, for strategic and complex work

Once GCSE feels natural, A-LEVEL is the logical next step. It is designed for tasks that are more layered or strategic, where you want the AI to operate less like a basic tool and more like a capable senior collaborator.

The first A stands for Assume the Role. Tell the AI who to be, such as a brand strategist, a customer who matches your target audience, or a senior business partner. The more specific the role, the more the output reflects genuine expertise rather than general knowledge.

The L stands for Leverage Tools and Functions. Instruct the AI to use what is available to it, including uploaded documents, memory, and earlier context in the conversation, rather than leaving it to guess.

The first E stands for Enforce Constraints. Build your requirements into the prompt from the start, including brand guidelines, legal parameters, formatting rules, and tone of voice. These should be part of the instruction rather than corrections you apply afterwards.

The V stands for Validate with Iteration. Ask the AI to review its own output, flag anything it is unsure about, and refine the response before returning it. This step meaningfully improves accuracy and reduces the amount of back and forth.

The second E stands for Embed References. Tell the AI to ground its response in real sources or the documents you have provided, rather than producing something that sounds plausible but cannot be verified. This is particularly important for anything data-driven or compliance-related.

The final L stands for Layer Instructions. Structure the prompt to produce several deliverables in one run, such as a strategy document and an executive summary, or a campaign brief and a shorter social version, all from a single well-constructed prompt.

Where GCSE produces one strong output, A-LEVEL produces a workflow. The prompt becomes less of a question and more of a brief, similar to the kind you would hand to a senior colleague you trusted to deliver something independently. In the session, this was where the biggest improvement happened. Once people started treating their prompts as proper briefs rather than quick questions, the quality of their output changed considerably.

Framework three: SPARK, for solving the problem first

GCSE and A-LEVEL are both about structuring your prompt. SPARK is different because it is about structuring your thinking session with AI. It is the framework I find most valuable to teach, because it changes what people use AI for.

Most people use AI reactively. They already know what they need to produce, and they use AI to produce it more quickly. SPARK is built for the situation where you are not yet sure what you need, where the problem is still unclear, and where you need to think it through before you can act.

The S stands for Set the Stage. Define the challenge using “How might we” framing. This opens up the problem rather than narrowing it too early. For a marketer it might be “How might we re-engage lapsed customers without leading with a discount.” The framing shapes everything that follows.

The P stands for Play with Perspectives. Ask the AI to consider the problem from different angles, such as your audience’s point of view, a competitor’s approach, or how a different industry has handled something similar. This is often where the most useful insight appears.

The A stands for Amplify Ideas. Generate as many ideas as possible without evaluating them yet. At this stage, quantity matters more than feasibility, so allow the AI to propose and combine ideas freely before you apply any judgement.

The R stands for Refine and Remix. Now apply critical thinking. Group the ideas from the previous stage, assess them against your real constraints such as budget, timeline, and objectives, and develop the ones with the most potential.

The K stands for Keep it Going. Finish with something concrete, including who owns what, by when, and how success will be measured. Use the AI to draft the action plan and to surface any dependencies or risks you may have overlooked.

SPARK changes how people think about the purpose of AI. It is not only a faster way to produce things you already know you need. It can act as a thinking partner that helps you work out what you need in the first place. That reframe is one of the most useful things people can take from a session like this.

Why this applies beyond marketing

One of the clearest things I noticed across both of my sessions is that none of this is specific to marketers. The same prompting problem appears in every function. HR teams ask for job descriptions and receive generic templates. Finance teams ask for data summaries and receive surface-level observations. Operations teams ask for process documentation and receive something too general to use.

The problem is the same in each case, and so is the solution. Give the AI enough context, clarity, and constraint to do useful work, and use a framework to make sure you are doing that consistently rather than only when you happen to get a good result.

What I would leave you with

The most interesting thing from teaching these sessions was how quickly people improved once they understood that the quality of the output is largely within their control. It does not start with the tool, it starts with how clearly you communicate what you need.

Start with GCSE to build solid foundations. Move to A-LEVEL when the work requires more depth and structure. Use SPARK when the problem itself is what needs working through first. None of this requires a technical background. It requires clarity about what you are asking for, and the discipline to build that clarity into how you use AI every time. That is a skill worth developing in any industry and at any level.

If you want more training

If this article has been useful and you would like to take your AI skills further, I would strongly recommend connecting with Kwadwo Benko directly. Alongside his work at Microsoft, he is on a personal mission to democratise AI skills across the UK, making tools like Microsoft Copilot accessible to anyone, regardless of background or industry.

He offers free, practical AI and Copilot training to schools, charities, community groups, small businesses, and anyone else who could benefit. The sessions are tailored rather than generic, focused on everyday impact across productivity, learning, creativity and careers, and require no prior experience.

His broader goal is to skill 100,000 Black people in AI by 2026, and 500,000 people overall by 2028. If you work in a school, charity, or community organisation, or you know a group that could benefit, you can register through his Free AI & Microsoft Copilot Training Initiative, or connect with Kwadwo directly on LinkedIn.

A brief note before we begin: I write this from an implementation perspective. My work sits closer to data, AI, and digital transformation than to policy analysis, so the argument that follows is less about governance theory and more about what technological change looks like when institutions have to absorb it.

Everyone is racing to get “AI-ready.” But AI was never the destination, it was the warning. The real question is whether your foundations can absorb whatever comes next. No country has fully cracked this yet, but the lessons of the last few years, and the choices Africa’s anchor economies make now, will determine who is ready when the next wave arrives.

The Race Everyone Is Running

Walk into any African ministry of digital affairs or any government innovation office today, and you will hear a version of the same sentence: “We need an AI strategy.”

The numbers explain the urgency. The African Union reported that over 83% of AI startup funding on the continent in the first quarter of 2025 flowed to just four countries: Kenya, Nigeria, South Africa, and Egypt. The AU adopted a Continental AI Strategy in 2024. National strategies followed in a cascade: Kenya in March 2025, Egypt already on its second iteration, Nigeria publishing drafts, South Africa finalising policy in 2026. Globally, the pattern is identical. By one count, the number of new national AI strategies launched in 2024 tripled compared to the year before.

Governments are, in other words, building AI stacks: procuring models, standing up chatbots, drafting governance frameworks, signing memoranda with hyperscalers.

