Hans van Linschoten: Sovereignty needs an operator

Africa’s digital sovereignty debate has matured. The question is no longer whether critical sectors such as healthcare should retain greater control over their data and AI systems, but who is actually capable of delivering that control in practice.

A recent Journal of Global Health Economics and Policy paper titled Cloud, control and diagnostic sovereignty: the political economy of AI-enabled health diagnostics in Africa, argues convincingly that data localisation alone is insufficient and that genuine diagnostic sovereignty requires the ability to audit, inspect and govern AI systems within national jurisdictions. It is an important contribution to the discussion. Yet it leaves one critical question largely unanswered: who builds and operates the infrastructure that makes those policy ambitions possible?

In this response, Hans van Linschoten argues that sovereignty is ultimately an operational challenge rather than simply a regulatory one. Policies can require local control, but without locally owned cloud operators capable of running secure, auditable AI platforms, sovereignty remains an aspiration rather than a capability. Drawing on the AfriQloud model, he explores what can realistically be delivered locally, where hyperscale cloud providers will continue to play an important role, and why Africa’s path to digital independence will almost certainly be hybrid rather than ideological.

What the paper found

In May this year, Carl Adams Kopati Gbali, Okechukwu Eze and Francisca Arbor, published a multi-case study of four African health technology firms – Helium Health, Ubenwa, Neural Labs Africa and Envisionit Deep AI – across Nigeria, Kenya and South Africa. The method is a desk study of thirty-six public documents, and the authors are careful to say so on almost every page: these are publicly observable patterns, not audited infrastructure records.

They identify three. Infrastructural lock-in: all four firms are publicly associated with AWS or Google Cloud for storage, training, inference and model updates. Data sovereignty deficits: no public evidence of locally governed training datasets or systematic local retraining. Governance asymmetry: the site of clinical impact and the site of technical control are in different jurisdictions.

The third finding is the one that should change how this debate is conducted. Diagnostic sovereignty, the authors argue, is not reducible to data ownership or data localisation. It requires the practical capacity to inspect datasets, test model performance on local populations, audit update histories and compel corrective action when a model underperforms. A copy of the data sitting in Lagos is not the same thing as the ability to audit the model that reads it.

What the paper does not say

The policy table at the heart of the article proposes five instruments:

• Tiered data sovereignty frameworks
• Regional health data trusts
• Staged validation sandboxes
• Public audit infrastructure
• Procurement rules

Every one is addressed to ministries, regulators and African Union institutions. Every one is a demand-side instrument. Not one of them names who supplies the capability that makes the demand enforceable.

This is the gap. A regulator can legislate a right to inspect model weights and update logs. It cannot exercise that right unless there exists, inside its jurisdiction, a platform that holds those artefacts and is operated by an entity it can license, inspect and – if necessary – sue. Absent that, “audit infrastructure” means a contractual clause against a company headquartered on another continent, enforced through a legal process no health ministry in the region has the resources to pursue.

We have a direct commercial interest in the answer, which is stated in full at the end of this piece. It is stated up front too, because the argument that follows should be read with it in view.

The sovereign add-on

The framing that survives contact with operational reality is not “replace the hyperscalers.” It is narrower and more answerable: what is the smallest set of functions that must sit under local jurisdiction for a regulator to do its job – and can that set be operated locally at acceptable quality and cost?

The paper’s own tiering answers the first half. Raw identifiable clinical data, governed training datasets, model weights, audit logs and update histories belong in the sovereign tier. De-identified analytics, burst training capacity, generic developer tooling and non-clinical workloads do not. That is a substantially smaller footprint than a hyperscaler region: a governed data store, steady-state inference capacity, and an audit plane.

This is what a sovereign add-on means in practice. Not a national cloud programme and not an import substitution project, but a node inside the jurisdiction, run by a locally licensed operator, carrying the regulated tier and interconnecting with global capacity for everything else.

