Research, white paper, EP-2409
Sovereignty Without Silos: Verifiable Compute and Africa's Data Localisation Debate
Eschon Research Unit,
Abstract
Forty-four African countries now have data protection laws, but a law on paper is not the same as a workable rule for where data can be processed. Kenya, Nigeria and Rwanda have chosen three different answers to the same question. At the same time, Africa holds roughly 0.6% of the world's installed data centre capacity, so a policy of "keep everything in-country" runs into a hard physical limit.
This paper reviews the legal position as of September 2026, then asks a practical question: when a regulator wants to know where data was processed, by whom, and under what code, what can a technical system actually prove? We assess three families of tools (federated learning, confidential computing, and zero-knowledge proofs) against the evidence, including two hardware attacks published this year that weaken the case for relying on any single one. We conclude with a layered evidence model that treats sovereignty as something you demonstrate, not something you assert.
This is a research paper, not legal advice. Every factual claim below is tied to a source listed at the end, and Section 7 lists what we could not verify.
1. Where the law stands
The continental picture has moved quickly, though unevenly.
- National laws. As of the end of 2025, 44 African countries had data protection laws, covering 80% of African Union member states. Thirty-eight had fully operational data protection authorities, and eleven member states had no law at all. Enforcement is becoming real: the same review reports more than 110 decisions from Kenya's regulator in 2025, and a fine of ₦766 million issued by Nigeria's regulator against a digital satellite provider.
- The Malabo Convention. The AU Convention on Cyber Security and Personal Data Protection entered into force on 8 June 2023, when the fifteenth state ratified. It is the only binding regional data protection treaty outside Europe, but fewer than a third of African states have ratified it, according to Chatham House.
- The AU Data Policy Framework. Adopted in 2022, it asks member states to develop a cross-border data flows mechanism that reflects differing levels of digital readiness. In December 2025 the AU Commission convened a validation workshop in Addis Ababa on three continental frameworks: data categorisation and sharing, cross-border data flows, and open data. The stated goal is a secured Digital Single Market by 2030.
- Trade and AI. The AfCFTA Digital Trade Protocol has been adopted, including an annex on cross-border data transfers. The AU's Continental AI Strategy, endorsed in July 2024, lists data sovereignty among its themes, with a first phase (2025 to 2026) focused on governance structures and national strategies.
The direction is toward harmonisation. The present reality is that each country still decides for itself, and the Future of Privacy Forum's review of five countries (Kenya, Nigeria, South Africa, Rwanda and Côte d'Ivoire) concluded that "the practical implementation of transfer tools remains uneven."
2. Three national answers
| Kenya | Nigeria | Rwanda | |
|---|---|---|---|
| Instrument | Data Protection Act 2019; Data Protection (General) Regulations 2021 | Data Protection Act 2023; General Application and Implementation Directive (GAID) 2025 | Law No. 058/2021 |
| Core approach | Transfers need safeguards; a defined set of state-interest processing must stay local | Transfers allowed via adequacy or approved instruments; no adequacy list yet | Personal data stored in Rwanda unless the regulator authorises storage abroad |
| Local processing rule | Yes, for six named purposes (below) | No general localisation rule found in the sources we reviewed | Yes, by default (Article 50) |
| Cloud abroad | Permitted with safeguards; sensitive data may need Commissioner approval | Permitted through the mechanisms above | Permitted with proper safeguards |
Kenya. The most precise rule is Regulation 26 of the 2021 Regulations, which we read in the official text. A controller or processor handling personal data for the strategic interest of the state must either "(a) process such personal data through a server and data centre located in Kenya; or (b) store at least one serving copy of the concerned personal data in a data centre located in Kenya." The listed purposes are civil registration and legal identity management; elections; public finance systems; protected computer systems under the Computer Misuse and Cybercrimes Act; early childhood and basic education; and primary or secondary health care. A Cabinet Secretary can also require a controller processing outside Kenya to comply.
Our reading is that Regulation 26 is, in the first instance, a copy requirement rather than a compute ban. Option (b) allows processing elsewhere so long as a serving copy sits in a Kenyan data centre. That distinction matters for the rest of this paper. Kenya's regulator had, as of August 2025, issued no adequacy decisions for any country, territory or sector, and Kenya's 2024 Cloud Policy is described by the Carnegie Endowment as requiring sensitive data to be stored in accredited local infrastructure.