This is understandable, and in many cases it reflects genuine urgency and political will. But it may also be a strategic miscalculation. Not because AI doesn’t matter, but because the AI stack is the wrong unit of planning. Governments that build for AI are building for a technology. Governments that build a dynamic tech stack are building for change itself.

The difference will define which states absorb the next thirty years of disruption and which states are forced to rebuild from scratch each time the next wave arrives.

What AI Actually Taught Us

ChatGPT reached 100 million users in two months. TikTok took nine months to reach the same milestone. Instagram took two and a half years. In twenty years tracking internet adoption, UBS analysts wrote at the time, they could not recall a faster ramp for any consumer application. Generative AI went from research curiosity to policy mandate in roughly eighteen months.

In doing so it ran an unplanned stress test on every government on earth and the results were instructive.

Governments with foundational infrastructure absorbed it quickly. Estonia, Singapore, and the UAE, which had invested in digital foundations for years, plugged AI into existing infrastructure within months. Singapore had treated digital capability as national infrastructure for a decade, with the overwhelming majority of government services already online and a national AI commitment running to 2027. The UAE had appointed a Minister of State for AI in 2017, five years before generative AI became a mainstream conversation; when the wave arrived, governance frameworks were already in place. Estonia added another layer to a thirty-year-old architecture built on one principle: digital identity and interoperable data first, applications later. Estonia is now contracting its post-quantum cryptography transition, preparing for the next wave while others are still working through this one.

Governments without those foundations found AI had nothing to stand on. Without clean data, without interoperable systems, without a functioning identity layer, they purchased tools that couldn’t connect to the information they needed to be useful. The AI itself was not the problem, the infrastructure underneath it was.

The lesson is important, and it points beyond AI: governments that absorbed this wave well had built foundations that didn’t care which technology arrived. They had not built for AI specifically, they had built for change.

That is the dynamic tech stack, and it matters now because AI is not the last wave. Agentic systems that act rather than answer are already deploying. Quantum computing will eventually require a wholesale upgrade of the cryptographic foundations underneath government systems, which is why the EU has set a 2030 deadline for post-quantum cryptography in critical infrastructure. The interval between waves is shrinking and transformation is no longer a project with an end date, it is the permanent operating environment.

The Framework: AI Stack vs. Dynamic Tech Stack

It is worth being precise about the distinction, because the two approaches look superficially similar (both involve cloud, data, and models) but they answer fundamentally different questions.

An AI stack answers: “How do we deploy artificial intelligence?”

A dynamic tech stack answers: “How do we absorb any technology we cannot yet name?”

The five foundations of a dynamic stack, in order of permanence:

1. A trust and identity layer: Verified digital identity for every resident and business, under transparent governance. This is the social contract rendered as infrastructure, and every subsequent technology, from AI agents to autonomous public services, authenticates against it.

2. A sovereign data layer: Data that the state owns, governs, and can move: in open formats, with clear lineage, privacy architecture, and quality standards. Models, vendors, and paradigms will change; the data layer is the one asset that compounds in value across all of them.

3. An orchestration and abstraction layer: Every capability exposed as an API. Every dependency swappable. This is what allowed well-prepared governments to avoid lock-in with any single AI vendor, and what will allow them to plug in quantum compute or agentic systems as they mature, without rebuilding from scratch.

4. An adaptive governance layer: Procurement frameworks that allow modular, specialist vendors rather than decade-long monolithic contracts. Regulation written for capabilities rather than specific products. Institutions that can issue, revise, and retire rules at the pace technology actually moves.

5. A human absorption layer: Workforce capability, leadership literacy, and public trust. Three of the five layers of a dynamic stack are not technical. This is why technology procurement, on its own, has never produced transformation.

A government with these five layers in place does not need to predict what comes next. It needs only to be ready to plug it in.

Africa’s Big Four and the Stack Question

Nowhere is this distinction more consequential right now than in Africa, and nowhere is the decision more visible than in its four anchor economies.

The concentration of digital capacity on the continent is significant. South Africa alone holds roughly 70% of Africa’s total digital infrastructure capacity. Egypt, Kenya, Morocco, and Nigeria account for a further 15%, leaving fifty other AU member states to share the remainder. The four biggest tech economies attract the overwhelming majority of AI investment. Whatever foundational architecture these four choose will shape, in practice, the trajectory of digital governance across the continent.

Before assessing each, it is worth stating the shared context honestly. All four are navigating infrastructure constraints that most wealthy-nation comparisons ignore: unreliable power grids, limited connectivity outside urban centres, thin technical workforces, legacy systems inherited from pre-digital eras, and the ongoing reality that nearly half a billion Africans still lack foundational legal identity. What these governments are attempting, under these conditions, is genuinely difficult. The question is not whether they are working hard enough. It is whether the strategic direction, given the constraints, is one that builds durable foundations or risks creating new dependencies.

The tension is visible in the policy record itself. Nigeria, Egypt, and Kenya have each published draft AI policies since early 2025 that explicitly name dependence on foreign technology companies as a threat to national security, while each continues to rely on those same companies for compute, capital, and expertise. This is not hypocrisy, it is the difficult reality of building sovereign capability in a world where the infrastructure you need is largely owned by others. The dynamic stack framework is one way to navigate that tension: own the permanent layers, rent the transient ones, and reduce dependency over time.

The geopolitical context sharpens the stakes. The Atlantic Council and others now describe a global “battle of the AI stacks”, with the United States and China actively competing to export their full technology verticals (chips, cloud, models, applications) to third countries. A government that adopts an exported AI stack wholesale is not building national capability. It is, in effect, leasing it. The terms of that lease, including the upgrade cycle, the data conditions, and the exit costs, are set elsewhere.

The dynamic stack is the structural alternative: own the identity layer, the data layer, and the governance architecture, and use market competition among vendors for the layers that change faster. With that structure in place, no single provider becomes indispensable.

How do the big four measure against that standard?

Nigeria is building the identity layer at genuine scale. With more than 130 million National Identification Numbers enrolled, Nigeria operates Africa’s largest digital ID system, backed by roughly $430 million from the World Bank and European partners. It is migrating to MOSIP, an open-source identity platform, a deliberate move toward reducing vendor dependency. The NIN already underpins SIM registration, banking verification, academic registration, and social protection schemes. This is meaningful progress on a genuinely hard problem at enormous scale. The next challenge, which Nigeria is actively grappling with, is building the ecosystem layer on top of that foundation: actual usage of digital credentials remains low, services remain fragmented across agencies, and the country currently loses an estimated $850 million a year to foreign hosting and digital service dependency. The foundation is real, the work of converting it into a functioning stack is ongoing.

