By Bora Varliyagci, CEO, Digitalthings

On 17 February 2026, the African Union Commission signed a Memorandum of Understanding at its Addis Ababa headquarters. The press release headline: “African Union Commission and Google Sign Landmark Partnership to Advance Africa’s Sovereign AI and Digital Capacity.”

The continental body’s flagship sovereign AI partnership is with Google. Google’s regional director for Sub-Saharan Africa described it as moving the continent “from digital access to digital agency.”

This is not a criticism of the AU or of Google. The MoU commits to real things — infrastructure, skills, centres of excellence, local-language support including Amharic, training for three million students and teachers by 2030. It is also, like most MoUs, non-binding, which is its own comment on how much of the sovereignty agenda currently exists on paper.

But it captures the tension African policymakers, investors and operators now have to resolve. Sovereignty has become the organising language of digital infrastructure policy across the continent. The infrastructure that language depends on remains overwhelmingly foreign-owned, foreign-operated and foreign-financed.

The question is not whether that tension exists. It plainly does. The question is which parts of the sovereignty agenda are worth paying for now, which can wait, and which are expensive symbolism.

Four definitions, none of them yours

The most significant development of 2026 is not any national policy. It is that sovereignty stopped being a unilateral posture and started organising into competing instruments — which do not agree on what the word means.

14 February, Munich. Canada and Germany launched a “Sovereign Technology Alliance,” framed as cooperation “among trusted partners” to “reduce strategic technology dependencies.”

20 February, New Delhi. At the India AI Impact Summit, OSTP Director Michael Kratsios set out the American position: “Real AI sovereignty means owning and using best-in-class technology for the benefit of your people, and charting your national destiny.” He urged strategic autonomy rather than full self-sufficiency, and rejected global governance of AI outright.

The vehicle is the American AI Exports Program — executive order July 2025, operating since October. New in February was the support package: a Commerce National Champions Initiative folding partner-country AI firms into customised American stacks, a Tech Prosperity Corps under the Peace Corps, and the part African finance ministries will notice — a Treasury-backed fund at the World Bank alongside new financing at the DFC and EXIM.

Read the condition. That financing helps developing partners overcome obstacles “as they import the American AI stack,” and Peace Corps placements go to countries procuring through the programme. This is tied export finance, not development finance. The World Bank fund was announced with no disclosed size, capitalisation or board process, and no African commitment against it that I can find. It may become significant. It is not yet money. On the same trip, Washington signed the New Delhi Declaration alongside China and roughly 89 other countries, having declined the equivalent at Paris in 2025 — the rejection of global governance is real, but narrower in practice than in rhetoric.

10 June, Singapore. The national standards bodies of Korea, Singapore, the UK, Australia and Canada signed a pact on AI pre-standardisation — the groundwork defining testing methodologies and conformity assessment before regulators codify them. Not a treaty. Arguably more consequential than one. Whoever writes the test writes the rule.

16 July, Shanghai. Twenty-nine countries established the World AI Cooperation Organization, with the UN Secretary-General in attendance. Founding members include Russia, Belarus, Cuba, Brazil and Venezuela, twelve Asian states, and ten African ones: Algeria, Cameroon, Republic of Congo, Ethiopia, Kenya, Lesotho, Mozambique, Senegal, South Africa and Zambia. The United States is not a member. Ten days earlier the UN had convened the first session of its own Global Dialogue on AI Governance in Geneva — the body Kratsios named directly when rejecting global governance in February.

Four instruments in five months, barely overlapping, and not the same kind of thing — two multilateral bodies, one standards agreement, one export programme with a financing wrapper. Each carries a definition of sovereignty that suits the party proposing it. None means self-sufficiency, and the American one says so explicitly.