Nigeria. The NDPC adopted the GAID on 20 March 2025, and it took effect on 19 September 2025 after a six-month transition. It sets out the factors the Commission weighs for adequacy but does not list approved countries. Instruments such as binding corporate rules and standard contractual clauses require the Commission's approval, and the transfer obligations apply to controllers and processors of major importance. Some commentary describes standard contractual clauses as immediately usable; we follow the more cautious reading, and readers should check the current position.
Rwanda. Article 50 of Law No. 058/2021 makes Rwanda the strictest of the three by default: storing personal data outside the country requires authorisation from the National Cyber Security Authority, which the regulator offers as a formal service.
The pattern across all three is that the law asks a question about the world (where is this data, who touched it, under which safeguards) and expects an answer that someone can check.
3. The capacity problem
If the answer to every localisation rule were "host it here," the continent would need somewhere to host it. The Africa Data Centres Association's 2026 economic report puts operational capacity at 360 MW, with 238 MW under construction and 656 MW planned, and says Africa represents about 0.6% of global installed capacity. It calls energy reliability "the central operational variable" and notes that in some markets, such as Nigeria, energy can be a significant share of operating cost.
Other market trackers give different totals (one puts current capacity above 400 MW), which reflects differences in what is counted and when. We use the ADCA figures because they come from the industry body and are the most conservative. The direction is the same in every source: strong growth from a very small base, concentrated in a few markets.
Our reading of these numbers is that blanket localisation is cheap to legislate and expensive to satisfy. A rule can be passed in a session; 656 MW of planned capacity takes years to build and power. Until then, the burden of compliance is likely to fall on whoever has the least room to absorb it.
4. Why sovereignty is turning into leverage
Data protection law is also starting to work as a negotiating position. Writing for Carnegie in June 2026, Jane Munga argues that African data protection laws have moved "from shield to lever": originally written to protect individuals, they now let governments question deals that treat health data as a freely extractable resource. The 2025 policy roundup records that a Kenyan court halted a bilateral health data framework with the United States. (The two sources quote different dollar values for the agreements involved, so we do not repeat either.)
Whatever one thinks of any single dispute, the implication for builders is clear. Institutions negotiating over data will increasingly need to show verifiable custody: what was done with the data, by which code, and where. Trust in a vendor's assurances will carry less weight than evidence.
5. What can technology actually prove?
Three families of tools are usually offered as answers. Each proves something different, and each has a limit that matters in African deployment conditions.
5.1 Federated learning: data stays put
In federated learning, a model travels to the data and only updates come back. This is the tool that best fits the legal problem, because raw records never leave the institution that holds them.
The African evidence is encouraging but early. A 2025 study of privacy-preserving learning for cross-border telemedicine spread 3,459 records across seven hospital nodes in Kenya, Tanzania and Uganda. Federated training improved accuracy by 0.0665 over the centralised baseline, and adding differential privacy (ε = 0.30) reduced the success rate of a model-inversion attack from 0.696 (centralised) to 0.638. The authors are explicit that the data was synthetic, generated with Synthea, and that no clinical validation was done. It shows the mechanism works; it does not show the numbers will hold on real patients.
A tuberculosis screening study spanning hospitals and research centres in eight African countries is a better guide to the real obstacles. Its authors report poor infrastructure, unreliable internet, limited digital literacy and weak AI regulation, and note that "some institutions were also reluctant to share model updates due to data control concerns." That last point is important: even sharing gradients can feel like a loss of control, so federated learning needs a way to show institutions what leaves their walls.
What it shows a regulator: raw data did not move. What it does not show: that the aggregator behaved correctly, or that updates cannot leak information.
5.2 Confidential computing: the hardware vouches for the code
Trusted execution environments (TEEs) such as Intel TDX and AMD SEV-SNP run code in encrypted memory and can produce an attestation, a signed statement of which code is running on which hardware. In principle that lets a regulator verify the processing environment without trusting the host.
Two results published this year show why "in principle" needs care.
- TEE.fail, presented at IEEE S&P 2026, extracts keys from Intel TDX and AMD SEV-SNP systems using a DDR5 memory-bus interposer built from off-the-shelf parts for under US$1,000, and shows how extracted attestation keys can be used to compromise Nvidia GPU confidential computing. The researchers report that Intel and AMD both exclude physical attacks from their threat models.