Egypt is making the case for sequenced, state-led delivery. Egypt stands out in continental assessments as one of the few countries with an AI strategy that has moved from policy document to active implementation, now in its second phase, with applied work in education, energy efficiency, and water management. The strength of Egypt’s approach is its clarity of sequencing and the institutional capacity to execute. The open question, from a dynamic-stack perspective, is how the architecture evolves as it matures: centrally coordinated systems can deploy with speed and coherence, but the most durable stacks also create conditions for distributed innovation, the startups and unexpected applications that emerge when infrastructure is open and accessible to builders beyond government.

South Africa has the infrastructure and is now building the architecture to match. Holding the large majority of the continent’s data-centre and connectivity capacity, South Africa is Africa’s most significant compute base. That is a considerable asset, and one the continent depends on. The policy process has taken longer than the infrastructure investment, with the national AI policy finalised in 2026 after an earlier draft was withdrawn for revision. That revision process, working through the implications carefully rather than publishing prematurely, reflects the seriousness with which South Africa is approaching the governance question. The challenge now is ensuring that formidable physical infrastructure is matched by the integration layers, the APIs, the identity architecture, the data standards, that turn raw capacity into a stack others can build on.

Kenya presents the most layered picture, because it arrived at some foundational elements earlier than almost any comparable economy, through a path nobody designed.

Kenya: An Early Foundation, and the Work Still Ahead

In the early 2000s, Kenya had limited formal financial infrastructure outside its major cities and one of the most unbanked populations in the world. What it had was widespread mobile phone adoption, and a population that began using prepaid airtime as an informal currency, trading minutes for cash and goods in the absence of accessible banking. Safaricom and Vodafone formalised that behaviour in 2007 as M-Pesa, and what followed became one of the most studied cases in development economics.

Kenya bypassed the conventional pathway of branch-banking infrastructure entirely. Mobile money accounts came to outnumber traditional bank accounts. Research from MIT and Georgetown estimated that M-Pesa alone lifted nearly 200,000 Kenyan households out of extreme poverty. A network of over a hundred thousand agents extended financial services into communities that physical infrastructure never reached.

Viewed through the dynamic stack framework, something deeper is visible. Kenya had inadvertently developed early versions of three foundational layers that most countries, including wealthy ones, were still building:

• A trust layer: Long before digital government services became standard elsewhere, ordinary Kenyans developed comfort and confidence in digital financial transactions. That cultural familiarity with digital services became, over time, one of Kenya’s most valuable assets for subsequent innovation.

• A transaction and data layer: A real-time record of economic activity that later made it possible to build AI-driven credit scoring for populations with no formal financial history.

• A distribution layer: The agent network became, in effect, a last-mile delivery infrastructure for any future digital service.

Subsequent developments built on that base. The government’s eCitizen platform, now integrating thousands of services and processing millions of applications, works in part because digital payment infrastructure and a baseline of digital trust already existed. Kenya’s strength in fintech, its position as a leading destination for AI venture investment on the continent, its March 2025 National AI Strategy positioning it as East Africa’s hub for AI development: all of it reflects the cumulative effect of a foundation that was laid early, even if unintentionally.

The more complex part of Kenya’s story is what the two decades since M-Pesa reveal about the difference between an accidental foundation and a deliberate, open stack.

Kenya still does not have a fully functioning national digital identity system. Huduma Namba, the initiative to consolidate identity records into a unified platform, faced significant legal challenges around privacy and data governance, and its successor is still in rollout. Government data remains fragmented across agencies. And the payments infrastructure that could have evolved into open national infrastructure remains largely a private platform: Kenyan developers have noted that India’s UPI, an open, publicly governed payments rail, has enabled thousands of fintechs to build at near-zero cost, while Kenya’s ecosystem depends on a dominant private actor as gatekeeper.

None of this diminishes what Kenya built. The point is a structural one; the layers that arrived by accident were built to serve a product, not to serve as public infrastructure. Converting them into a sovereign, open foundation requires intentional decisions that market forces alone will not produce. That is the work Kenya is now undertaking, and the question of whether it does so through open architecture will shape how much of its early advantage compounds into the next wave.

The Kenyan experience offers a clear and generalisable principle: an early foundation is a significant advantage, but only open architecture sustains the compounding.

Rwanda: What Deliberate Architecture Looks Like

Rwanda is not among Africa’s big four by economic scale, but it belongs in this analysis because it represents the clearest available example of what intentional, layered digital public infrastructure looks like in an African context.

Beginning with the IremboGov platform in 2015, Rwanda set out to build digital government as a governance strategy, not as a technology procurement exercise. IremboGov now hosts over six hundred government services, with roughly 80% of applications completed fully online. Rwanda’s newer payments infrastructure settles transactions across all banks and mobile wallets in under fifteen seconds and was deliberately built as open-source, openly governed infrastructure, designed to integrate savings cooperatives and microfinance institutions that commercial platforms had historically excluded. Interoperable electronic health records are being connected into the same identity and service architecture.

The significance of Rwanda’s approach is less about the specific platforms and more about the sequencing logic: identity first, then payments, then services, then data interoperability. And at each layer, a deliberate choice toward openness and public governance rather than proprietary vendor architecture. Rwanda is not leading with AI procurement, it is building the layers on which AI, and whatever follows it, will run.

Rwanda’s model is not without its own questions. It is a small, highly centralised state, and the degree of central coordination that made its digital strategy coherent may not translate directly to larger, more federated economies. And it too remains years from a complete stack. But as an illustration of architectural intent, of what it looks like to build layers deliberately and openly, it is the most instructive example currently available on the continent.

The contrast with Kenya is instructive precisely because both countries started from comparable positions. Kenya had an earlier and larger foundation, arrived at accidentally. Rwanda began its deliberate build later, with fewer resources. Neither has a complete stack today; but Rwanda’s trajectory is more predictable, because it is being driven by architecture rather than by accident.

The more pressing question for the big four is not whether they have the resources Rwanda lacked, but whether they are directing the resources they have toward the permanent layers or the transient ones.

