We have spent two years debating whether AI belongs in low- and middle-income country (LMIC) healthcare. That debate is over. AI is already running, whether we sanctioned it or not.
- Health workers paste patient symptoms into ChatGPT.
- WhatsApp chatbots triage pregnancy questions.
- Predictive models forecast outbreaks.
The real question is whether we’re paying attention to the parts of this stack that will determine whether AI helps or harms the people we serve.
Most of us live in the messy middle. We see the radiologist shortages and want the diagnostic gain. We see data extraction and want sovereignty protections. Pretending those goals are easy to reconcile, or that one cancels the other, is how the sector keeps producing strategies that age badly.
I moderated a Global Digital Health Network webinar recently on AI ethics in LMIC healthcare with Hannah Cooper Klein. She opened with a thesis that should have ended one conversation and started a different one.
Malawi has roughly one radiologist per 8.8 million people. In a country that is under-resourced, refusing to use AI for image interpretation is not the cautious ethical position. It is the negligent one, and the point keeps getting clearer.
But Hannah did not stop there, and this is where most of our sector’s AI discourse falls apart.
Even if we designed the perfect AI radiology tool for Malawi tomorrow, Malawi cannot run it. The country does not have the compute capacity. Neither do most of its neighbors. While we publish another framework on responsible AI, the physical infrastructure that would let any of those frameworks matter is being built somewhere else, for someone else.
AI is not just a model
Hannah made it clear that AI is not just a model. AI is compute, cloud, chips, data centers, energy, procurement power, cybersecurity, and governance. Most of our sector’s discourse engages with the model layer because that is what gets demoed at conferences.
The seven other layers determine whether the model layer can be used responsibly, sustainably, or at all.
The global AI race is creating a hierarchy of access. Even US academic researchers struggle to get enough compute, which is why the National AI Research Resource Pilot exists: the largest cloud providers, frontier AI labs, and national security actors are absorbing capacity faster than it can be built.
Africa sits at the bottom of that hierarchy.
At AfricaCom 2024, African Telecommunications Union Secretary General John Omo stated plainly that the whole of sub-Saharan Africa has less cloud capacity than Switzerland. Asking African health systems to “adopt AI” in that context is asking them to depend on infrastructure they do not own, cannot easily procure, may not fully trust, and have limited leverage to govern.
Hannah was hopeful, naively, that US-China competition would push both sides to build out African digital infrastructure as a condition of market access. That is not happening. Big tech wants African consumers, not African capacity, and infrastructure investment will not arrive through commercial logic alone.
Look at one of the largest bets in this space.
In January 2026, the Gates Foundation and OpenAI announced Horizon 1000, a $50 million commitment to deploy AI tools across 1,000 primary healthcare clinics in Africa by 2028, starting in Rwanda.
The framing is correct on the workforce shortage, sub-Saharan Africa’s roughly 5.6 million health worker gap. The framing is silent on where the inference runs, who controls the data, and what happens when the pilot ends. A $50 million tools rollout without an infrastructure plan underneath it is a deployment, not a strategy.
Hannah’s answer was the most useful thing said in the hour.
A single small country negotiating with a hyperscaler has almost no leverage. Twenty countries negotiating jointly for public-interest health use cases is a different conversation. The African Union’s Continental AI Strategy, endorsed in July 2024, gestures at this. None of the practitioners I work with believe it is being operationalized at the speed the technology is moving.
We are losing the data sovereignty discussion
The compute conversation cannot be separated from where data lives and who controls it. You cannot ask African ministries to ship health data to foreign clouds without confronting the long history of LMIC data being collected, used, and monetized without meaningful consent.
The Kenya-US Health Cooperation Framework is the live case study.
The five-year, $1.6 billion agreement signed in December 2025 included a Data Sharing Agreement that, whistleblowers and civil society argued, would have given the United States access to Kenya’s national health database under US federal law.
The Consumer Federation of Kenya sued. Kenya’s High Court suspended the data-sharing components pending constitutional review, and nearly 50 African civil society organizations called on heads of state to demand equity and sovereignty in their bilateral US health deals.