South Africa and Kenya, the continent’s two most developed digital infrastructure markets, have joined a body headquartered in Shanghai. It is worth asking how deliberately — and noting that the alternative to Western dependence is not independence. Huawei remains one of Africa’s largest infrastructure providers; G42 is Emirati. Shenzhen and Abu Dhabi are different counterparties with different risk profiles, not an escape from counterparty risk.

One uncomfortable point of agreement: Kratsios’s advice — strategic autonomy, not self-sufficiency — is substantively where this analysis lands on the evidence. That does not make it disinterested; it arrives attached to a sales channel. It does mean the maximalist position circulating domestically is one no major power recommends. China is the only party seriously attempting full-stack self-sufficiency, from an accumulated base no African market has. The more transferable model is the summit’s host: India’s application-layer-first approach, twenty-two languages, built on existing digital public infrastructure rather than frontier training — which drew far less attention than the American announcements it shared a stage with.

What the capacity gap actually looks like

Any honest sovereignty discussion starts with the physical position.

Africa holds less than 1% of global data centre capacity while accounting for roughly 18% of the world’s population. Beyond that, the datasets diverge and it is worth being straight about it: a July 2026 Fortren & Company report puts operational capacity above 500 MW with 890 MW in the pipeline, while the Africa Data Centres Association’s 2026 Economic Report puts active capacity at 360 MW, with 238 MW under construction and 656 MW planned, and Africa’s global share at 0.6%. Both agree South Africa accounts for the clear majority of what is operational.

Either way, the order of magnitude is the point. France alone ended 2024 with roughly 714 MW installed. The entire African continent, on the most generous dataset, has less — serving eighteen times the population.

This is what most sovereignty rhetoric skates past. You cannot mandate the repatriation of workloads to infrastructure that does not exist. Nigeria’s own regulator says so plainly: NITDA Director-General Kashifu Inuwa puts more than 85% of Nigerian workloads on public cloud, with most of the supporting infrastructure sitting outside the country. Nigeria hosts just 22% of its thousand most-accessed websites locally — below the Sub-Saharan African average of 34%.

The distribution matters as much as the total. Roughly half of Sub-Saharan Africa’s data centres sit in three markets — South Africa, Nigeria and Kenya — and South Africa alone holds the clear majority of operational megawatts. For the other forty-odd countries, “data residency” and “in-country processing” are not the same requirement, and conflating them writes mandates that cannot be met. A residency rule in a market with no compliant facility is not a sovereignty policy. It is an import ban on cloud services, and it will be priced as one.

What the frontier race actually tells us

The standard argument against African AI ambition is capital scale, and the headline numbers support it. Stanford’s AI Index recorded US private AI investment of $109.1 billion in 2024 against China’s $9.3 billion, and 40 notable models from US institutions against 15 from China and three from Europe. Stanford’s 2026 edition puts the 2025 gap wider still — $285.9bn against $12.4bn.

Yet on capability, the gap has all but closed. The 2025 Index showed the US–China gap on MMLU, MATH and HumanEval collapsing from 17.5, 24.3 and 31.6 percentage points at end-2023 to 0.3, 1.6 and 3.7 by end-2024. The 2026 edition puts the leading US model just 2.7% ahead on Arena Elo.

Two honest caveats before anyone builds a strategy on this. First, Stanford itself warns that the Chinese investment figure understates actual spending, because state guidance funds and state-initiated vehicles do not appear in private-capital databases — the report says explicitly that the ratio “may be less dramatic than it appears.” Second, China closed that gap holding roughly 70% of global generative AI patent families (WIPO), frontier-scale domestic labs, a semiconductor programme and an electricity system with reserve margin to spare. Those are capital-formed assets accumulated over decades.

So the correct reading is narrower than the optimistic one, and less comfortable: accumulated stock in talent, research and energy matters more than current-year investment flows. That is not an argument that Africa can enter the frontier race cheaply. It is an argument about where the returns actually are.