The Choice Ahead

The AI wave was an accelerated version of a pattern that has repeated with every major technology cycle: the organisations and governments that absorbed it best were not the ones that had predicted AI specifically. They were the ones that had built foundations capable of absorbing change in general. Estonia, Singapore, and the UAE benefited from decisions made years and in some cases decades before generative AI existed. Kenya benefited, partially, from an accident that happened to produce foundational layers.

What is different now is the pace. The interval between technology waves is compressing, and the consequences of being caught unprepared compound with each cycle. A government still rebuilding foundational infrastructure when the next wave arrives will face a more compressed timeline than the one AI provided.

Africa’s big four face a version of this decision that will have consequences well beyond their borders, given the concentration of the continent’s digital capacity in their hands. The path of least resistance is to continue acquiring AI capabilities from the major technology providers, achieving real and visible results in the near term, while the underlying architecture remains dependent on external infrastructure and external terms. Their own published policies name this risk clearly. The harder path is to prioritise the permanent layers: sovereign identity architecture, open data standards, interoperable payments rails, governance frameworks built for capabilities rather than products, and the procurement reforms that allow modular, competitive vendor relationships rather than long-term lock-in. That path is slower to show results, it is also the one that compounds.

The $720 million Cassava-Nvidia partnership to build AI infrastructure across the big four economies represents exactly the kind of investment the continent needs. The question is whether the states receiving that infrastructure own the layers that allow them to use it on their own terms, or whether they are, in effect, plugging external infrastructure into external architecture with limited national control over either.

One continental analysis asked whether Africa’s rush of AI strategies was putting the cart before the horse. The more precise framing may be this: AI is one horse among many that will need to be harnessed over the coming decades, the question is who builds the road.

For government and policy leaders, the practical agenda follows from the framework:

Audit by layer, not by technology: Rather than asking “what is our AI readiness?”, ask: who governs our identity layer? Is our data held in formats we control? Can we change any technology vendor within a reasonable timeframe? Is our cryptographic infrastructure designed to be upgraded as standards evolve?

Fund the permanent, create market conditions for the transient: Public investment belongs in identity, data governance, open payment rails, and human capability. Models, applications, and services are best sourced competitively. Nigeria’s move toward open-source identity infrastructure and Rwanda’s open payments architecture are examples of this logic applied. The test for any new investment: does this increase or decrease our structural dependency?

Design governance for capabilities, not for products: Frameworks written specifically for current AI tools will need to be rewritten when those tools change. Governance anchored in principles (accountability for automated decisions, rights over personal data, transparency in public-sector use) will apply to AI, to agentic systems, and to whatever follows.

Treat public trust as a foundational layer: The legal and governance challenges that slowed Kenya’s identity programme and required South Africa to revise its AI policy are not obstacles to digital transformation. They are signals that the governance layer needs to be built with the same rigour as the technical one. Public trust, once lost, is the hardest layer to rebuild.

The next technology wave is already forming. Governments will not choose whether it arrives, they will only determine what it lands on.

Build for the technology you can see, and you will rebuild with every cycle. Build for the capacity to absorb change, and the investment compounds across every wave that follows.

Sources and further reading:

• African Union: Africa Declares AI a Strategic Priority (2025)

• Microsoft: Global AI Adoption in 2025 (2026)

• Oxford Insights: Government AI Readiness Index 2025

• Carnegie Endowment: Understanding Africa’s AI Governance Landscape (2025)

• Rest of World: Africa’s AI plans still depend on Google, Microsoft, Nvidia, Meta (2026)

• Atlantic Council: Eight ways AI will shape geopolitics in 2026 (2026)

• GovInsider: How Rwanda built a credible, execution-ready DPI (2026)

• Fintech Association of Kenya: At 18, M-Pesa Faces Its Adult Future

• PMI: M-Pesa, Most Influential Projects

• Biometric Update: Nigeria’s digital ID passes 130 million (2026)

• Global Center on AI Governance: Are African Countries Putting the Cart Before the Horse? (2025)

• Cybernetica: Estonia’s transition to quantum-safe e-governance

• In On Africa: Governing the digital economy (2026)

Microsoft Build 2026 wrapped this week, and the theme running through every announcement was the same: we are moving from AI as a tool you use to AI as an infrastructure layer that works continuously on your behalf. Satya Nadella framed it as a question every developer, company, and institution now needs to answer:

This article is my attempt to pull out the announcements I found most significant, explain what they actually mean in plain terms, and offer my own take on the implications for businesses, individuals, and specifically for Africa.

One caveat worth stating once, upfront, and letting stand for every Africa section that follows: access to technology does not equal adoption on the continent. Throughout this article I try to distinguish between what is technically possible and what is institutionally plausible, because those are two different things. Procurement cycles, compute availability outside major hubs, data sovereignty regulations, and varying levels of enterprise maturity all shape how and whether these tools get deployed in practice. The opportunity is genuine. The path to realising it is uneven.

This is a detailed article. Use the links below to navigate to the sections most relevant to you.

What's covered:

• Microsoft Discovery

• Majorana 2

• MAI Models: Microsoft’s In-House AI Family

• Frontier Tuning

• Microsoft Scout

• Project Solara

• A Final Thought

Microsoft Discovery

Science has always been constrained by bandwidth. A researcher can only run so many experiments, test so many hypotheses, and review so much literature in a working day. Microsoft Discovery, now generally available, is built around the idea that agentic AI can change that equation fundamentally.

At its core is the Discovery Engine: a system of specialised agents that mimic the scientific method across large bodies of knowledge, generating hypotheses, validating theories, and running that loop continuously without waiting to be prompted. This is not a search tool or a summarisation layer. It is an autonomous research collaborator that can work through problems at a speed and scale no human team can match. GSK is iterating on drug discovery. Syensqo is accelerating semiconductor R&D.

For enterprises in regulated industries, including food, energy, and pharmaceuticals, the governance layer matters as much as the capability. Discovery ships with built-in controls for security, transparency, and audit, which means it is deployable in environments where a general-purpose AI tool would never pass compliance review.

There is also a preview of a standalone Discovery app, designed to bring these capabilities to individual researchers, students, and academic labs. You only need a GitHub Copilot account to get started.