AfriQloud is built to roughly that specification: country operating companies under local majority ownership, with a minimum viable server node rather than a region-scale build as the unit of deployment. Per-node capital cost in our modelling is in the order of $339,000 or even less. The point of the small unit is that a teaching hospital group, a national insurer or a private radiology network can anchor one without a sovereign wealth fund behind it. That is a deliberate contrast with hyperscale economics, which require market density most African countries do not yet have and will not have for some years.

Can the same quality be realised locally?

Not across the board – and any vendor who tells a health ministry otherwise should be kept away from clinical workloads.

Where the gap is real:

• Managed service breadth. A hyperscaler ships hundreds of managed services. A sovereign node ships compute, storage, orchestration and inference. A team that built its pipeline on managed ML tooling faces genuine re-platforming cost, and that cost falls on the innovator, not the regulator demanding the move.
• Accelerator supply. GPUs are allocated globally, and priority follows volume. A regional operator queues behind buyers an order of magnitude larger. This constrains retraining cadence more than it constrains inference, but it constrains it.
• Failure domains. Multi-region redundancy is hard to replicate at small scale, and grid stability and single-path transit remain live constraints in several markets. There is a serious argument that in-hospital inference is more robust than a transcontinental dependency – but it has to be made with published uptime data, not asserted.
• Compliance attestation. Hyperscalers hold deep portfolios of audited certifications. A local operator starts near zero and has to earn ISO 27001 and health-sector equivalents before a serious hospital group will sign anything.
• Engineering depth. Site reliability talent, not hardware, is the binding constraint on most local infrastructure builds, and it is the hardest one to fix on a project timeline.

  Where parity is achievable, and is what actually matters:

  • Data residency plus inference latency and availability for a defined clinical workload. This is a bounded engineering problem, not a platform-breadth problem.
  • Regulator-inspectable audit logs and model update histories. This is arguably easier locally, because the operator holding them is licensable in the jurisdiction rather than reachable only by contract.
  • Local retraining on locally governed datasets – the very practice the paper found unevidenced across all four cases.

  So the right question is not whether a sovereign node matches AWS or GCP. It is whether it is good enough at the specific things clinical governance requires, at a cost a health system can actually carry. Everywhere else, parity is not worth paying for.

  What still has to be proven

  We should hold ourselves to the standard the authors of the article set for themselves. To our knowledge there is no published, peer-reviewed evidence of clinical-grade AI diagnostics running at scale on African-owned sovereign infrastructure with documented regulator audit access. The architecture exists and the economics are modelled. The clinical and operational evidence does not yet exist, and pretending otherwise would repeat exactly the transparency problem the paper identifies in the four firms it examined.

  What would settle it is a controlled deployment: one imaging or triage workload, one jurisdiction, a hybrid architecture with the regulated tier on a local node, and published measurement – uptime, inference latency, documented regulator access to the model update history, and clinical performance benchmarked against the same tool running on hyperscale infrastructure. Twelve to eighteen months. It should be designed and reported by people with no commercial interest in the result (which explicitly includes us).

  Accountable interdependence

  We borrow the authors’ own phrase because it is the correct target. The realistic path is hybrid: global capacity for what genuinely benefits from scale, local infrastructure for what needs jurisdiction. And the lever is the fifth instrument in their table, procurement. Ministries and donor-funded programmes that attach local hosting of the regulated tier, local clinical validation and regulator audit access to their tenders create the demand that makes local operators viable at all.

  Without that anchor demand, the capability does not get built, the evidence never gets generated, and a well-argued set of policy recommendations stays on paper while the diagnostic pipeline continues to run somewhere else.

  Africa simply can’t copy the Western (American model). Even Europe with its resources, has a difficult time catching up. The path forward on the continent is distributed capacity and more work to be done on smaller, (highly) specific models and probably different approaches all together. If we don’t start local development on local infrastructure, the sovereignty gap will only widen and digital indpendence will become a fairytale.

  Hans van Linschoten is the CEO of whitesky.cloud, a European sovereign cloud platform, and Co-founder of AfriQloud, a pan-African sovereign cloud and edge AI infrastructure initiative.