Algorithmic efficiency and open weights have lowered the cost of building on top of the frontier far faster than they have lowered the cost of reaching it. The strongest returns on African AI spending therefore sit at the application, fine-tuning, data and inference layer — not the training layer. That is a defensible strategic conclusion. “Africa can compete on models” is not.

Where sovereignty is real

Regulated data categories. Nigeria’s Central Bank issued Circular PSS/DIR/PUB/CIR/001/004 on 15 June 2026, requiring payment transaction data generated in Nigeria to be stored and managed in Nigeria, with full compliance by 1 January 2027. Narrow, specific, defensible. Payments data is systemically sensitive, the supervisory case is clear, and the scope is bounded enough to implement.

Operational control, not ownership. The distinction that matters most is the one most often collapsed. Reporting indicates that in some North African markets, governments built data centres domestically and then outsourced their management to third-party providers — reproducing vendor lock-in inside a nationally-owned building. Owning the box is not control. Controlling the operating model, the exit rights and the data access terms is.

Procurement terms — the cheapest lever, and an untested one. Exit rights, data portability, audit access, subprocessor disclosure and price escalation caps require no capital expenditure. Most African public-sector cloud contracts contain none of them, and inserting them costs nothing but drafting.

The same scepticism owed to collective bargaining below is owed here. Cheap to draft is not the same as achievable. No African state has yet held hard procurement terms against a hyperscaler at scale, and the reason is structural: leverage comes from credible substitutability, and where local capacity does not exist, the threat to walk is not credible. Nigeria is testing the adjacent question — NITDA has told global providers that waivers are ending and it wants a localisation roadmap.

The precedent worth borrowing is the EU Data Act’s mandatory cloud-switching and egress-fee provisions — this exact lever, legislated and enforced. The two clauses to fight for first are exit rights and data portability: they convert ownership into control, and a contract that cannot be exited is not a sovereign arrangement regardless of where the servers sit.

So: the best risk-adjusted opening move — cheap to attempt, meaningful upside, no capital downside — and the one most likely to go unattempted, because the binding constraint is the same thin drafting capacity discussed below. Not the easiest win. The cheapest to try. Watch NITDA’s October 2026 certification platform and the CBN’s 1 January 2027 deadline for the first real evidence either way.

Coordination — as an aspiration, not an achievement. The top three cloud providers command roughly 63% of global cloud infrastructure spend; with Oracle, Alibaba and Huawei, over 70%. A single African state negotiating against that concentration has little leverage.

I want to be careful here, because the evidence cuts against the optimistic reading. Under the US America First Global Health Strategy, Ghana, Zambia and Zimbabwe declined bilateral arrangements carrying health data-surveillance provisions. But 21 African countries signed — including Nigeria, whose five-year MoU runs to 2030 and carries a $2.1 billion US commitment alongside $3.0 billion in Nigerian domestic co-investment. (South Africa has not signed, for reasons unrelated to data.) That is not collective bargaining working. It is precisely the fragmentation-and-isolation dynamic that makes collective bargaining necessary.

The instruments to fix it exist on paper — the AU Data Policy Framework, the Malabo Convention, the AfCFTA Digital Trade Protocol. None has yet been used to negotiate infrastructure terms at scale. Until one is, “collective bargaining” is a recommendation, not a description.

Where it is overplayed

Sovereign frontier model training. No African market currently has the power availability, GPU access, engineering depth or capital stack to train frontier models competitively. Fine-tuning, local-language adaptation and inference are entirely credible. Frontier pre-training is not, and framing national AI strategy around it spends scarce political capital badly.

Sovereignty without a power plan. Nigeria’s data centre estate is concentrated in Lagos — roughly 18 of about 21 operational facilities — against a grid averaging four hours of supply a day and dependent on diesel backup. The national grid collapsed twice in a single week in January 2026. A sovereignty strategy that does not begin with a bankable power solution is not a strategy.