For enterprises, the value is in R&D acceleration and the reduction of the time between problem identification and validated solution. Industries with long research cycles, including energy, materials science, and pharmaceuticals, stand to gain the most. For individual researchers and students, the app version is the more significant announcement. Access to agentic research infrastructure that previously required institutional resources is a genuine leveller.

For Africa, the opportunity is most visible in healthcare, where local disease profiles, drug resistance patterns, and demographic data remain underrepresented in global research, and in agriculture, where climate variability is creating highly localised challenges that global models are poorly equipped to address. Africa carries a disproportionate share of the world’s disease burden but operates with a fraction of the research infrastructure available in Europe or North America. A tool that allows a small team to run hypothesis-generation and validation loops at the scale of a much larger lab has real potential here.

The technically possible impact is significant. The institutionally plausible adoption path is narrower. Most African universities operate on procurement cycles that are slow relative to how quickly these tools are evolving, and compute availability outside a handful of major cities remains inconsistent. The standalone Discovery app lowers the barrier considerably, but realising its potential will depend on whether institutions can build the data infrastructure and research workflows to use it meaningfully, not just access it.

Majorana 2

Quantum computing has spent years promising a future that always seemed just out of reach. Majorana 2, unveiled at Build 2026, is the most credible signal yet that the timeline is compressing.

The headline numbers are striking. Majorana 2’s qubits are 1,000 times more reliable than the previous generation. Where conventional quantum systems measure qubit lifetimes in microseconds, Majorana 2 offers an average of 20 seconds, with instances lasting up to a minute. Microsoft uses a useful analogy: comparable to inventing a phone battery that lasts nearly three years instead of a single day. The chip is also remarkably small, with each qubit measuring just one hundredth of a millimetre, putting the long-term goal of a million qubits on a single chip within the realm of engineering rather than theory.

What makes this announcement particularly interesting is how it got here. Microsoft used its own agentic AI, specifically Microsoft Discovery, to accelerate the chip’s development, automating testing and fabrication workflows in ways that compressed the timeline significantly. The target of a commercially viable quantum computer by 2029 has been pulled forward from earlier projections. Microsoft’s quantum roadmap now includes a Quantum Starling prototype by 2028, integrating 200 logical qubits, as a step toward large-scale commercialisation by 2033.

It is worth noting that scrutiny remains. Researchers have previously called for reproducible data on topological Majorana qubits, and independent verification of these claims will matter as the roadmap progresses.

For most businesses, Majorana 2 is not an immediate operational concern. It is a signal about where computational power is heading and what will become possible over the next decade. Industries with genuinely intractable optimisation problems, including logistics, pharmaceuticals, financial modelling, and materials science, should be tracking this closely and beginning to build internal understanding of where quantum advantage will apply to their specific workflows. For individuals, particularly those in research, engineering, or data-intensive fields, quantum literacy is becoming a meaningful career differentiator.

The applications Microsoft highlights, including global health, food supply, sustainability, and energy production, map closely onto Africa’s most pressing structural challenges. Quantum computing at commercial scale could model drug interactions for diseases disproportionately affecting African populations, optimise agricultural yields under climate stress, and run energy grid simulations that current classical computing cannot handle at the required complexity.

The technically possible impact is long-term but genuinely relevant. The institutionally plausible adoption path is less about the technology itself and more about readiness. The 2029 commercialisation target means real-world deployment is still years away, which is actually an opportunity: African research institutions and governments with long planning horizons have time to build quantum literacy and put in place the institutional frameworks that would allow them to deploy these tools when they arrive, rather than scrambling to catch up after the fact. The countries that treat this as a planning question now rather than a procurement question later will be better positioned.

MAI Models: Microsoft’s In-House AI Family

For years, Microsoft’s AI story has been synonymous with OpenAI. GPT-4, DALL-E, Codex: the relationship has been so central that it was easy to assume Microsoft’s AI ambitions were largely derivative. The MAI model family, expanded significantly at Build 2026, tells a different story.

The headline addition is MAI-Thinking-1, Microsoft’s first reasoning model. At 35 billion active parameters it is mid-sized by frontier standards, but it was built from the ground up on clean data, without distillation from third-party frontier models. That detail matters significantly. It means Microsoft owns the full training lineage, which has direct implications for enterprise compliance, data governance, and long-term model control. The model is designed for complex multi-step instructions, long-context reasoning, and code generation, and it is positioned explicitly on cost efficiency rather than raw capability.

The broader family now covers the full stack of developer needs. MAI-Image-2.5 handles both text-to-image and image-to-image workloads. MAI-Transcribe-1.5 delivers state-of-the-art speech-to-text accuracy. MAI-Voice-2 now supports more than ten additional languages. MAI-Code-1-Flash is purpose-built for GitHub Copilot and VS Code. All image, transcription, and voice models are generally available now on Microsoft Foundry and MAI Playground.

For enterprises already inside the Microsoft ecosystem, the MAI family reduces dependence on third-party model providers and gives procurement and compliance teams a cleaner story to tell. For developers, MAI-Code-1-Flash and MAI-Thinking-1 offer capable, cost-efficient alternatives to frontier models for workloads that do not require the heaviest compute. Cost is a meaningful part of this story: MAI-Thinking-1 is explicitly positioned on low-token efficiency, a direct response to the compounding expense enterprises face when running frontier models at scale. For organisations that have started to feel the real cost of AI in production, a capable mid-sized model that performs well on targeted workloads at lower token cost is not a compromise but the practically sound choice.

It is also worth being clear about the strategic significance of MAI-Thinking-1 being built without distillation from third-party frontier models. This is not a technical footnote. It represents Microsoft building the capability to own its full AI stack, from silicon through to model, application, and agent. The OpenAI partnership remains commercially valuable, but Microsoft is no longer building its AI future on a single relationship it does not control. With full ownership of its inference pipeline, Microsoft can price aggressively in ways that a reseller arrangement with OpenAI would not permit.

MAI-Voice-2's expanded language support is the most immediately relevant announcement in the African context. The addition of new languages and voice options matters in markets where voice is often the most practical interface for AI, particularly in contexts with lower literacy rates or limited keyboard access. MAI-Transcribe-1.5 similarly opens possibilities for transcription and documentation in local languages that have been poorly served by existing tools.