- DDRop, reported on 14 September 2026 ahead of presentation at ACM CCS in November, comes from researchers at KU Leuven, ETH Zurich, Durham University and Google. According to that report, it uses an interposer costing under US$200 to silently drop memory writes, can produce forged attestations, and affects Intel TDX, Intel Scalable SGX and AMD SEV-SNP. Intel and AMD were informed in advance and, per the report, offered no mitigation guidance. We have read the secondary reporting only and will revisit this when the paper is public.
The vendors' position is defensible: these attacks require physical access. But physical access is exactly the scenario a sovereignty argument has to cover, since the question is whether a foreign or third-party host can be trusted with the machine. A TEE attestation is therefore useful evidence but not proof of custody.
What it shows a regulator: the intended code was running, if the hardware is honest. What it does not show: that the hardware was not tampered with.
5.3 Zero-knowledge proofs: mathematics instead of trust
Zero-knowledge proofs let a party prove a computation was done correctly without revealing the inputs. They remove the trust-in-hardware problem entirely, which makes them the strongest of the three in theory.
They are also the least mature. A survey of the field, accepted by Artificial Intelligence Review in 2026 and covering research from 2017 to August 2025, identifies "limited circuit expressiveness," "high proving cost" and "deployment complexity" as the main obstacles. In practice this means proofs are feasible for small models and specific checks, and not yet for routinely proving the inference of a large language model.
What it shows a regulator: a specific computation was performed correctly. What it does not yet do: scale to the models institutions want to use.
5.4 Summary
| Technique | What it can show | Main limit today | Maturity |
|---|---|---|---|
| Federated learning | Raw data never left the source | Trust in the aggregator; update leakage; weak connectivity | Deployed in pilots |
| Confidential computing | Approved code ran on attested hardware | Physical-access attacks can forge attestation | Widely available, recently weakened |
| Zero-knowledge proofs | A stated computation was correct | Proving cost limits model size | Research to early production |
No single row is enough. The useful observation is that their weaknesses do not overlap.
6. A layered evidence model
Our proposal follows from the table: prove the process, not just the location, and stack evidence so that no single failure removes the whole case.
- Keep raw data at the source. Use federated or other split-computation designs so records stay with the institution that holds them. This addresses the most common legal exposure and the control concern raised in the eight-country study.
- Make the local copy do the legal work. Where a rule such as Kenya's Regulation 26 turns on an in-country serving copy, design for that directly: an in-country copy under local custody, with processing wherever is practical. Treat the copy and the compute as separate questions.
- Attest the code, and assume the host may be hostile. Use TEE attestation as one signal among several. Pair it with physical custody controls, tamper monitoring, and independent verification that the attestation keys have not been compromised. Given TEE.fail and DDRop, do not let attestation be the only control.
- Prove the small things cryptographically. Use zero-knowledge proofs where they are practical today: bounds on model updates, correctness of aggregation steps, and membership in an approved set. Eschon's own BFT-FedAvg++ work follows this approach, using range proofs on updates rather than trying to prove a whole model.
- Publish an audit trail a regulator can verify. Append-only, signed logs of what code ran, on what class of data, under which policy, without exposing the data itself. This is the layer that turns technical properties into something a data protection authority can actually review.
Each layer answers a question a regulator might ask. Where is the data? (Layers 1 and 2.) What code touched it? (Layers 3 and 5.) Was the aggregation honest? (Layer 4.) Can we check that later? (Layer 5.)
7. What we could not verify
Sovereignty claims are easy to overstate, so we list the limits of this paper.
- Legal interpretation. Our reading of Kenya's Regulation 26 as a copy requirement rests on the official text but has not been tested with the Office of the Data Protection Commissioner. Nigeria's treatment of standard contractual clauses is described differently by different commentators. Please consult counsel before relying on either.
- AU frameworks. The three continental data frameworks were validated in December 2025. We found no confirmation of formal adoption at the time of writing.
- Malabo ratifications. Published lists of ratifying states disagree with one another, so we cite only the entry-into-force date and the "fewer than a third" statement, and give no country list.