Sovereignty as a free good. It is not. NITDA’s own Director-General concedes local cloud in Nigeria costs more than offshore alternatives. A residency mandate is, in the near term, a tax on domestic banks, fintechs and startups. That may well be worth paying for payments data. It is worth paying deliberately, with the cost booked, rather than assumed away.

Ambition running ahead of institutional capacity. On 26 April 2026, South Africa’s Minister of Communications and Digital Technologies withdrew the 86-page Draft National AI Policy after researchers identified at least six fabricated academic citations in its bibliography — apparently AI-generated and unverified. Two officials were suspended roughly a month later, after further media pressure.

The narrow lesson is the right one: policy-drafting capacity across the continent is thinner than the ambition of the documents being drafted. That is a resourcing problem, and it is fixable. But it should temper expectations about how quickly sophisticated sovereignty regimes can be written, let alone administered.

Europe is the cautionary comparator here, and it is telling that it rarely features in African sovereignty discussions. Gaia-X, the EU sovereign cloud push and national-champion joint ventures like Bleu and S3NS have had far more capital, far deeper institutional capacity and more than five years. Hyperscaler share in Europe has risen over that period. Sovereignty is difficult even when you can afford it.

What sovereignty costs: the Kenya case

The clearest price signal came in 2026.

Microsoft and G42 announced a $1 billion geothermal-powered data centre at Olkaria in Kenya in May 2024 — an initial 100 MW scaling toward 1 GW, and one of the most credible green-compute propositions on the continent. By 2026 it had stalled, for two reasons that are usually reported separately and belong together.

The first is power. President Ruto put it bluntly: to switch on that data centre, “we would need to shut off power for half the country.” A gigawatt of load does not fit Kenya’s grid, and no amount of policy ambition changes that.

The second is demand. Reporting indicates Microsoft and G42 asked the government to commit to purchasing a fixed volume of cloud capacity annually. Kenya declined. The project is not dead — Kenya’s principal secretary for information, John Tanui, says it is “not failed or withdrawn,” and a right-sized first phase that fits Olkaria’s supply is the obvious path back.

Both constraints are real and they compound. A government that refuses to underwrite demand it cannot forecast is exercising exactly the fiscal discipline sovereignty is meant to protect. A grid that cannot carry the load is a physical fact that no negotiation resolves.

That is the actual trade: not sovereignty versus dependency, but sovereignty versus schedule — and only for as long as the power constraint stands.

What this means for investors and operators

Sovereignty has moved from policy commentary into the investment case. Three implications.

Sovereignty requirements are a cost line, not a footnote. Residency mandates, local operator requirements and certification regimes change capex, opex and time-to-revenue. Nigeria’s National Sovereign Cloud Initiative — three regulatory instruments signed by NITDA with Galaxy Backbone on 5 August 2026, with a national certification platform going live from October 2026 — will require cloud service providers, data centre operators, managed service providers, AI infrastructure providers and sovereign compute providers to register and certify. A real gate, with a real date.

Regulatory risk is now bidirectional. The traditional African infrastructure risk was regulatory absence. The emerging risk is regulatory divergence — a Nigerian workload that cannot legally sit in a Kenyan facility, fragmenting exactly the scale economics that make regional data centre investment work. For regional platforms such as Teraco, Africa Data Centres and Raxio, and for DFI financiers such as IFC, BII and Africa50, this is a first-order modelling question, not a compliance footnote.

Anchor-tenancy assumptions need re-testing. Kenya shows governments will not automatically underwrite the demand that sovereign infrastructure business cases assume.

The questions we now put to every data centre and cloud investment case:

• Which specific data categories face residency mandates, and on what compliance timeline?
• Does the operating model deliver control, or only ownership? Where are the exit rights?
• Is the certification pathway clear, and is the regulator resourced to process it?
• Does the demand case rest on a public-sector anchor tenant — and has that commitment actually been secured?
• Can the power solution support the sovereignty commitment, or does the commitment collapse when the grid does?
• If regional workload consolidation is blocked by divergent residency rules, does the case still clear?
• Which bloc’s standards, financing and certification regime is this asset being built for — and what happens to the case if the host government switches?