Of the announcements in this article, the MAI family sits closest to the institutionally plausible end of the adoption curve. The models are accessible through existing cloud infrastructure, do not require significant on-premise compute, and the voice and transcription tools in particular address problems that African enterprises and developers are already trying to solve. The data sovereignty argument is also relevant here: access to a range of models without dependence on a single provider gives African institutions more negotiating room as governments across the continent become more deliberate about how they engage with global AI infrastructure.

Frontier Tuning

Foundation models are powerful, but they are built for everyone, which means they are optimised for no one in particular. They do not know your internal terminology, your APIs, your compliance requirements, or the specific way your organisation makes decisions. Frontier Tuning, now open to select early partners, is Microsoft's answer to that gap.

Rather than adapting your workflows to fit a generalised model, Frontier Tuning lets you train domain-specific models on your own data, workflows, policies, and house style. The result is a model that understands your context from the start, which means higher quality outputs, fewer correction loops, and lower token costs over time.

What makes Frontier Tuning architecturally significant is that it runs entirely within your own tenant. Your data does not leave your environment. The model you produce is yours. It ships with enterprise reinforcement learning environments, meaning the model learns directly from real workflows, task completions, decision reviews, and validated outcomes, rather than from synthetic benchmarks. You own both the model and the learning loop.

For enterprises, this is one of the most commercially significant announcements at Build. The combination of domain-specific tuning, tenant-level isolation, and an owned learning loop addresses three of the most common objections to deploying AI in regulated or sensitive environments: performance, data sovereignty, and auditability. Organisations in financial services, legal, healthcare, and government now have a credible path to models that perform at frontier level on their specific workloads without the risk of sending sensitive data to a third-party training pipeline. For individuals and smaller teams, the early partner model means access is limited for now, but the direction of travel is clear.

It is also worth being clear-eyed about the lock-in dimension. Once an organisation has invested in tuning a model on its own data and workflows, switching costs increase considerably. Frontier Tuning is genuinely empowering, and it is also a well-designed retention mechanism. Customers should understand what they are building and who benefits most from the compounding value of that learning loop over time.

The most compelling use cases on the continent are in sectors where global foundation models consistently underperform due to data gaps. Healthcare is the clearest example. A hospital network in Lagos or Nairobi running a model fine-tuned on its own patient data, local disease profiles, and clinical workflows will get meaningfully better outputs than one relying on a generalised model trained predominantly on Western medical literature. The same logic applies to financial services, where local credit behaviour, informal economy patterns, and regulatory frameworks differ significantly from the training distribution of most foundation models.

The gap between technically possible and institutionally plausible is wide here. Frontier Tuning requires clean, well-structured organisational data to be effective, and that is a significant constraint in markets where data infrastructure is still being built. Outside of the most mature enterprises in fintech and telecommunications, the institutional readiness required to deploy this effectively is not yet widespread. That does not diminish the opportunity, but it does mean the adoption curve will be steeper than the technical capability suggests.

Microsoft Scout

Most AI at work today is reactive. You open a tool, type a prompt, get an output, and close it. Microsoft Scout is a different proposition. It is Microsoft's first Autopilot agent, one that operates continuously in the background, under its own governed identity, taking action on your behalf across Teams, Outlook, OneDrive, SharePoint, and your local device.

Every action Scout takes is attributed to a governed Entra identity, not a shared anonymous service account. That means everything it does is traceable, auditable, and operates within the policy rules an administrator sets. Scout also runs on Work IQ, a context engine that learns the people, files, workflows, and patterns specific to how you work.

Scout can monitor a GitHub discussion, identify the right feature owners across Microsoft 365, and open tracking conversations in Teams without being prompted. It can check your calendar, identify conflicts, and set an out-of-office block from a single instruction. The cumulative value of an always-on agent comes less from any single dramatic action and more from the continuous reduction of small coordination tasks that collectively consume a significant amount of time. Scout is currently available to Frontier organisations through an early experimental release.

For businesses, Scout represents the beginning of a meaningful shift from AI as a productivity tool to AI as an operational layer. A productivity tool makes individuals faster. An operational layer changes how work flows through an organisation. The compounding effect of removing coordination overhead at scale, across every employee, every day, is easy to underestimate. For individuals, the more immediate question is about trust and calibration. Scout’s value depends entirely on Work IQ learning your context accurately, and that takes time. Early adopters should expect a settling-in period and invest in setting clear policy boundaries before deployment.

The opportunity for Africa is in organisations managing high volumes of coordination work across distributed teams, which describes a significant proportion of how large African enterprises, NGOs, and government agencies actually operate, from pan-African banks coordinating across Lagos, Nairobi, and Johannesburg to NGOs managing distributed field teams in East and West Africa. An agent that can monitor communications, track task ownership, and surface the right information to the right person without manual chasing has real operational value in environments where coordination overhead is high and administrative capacity is stretched.

Of all the announcements in this article, Scout sits closest to the institutionally plausible end of the adoption curve for a specific reason: it does not require significant new data infrastructure or regulatory navigation to get started. It operates within tools that large African enterprises are already using, and the governed identity architecture addresses the accountability concerns that would otherwise slow deployment in institutional settings. The more realistic constraint is organisational: Scout requires a level of Microsoft 365 maturity and IT governance that is present in larger enterprises and multinationals operating on the continent, but less common in mid-sized or public sector organisations. For those institutions, the value is clear but the readiness gap is real.

Project Solara

Project Solara is Microsoft's early-stage answer to a hardware question the agent era is beginning to ask: what does a device purpose-built for ambient AI actually look like?

The two reference designs shown at Build give a concrete sense of the direction. One is a badge device, wearable and voice-first, powered by next-generation Qualcomm silicon, designed for the moments when you are away from a desk. The other is a desk device, always on and contextually aware, powered by a MediaTek IoT SoC, designed to support thinking and planning without requiring you to open an application or pick up a phone. The platform is explicitly open and multi-agent, designed as infrastructure for an ecosystem rather than a single assistant. It is important to note that these are concept reference designs, not shipping products.

For businesses, the desk device concept points toward a future where dedicated agent hardware becomes a standard fixture of the workplace, separate from the laptop or phone and always contextually aware. It is worth tracking as an early signal of where enterprise hardware procurement may be heading. For individuals, the wearable concept points toward a future where the interface to AI becomes as unobtrusive as a badge clipped to a collar, shifting the nature of the interaction from deliberate prompting to continuous ambient assistance.