- Federated learning evidence. The East Africa result uses synthetic data. We are not aware of a published multi-country African study that combines federated learning, differential privacy and real clinical data at scale.
- DDRop. We relied on secondary reporting of an unpublished paper.
- Capacity figures. Trackers disagree; we used the industry association's numbers.
- The layered model itself. It is a design proposal. We have not yet tested it against a regulator's review, and we do not claim it has been accepted by any authority.
8. Recommendations
For regulators. Publish the evidence you would accept, not only the rule. A short list of acceptable attestation and audit artefacts would let institutions comply without guessing. Where practical, distinguish "must be stored here" from "must be computed here," as Kenya's Regulation 26 already does.
For institutions holding sensitive data. Ask vendors for evidence, not assurances: which code ran, on what hardware, under whose physical control, with what audit trail. Treat any single-technology answer, especially a TEE-only one, as incomplete.
For builders. Design for weak connectivity and small data centres from the start, and make what leaves the institution inspectable. The federated learning studies suggest that showing people what is shared matters as much as the privacy guarantee itself.
For funders and policymakers. Capacity and energy are the binding constraint. Localisation rules without matching investment in reliable, in-country compute will push institutions toward exemptions, or toward silence.
9. Conclusion
The debate is often framed as localisation against openness. The evidence points to a third option: sovereignty as a property you can demonstrate. Laws across the continent already ask where data is and who handled it. What is missing is a shared, checkable way to answer. The tools to build one exist, but none is sufficient alone, and this year's hardware research is a reminder that trust in any single layer is a liability. Layering imperfect evidence, and publishing it, is the most credible path we can see.
Sources
Law and policy
- Digital Policy Alert. The Year of the Teeth: Data Protection in Africa Roundup, 2025, Projections for 2026. 11 January 2026. digitalpolicyalert.org
- African Union. AU Commences Validation of Data Governance Frameworks to Accelerate Digital Single Market by 2030. December 2025. au.int
- African Union. Continental Artificial Intelligence Strategy. July 2024. au.int
- Future of Privacy Forum. The African Union's Data Policy Framework. fpf.org
- Future of Privacy Forum. Cross-Border Data Flows in Africa. 6 June 2025. fpf.org
- Chatham House. The AU can help African countries adopt the UN cybercrime convention. But the challenges are significant. October 2025. chathamhouse.org
- Malabo Convention (entry into force). Wikipedia
- Republic of Kenya. The Data Protection (General) Regulations, 2021, Legal Notice No. 263, Regulation 26. odpc.go.ke
- Data Governance Africa. Cross-Border Data Transfers (Kenya). 8 August 2025. datagovernance.africa
- DLA Piper. Nigeria: NDPC issues GAID, key compliance insights. June 2025. privacymatters.dlapiper.com
- Rwanda Data Protection and Privacy Office. Cloud Storage Under Rwanda's Data Protection Law. dpo.gov.rw
- Munga, J. Locked In: How African Data Protection Laws Move from Shield to Lever. Carnegie Endowment for International Peace, via Lawfare, 1 June 2026. lawfaremedia.org
Infrastructure
- Africa Data Centres Association. Data Centres in Africa 2026: The Economic Report, as reported by Zawya. 2026. zawya.com
Technical
- Privacy-preserving data governance in cross-border telemedicine using federated learning and differential privacy in Kenya. Internet of Things and Cloud Computing, 13(3), 2025. doi:10.11648/j.iotcc.20251303.12
- Fabila, J., Garrucho, L., Campello, V. M., Martín-Isla, C., Lekadir, K. Federated learning in low-resource settings: a chest imaging study in Africa. arXiv:2505.14217
- Peng, Z. et al. A Survey of Zero-Knowledge Proof Based Verifiable Machine Learning. Artificial Intelligence Review, 2026. arXiv:2502.18535
- TEE.fail: Breaking Trusted Execution Environments via DDR5 Memory Bus Interposition. IEEE S&P 2026. tee.fail
- New DDRop Attack Breaks Intel TDX and AMD SEV-SNP Confidential Computing. The Hacker News, 14 September 2026. thehackernews.com
- Eschon Research Unit. BFT-FedAvg++: Byzantine-Resilient Federated Aggregation. March 2026. /research/bft-fedavg-plus-plus