At Digitalthings we run these questions across data centre due diligence, carve-out feasibility, multi-country TCO modelling and transaction support in African markets. What we consistently find is that sovereignty requirements are the least well-modelled variable in the investment case — treated as policy background when they are direct inputs into capex, licensing timelines, addressable market and exit optionality.

The real question

Africa’s sovereignty debate is framed as a question of ambition. It is better framed as a question of sequencing.

The continent does not need to choose between dependency and autonomy. It needs to work out which sovereignty is worth paying for now, which can wait, and which is being demanded without a plan to fund the power, skills and institutional capacity that would make it real.

And it now needs to do that while four separate instruments compete to define the term, each carrying a definition that suits the party proposing it. Washington offers announced-but-unsized financing tied to procuring the American stack. Shanghai offers a governance seat. Berlin and Ottawa offer a club of the already-capable. Each is worth evaluating. None is neutral, and none is a substitute for a definition Africa writes itself — which is what the AU Data Policy Framework, Malabo and the AfCFTA Digital Trade Protocol were meant to be, and have not yet been used as.

Ten African states signed into a new global AI governance body in July. The infrastructure those commitments will be judged against still runs, overwhelmingly, on somebody else’s estate.

Closing that gap is not a communications exercise. It is a capital allocation problem, and it deserves to be analysed as one.

---

Bora Varliyagci is CEO of Digitalthings, a boutique strategy and investment advisory firm specialising in digital infrastructure, cloud, platform and IoT solutions, broadband networks and services across Africa.