The badge device concept has a stronger product-market fit in Africa than it might initially appear. Voice-first, hands-free interaction is more practical in environments where screen-based interaction is constrained by cost, connectivity, or the physical nature of the work being done. A field worker, a healthcare community worker, a smallholder farmer: these are users for whom ambient AI that can answer questions and log information without requiring them to stop and navigate an interface represents a genuinely different kind of access. The open, multi-agent architecture also gives African developers more room to build locally relevant experiences on top of the hardware layer when it eventually arrives. That said, meaningful adoption will depend on connectivity infrastructure, hardware affordability, and a mature local developer ecosystem, none of which are uniformly in place across the continent today.

A Final Thought

Taken together, the announcements at Build 2026 tell a coherent story. Microsoft is assembling a full-stack AI platform, with its own models, its own quantum roadmap, its own agent runtime, its own ambient hardware, and its own tuning infrastructure. Each piece is designed to work with the others, and each reduces Microsoft’s dependence on any single external partner.

For institutions across Africa, the opportunity is real but so is the risk of engaging with this stack on terms set entirely by others. This is worth being concrete about. When you tune a model on your organisational data, Microsoft’s infrastructure gets better at serving organisations like yours. When Scout learns your workflows, that context lives inside a governed Microsoft environment. When your research institution runs Discovery, the hypotheses and validation loops generated on your data compound inside a platform you do not own. None of this makes these tools less valuable. It makes the terms of engagement more important.

The questions worth asking are specific: who owns the model after tuning? What data leaves your environment and when? What happens to your learning loop if you switch providers? And where does pricing power sit as these tools become operationally embedded?

The value of any of these tools will ultimately be determined not by their sophistication, but by who controls the data, the workflows, and the learning loops that run on top of them. The earlier you ask that question, the more leverage you will have over the answer.

On 30 May 2026, I was invited to attend the Oxford Frontier AI Forum 2026, hosted by the Oxford AI Society. One of the sessions that stayed with me was the Sovereign AI Seminar: an exploration of UK sovereign AI and institutional leadership. The room was full of investors, engineers, founders, and policy thinkers, and the discussion kept circling back to a deceptively simple question: what does it actually mean for a country to have sovereign AI?

The conversation was focused on the UK, but the questions it raised apply just as directly to Africa. This article recaps the key themes from that discussion and asks whether the frameworks being debated at forums like this one translate to a continent navigating the same questions with a very different set of starting conditions.

One answer kept surfacing, dressed in different words: sovereignty is not self-sufficiency, it is leverage.

That framing should matter enormously to Africa. Not as a borrowed European concern, but as a lens through which the continent’s own digital future looks clearer and more urgent.

Rethinking Sovereignty

Most conversations about AI sovereignty start in the wrong place, with infrastructure tallies, GPU counts, and headline national strategies. A more useful starting point is a simpler question: what does leverage actually look like, and where does Africa already have it?

A metaphor from one of the speakers framed the choice well. At one end sits the Tower of Babel: centralised, hyperscaler-led AI, powerful and resource-intensive, concentrated in few hands. At the other sits Jerusalem: distributed, human-scale, built brick by brick. The instinct for many African nations has been to reach for Babel, to chase data centres and headline strategies, when the numbers suggest a Jerusalem path may be more durable, and more honest about where real leverage lies.

The panel was direct: no country achieves full AI independence. The EU depends on American chips. America depends on Taiwanese manufacturing. The UK built a world-class AI sector without owning a single frontier model from scratch. What matters is not control of every layer. What matters is having enough leverage at the right layers to protect your interests and deliver value to your people.

Where Africa Actually Stands

The numbers are sobering, but they are also clarifying. Africa currently holds less than 1% of global data centre capacity. An estimated 80% of the continent’s internet traffic is processed on foreign servers, most of them in Europe. Nigeria, Kenya, and South Africa together account for 41% of Africa’s 223 data centres, leaving the rest of the continent, 35 countries and hundreds of millions of users, sharing fewer than 130 facilities.

In April 2025, 54 African nations signed the Africa Declaration on Artificial Intelligence in Kigali, expressing a desire to establish a $60 billion AI Fund targeting infrastructure, talent, and startups. A bold statement of intent. Over a year later, governance structures, funding sources, and implementation details remain largely undefined, while three-quarters of all global AI venture capital continues to flow to the US alone. The gap between ambition and architecture remains wide.

There are, however, signs that homegrown actors are starting to fill it. In July 2025, MTN Nigeria launched the Dabengwa Sifiso Data Centre in Lagos, currently the largest Tier III facility in West Africa, with cloud infrastructure the company claims rivals offerings from Amazon, Microsoft, and Google. The $235 million project, led by MTN’s infrastructure subsidiary Bayobab, is explicitly designed for AI processing readiness. MTN has since confirmed South Africa and Nigeria as priority markets for greenfield data centre development under its Ambition 2030 strategy. This matters not just as a business story, but as a sovereignty signal: an African-headquartered operator building infrastructure on the continent, under African jurisdiction, rather than waiting for a hyperscaler to arrive with conditions attached.

The contrast with the Microsoft-G42 Kenya project is instructive. That facility, positioned as a flagship of East Africa’s AI ambitions, stalled when Kenya’s government pulled back from the computing commitments the companies demanded. President Ruto’s explanation was blunt: powering it would require shutting off electricity for half the country. This is not a story about African failure. It is a story about what happens when a sovereignty strategy is built around the Babel model, chasing hyperscale infrastructure rather than building from genuine leverage.

Energy is the constraint that sits underneath this tension. Data centres are not abstract digital assets; they are physical systems dependent on stable and scalable electricity supply. In many African contexts, compute ambition is now advancing faster than the energy infrastructure required to sustain it. Until that gap is addressed, sovereignty strategies will continue to collide with physical limits, regardless of capital, intent, or partnership structure.

The Leverage Africa Already Has

If sovereignty is about leverage rather than self-sufficiency, the question changes entirely. It stops being can Africa build its own large language models? and starts being where in the AI stack does Africa have real, durable power?

The answer is more interesting than most conversations admit.

Data diversity: African languages, healthcare demographics, agricultural contexts, and financial behaviours represent training data that global AI systems are acutely underrepresented in. Rwanda’s ICT Minister Paula Ingabire made the point plainly in April 2026: AI tools trained on foreign datasets misdiagnose diseases, misread economic trends, and misclassify agricultural outputs. That is not a gap to apologise for. It is a leverage point. Whoever builds AI systems that work reliably across Swahili, Yoruba, Hausa, Amharic, and Igbo will not just serve Africa. They will build something the world’s largest labs have consistently failed to.

Regulatory positioning: Nigeria’s National Data Protection Commission issued its largest fine to date in July 2025, penalising Multichoice for unlawful data transfers. Kenya’s Data Protection Commissioner has similarly begun enforcement. By early 2026, 44 African countries had implemented data protection laws. The regulatory infrastructure exists. The question is whether it is deployed strategically, to create leverage, or bureaucratically, to create friction.

The open-source opportunity: Several African AI initiatives are being built on open-source models fine-tuned on local data, fitting models to the demographic profile of local healthcare systems, training on local languages, and reflecting local context. This is not a consolation prize for lacking compute. It is, arguably, the smarter architecture. The UK’s own competitive edge in AI is not in the size of its models. It is in the quality of its applications and its talent for doing more with less. Africa can own that same position.

Of these three leverage points, regulatory positioning is the most immediately actionable: 44 countries already have the legal infrastructure and the question is political will. Data diversity is the most strategically durable, compounding over time in ways outside actors cannot replicate. The open-source path is the most practically accessible right now, requiring neither large capital nor government coordination to begin. Energy remains the binding constraint underneath all three. Sequence matters: build on open-source foundations while constructing the data and regulatory assets that will eventually make African AI genuinely irreplaceable rather than merely locally relevant.

The Middle Power Question

At the Forum, one of the more interesting moments was a brief reference to Canada, listed alongside Singapore as a potential partner for the UK to work with. The UK, the panel noted, is a middle power. It does not have the scale of the US or China, but it has real assets: research talent, a strong regulatory voice, and an existing network of trusted allies. The question is not whether middle powers can play in AI. It is how they play, and on whose terms.

Most African nations are not yet middle powers in the AI stack. They are earlier than that, still building the foundations of digital infrastructure, regulatory capacity, and technical talent that middle power status requires. But that is precisely what makes the framing useful. The middle power position is not where Africa starts. It is where the continent could realistically aim, if the right choices are made now. The instinct for many African nations is to resolve this by picking a side: align with US hyperscalers, or look east toward Chinese infrastructure investment. But the Oxford discussion gestures at a third path, one built on collective positioning, shared leverage, and clarity about where your genuine strengths lie.

The African Union’s Continental AI Strategy, the AfCFTA Digital Trade Protocol, and Smart Africa's AI Council are the architecture of exactly that kind of positioning. The risk is not that Africa lacks the frameworks. It is that the frameworks remain disconnected from each other, and from the commercial and technical actors who could give them teeth. What a nascent power building toward middle power status actually needs is not more strategies. It is intentional choices, made clearly, and defended across political cycles.

What the Middle Ground Actually Looks Like

Morocco may offer the clearest proof of concept so far. In June 2025, a consortium announced a 500-megawatt, renewables-powered AI infrastructure project on the Atlantic coast, under Moroccan jurisdiction and not subject to the US CLOUD Act. It is designed to export compute across Europe, the Middle East, and Africa. That is not self-sufficiency. It is strategic positioning at a chokepoint.

Nigeria and Kenya are attempting something similar at the policy layer. Both released national AI strategies in early 2025 that explicitly name dependence on US tech companies as a strategic risk. Kenya hosted a joint OECD-UK AI governance workshop in Nairobi in March 2026, signalling its intent to be a rule-maker, not just a rule-taker, in the emerging global AI order.

And then there is MTN’s move, significant precisely because it did not come from government. It came from a commercial actor making a strategic bet that African infrastructure, owned and operated on African terms, is the better long-term position. When private capital starts to reach the same conclusion as policy, the middle ground stops being a theory and starts being a market.

None of this is enough yet. But it points toward the right architecture: not building walls, but building choices.

The Deeper Question

The Sovereign AI Seminar ended with a question from the floor about energy, which the panellists called the elephant in the room. Talent, regulation, compute: all of these can be addressed at the margin. Energy is the hard constraint, particularly for a continent where data centre ambitions consistently run ahead of grid capacity. The Kenya-Microsoft stall was, at root, an energy story as much as a sovereignty one.

That observation cuts both ways for Africa. It is a real limitation. But it is also an opportunity. The continent holds some of the world’s largest renewable energy potential, and the countries that invest in energy infrastructure on their own terms are not just solving a constraint. They are building the foundation that makes everything else possible.

The middle ground in the end is not a compromise position. It is a deliberate one. Know where you have leverage and build there first. Defend it across political cycles, not just within them. And resist the pull of the Babel model, not because scale is wrong, but because borrowed scale, on someone else’s terms, is the opposite of sovereignty. The countries that solve the energy-compute equation on their own infrastructure will not just achieve AI sovereignty, they will have something the world will want to buy.

A Final Thought

The conversation at Oxford was largely about the UK as a single entity navigating a single question. But Africa is not one country. It is 54, each with a different resource base, regulatory maturity, energy profile, and relationship with technology. The most interesting version of the African AI sovereignty conversation is not a continental one. It is a country-by-country one, where each nation asks a specific question: where in the AI stack do we actually have leverage?

South Africa, with the continent’s largest installed energy capacity and its most mature enterprise technology ecosystem, is already attracting global cloud infrastructure investment, with Microsoft, Amazon, and Alibaba operating cloud services there, and MTN selecting it as one of two priority markets for its AI data centre build-out. Rwanda’s governance institutions and regulatory track record position it to lead on data standards and AI policy frameworks. Nigeria’s population size, language diversity, and consumer market give it significant leverage at the data and applications layer. Other countries will find their own positions, whether in renewable energy, talent development, research, digital trade, or specialised datasets.

The mistake would be to assume sovereignty requires every country to do everything. No nation, not even the United States, controls every layer of the AI stack. Sovereignty is not built through self-sufficiency. It is built through leverage.

For Africa, that means identifying where each country has durable advantages, investing deliberately at those layers, and connecting those strengths into something that functions as a continental ecosystem. The question is not whether Africa can replicate Silicon Valley or Shenzhen. It is whether African countries can build the pieces of the future that the rest of the world cannot easily replace.

That is the conversation worth having. Not how Africa catches up, but where Africa leads.