---

Sources

1. African Union Commission, “African Union Commission and Google Sign Landmark Partnership to Advance Africa’s Sovereign AI and Digital Capacity,” 17 February 2026 — https://au.int/en/pressreleases/20260217/auc-google-sign-partnership-advance-africas-sovereign-ai-digital-capacity
2. Government of Canada, “Canada and Germany sign AI joint declaration and launch Sovereign Technology Alliance,” 14 February 2026 — https://www.canada.ca/en/innovation-science-economic-development/news/2026/02/canada-and-germany-sign-ai-joint-declaration-and-launch-sovereign-technology-alliance.html
3. Standards Council of Canada / Korea Times, five-body AI pre-standardisation MoU, Singapore, 10 June 2026 — https://www.koreatimes.co.kr/economy/policy/20260612/5-nations-sign-ai-alliance-pact-to-shape-global-tech-rules
4. Xinhua / Global Times, World AI Cooperation Organization founding agreement, 16 July 2026 — https://english.news.cn/20260716/2d364f232dd7456c8edf1d67d5125d37/c.html 4a. The White House, “U.S. Promotes AI Adoption, Sovereignty, and Exports at India AI Impact Summit,” 20 February 2026 — https://www.whitehouse.gov/releases/2026/02/u-s-promotes-ai-adoption-sovereignty-and-exports-at-india-ai-impact-summit/ 4b. The White House, “Remarks by Director Michael Kratsios at the India AI Impact Summit,” 20 February 2026 — https://www.whitehouse.gov/releases/2026/02/remarks-by-director-michael-kratsios-at-the-india-ai-impact-summit/ 4c. Executive Order 14320, Promoting the Export of the American AI Technology Stack, 23 July 2025 — https://www.whitehouse.gov/presidential-actions/2025/07/promoting-the-export-of-the-american-ai-technology-stack/ 4e. UN Global Dialogue on AI Governance, first session, Geneva, 6–7 July 2026 — https://www.un.org/global-dialogue-ai-governance/ 4f. Press Information Bureau, Government of India, “AI Impact Summit 2026 Concludes with Adoption of New Delhi Declaration” — https://www.pib.gov.in/PressReleasePage.aspx?PRID=2231208 4g. US Peace Corps, Tech Prosperity Corps — https://www.peacecorps.gov/ways-to-serve/tech-corps/ 4d. American AI Exports Program — https://aiexports.gov/
5. Stanford HAI, AI Index Report 2025 — https://hai.stanford.edu/ai-index/2025-ai-index-report
6. Stanford HAI, AI Index Report 2026 — https://hai.stanford.edu/ai-index/2026-ai-index-report
7. Africa Data Centres Association, Data Centres in Africa 2026: The Economic Report — https://africadca.org/wp-content/uploads/2026/02/DataCentresInAfrica2026-EN.pdf
8. Fortren & Company via Nairametrics, “Africa’s data centre capacity exceeds 500 MW with 890 MW in pipeline,” 27 July 2026 — https://nairametrics.com/2026/07/27/africas-data-centre-capacity-exceeds-500-mw-with-890-mw-in-pipeline-report/
9. Central Bank of Nigeria, Circular on Introduction of Market Structure Requirements, Data Localisation, Ultimate Beneficial Ownership Disclosure, and Systemic Oversight Measures in the Nigeria Payments System, PSS/DIR/PUB/CIR/001/004, 15 June 2026 — https://www.cbn.gov.ng/Out/2026/CCD/
10. TechCabal, “Nigeria pushes to keep more of its cloud infrastructure local,” 10 August 2026 — https://techcabal.com/2026/08/10/nigeria-pushes-local-cloud-as-85-of-workloads-use-public-clouds/
11. Nairametrics, “NITDA signs sovereign cloud framework to boost AI, data centre investment,” 5 August 2026 — https://nairametrics.com/2026/08/05/nitda-signs-sovereign-cloud-framework-to-boost-ai-data-centre-investment/
12. Fasken, “The Hallucinatory Irony of it All: Minister Withdraws Draft National AI Policy,” April 2026 — https://www.fasken.com/en/knowledge/2026/04/the-hallucinatory-irony-of-it-all-minister-withdraws-draft-national-ai-policy
13. Tom’s Hardware, “Microsoft’s $1 billion Kenya data center stalls over disagreements on power capacity” — https://www.tomshardware.com/tech-industry/microsofts-1-billion-kenya-data-center-stalls-over-disagreements-on-power-capacity
14. Bloomberg, “Microsoft and G42’s Kenya Data Center Appears at a Standstill,” 11 May 2026 (paywalled) — https://www.bloomberg.com/news/newsletters/2026-05-11/microsoft-and-g42-s-kenya-data-center-appears-at-a-standstill 14a. WIPO, Patent Landscape Report: Generative Artificial Intelligence (2024) — https://www.wipo.int/web-publications/patent-landscape-report-generative-artificial-intelligence-genai/en/index.html
15. African Business, “US aid-for-data deals meet African resistance,” July 2026 — https://african.business/2026/07/technology-information/us-aid-for-data-deals-meet-african-resistance
16. KFF, America First bilateral global health agreements tracker — https://www.kff.org/global-health-policy/kff-tracker-america-first-mou-bilateral-global-health-agreements/
17. Synergy Research via Statista, global cloud infrastructure market share, Q1 2026 — https://www.statista.com/chart/18819/worldwide-market-share-of-leading-cloud-infrastructure-service-providers/
18. New America, “Africa’s Digital Sovereignty Trap: The Data Center Dilemma” — https://www.newamerica.org/insights/africas-digital-sovereignty-trap/
19. Rest of World, “Pushing back from Big Tech: Africa’s hard road to AI sovereignty,” 21 May 2026 — https://restofworld.org/2026/africa-ai-sovereignty-big-tech/
20. TechCabal Insights, “Can Nigeria’s data centres power Africa’s AI future?” — https://insights.techcabal.com/can-nigerias-data-centres-power-africas-ai-future/