Civil society pressure got the agreement amended to specify that Kenyan law prevails. That is a real win. But the underlying dynamic is unchanged: data flows still go north, value creation still happens elsewhere, and the legal protections still depend on courts and ministries with far less capacity than the entities they negotiate with.
As the IMF AI Preparedness Index makes uncomfortably clear, low-income countries score around 0.32 on AI readiness against advanced economies’ 0.68. We are not negotiating from parity.
It also helps to be honest that strong legal regimes are not, by themselves, doing the work we pretend they are. The webinar surfaced what one participant called the “click through problem”.
Faced with a 50-page terms-of-service screen on a phone, almost no one reads it. They click accept, because the alternative is not using the service. The same dynamic plays out when a community health worker reads consent language to a patient at clinic volume.
GDPR, Kenya’s Data Protection Act, and India’s Digital Personal Data Protection Act all exist. None of them protects anyone if the operational layer reduces consent to a tap-through.
What practitioners should do
We spend disproportionate energy on app-level ethics, the bias audit of a chatbot, the hallucination rate of an LLM, Those questions matter. But they are not sufficient. In LMIC healthcare, the deeper ethical questions are increasingly about infrastructure: compute, cloud, cybersecurity, procurement power, data governance, and whether governments can access, evaluate, adapt, sustain these systems.
During our discussion Hannah suggested several practical steps that would shift the conversation from what responsible AI should mean in theory to what governments need in practice: access to compute, trusted cloud pathways, procurement leverage, evaluation infrastructure, as well as the ability to walk away from systems that no longer serve them:
1. Stop treating compute as a back-office concern.
If an AI health project relies on foreign cloud infrastructure, that is not a technical footnote. It is a sustainability, security, and sovereignty question. Practitioners should document where data are hosted, what compute is required, what contractual protections exist, who can access the data, how the system will be monitored, and what happens if the vendor relationship ends. Ministries should be briefed honestly on these tradeoffs before tools are deployed, not after pilots have already created dependency.
2. Fund an African-led cloud and compute convening.
Individual country compute build-outs are unlikely to be realistic in the near term. But African governments do need practical, defensible frameworks for adopting cloud and generative AI on terms they can hold politically and operationally. Hannah suggested an Africa-wide process where governments could compare when they have granted exceptions for cloud hosting and why; hear from technical, legal, procurement, and cybersecurity experts on the real options; and begin forming a negotiating bloc with hyperscalers and AI providers.
That process could focus on choosing cloud providers, single-cloud versus multi-cloud strategies, model contract terms, switching providers, data access and security, and appropriate generative AI use cases in health. Done well, it could produce shared principles and best practices, organizational and regulatory roadmaps, an ongoing peer network for African public-sector leaders, and a set of Principles on Cloud Computing and Data Sovereignty drafted by African leaders in digital technology, healthcare, and public infrastructure.
3. Build regional public-interest compute capacity.
The goal is probably not sovereign frontier AI infrastructure in every country. But it is reasonable to aim for enough regional compute capacity to adapt, evaluate, govern, and run priority public-interest use cases. African institutions should not only be consumers of tools built elsewhere; they need enough infrastructure to test models against local tasks, support local-language and clinical workflow adaptation, and reduce total dependence on systems controlled outside the region.
4. Require exit rights and portability.
Governments should not adopt AI systems they cannot leave. Every AI and cloud contract should include data export rights, open standards, APIs, documentation, transition support, non-punitive termination, and clear commitments that health data will not be used to train external models without explicit consent. Where appropriate, contracts should also address model and prompt logs, auditability, security review, and continuity of service. One ethical question is whether a country can adopt an AI system safely. Another is whether it can leave that system without losing access to its own data, workflows, or institutional memory.
5. Build fit-for-purpose testing and evaluation systems.
Before AI tools are scaled in LMIC healthcare settings, countries need local evaluation environments: test datasets, red-teaming protocols, language testing, clinical workflow testing, and post-deployment monitoring. Generic benchmarks are not enough. A model that performs well on a medical exam or English-language benchmark has not proven that it can support a nurse, community health worker, district planner, or ministry team working with local guidelines, incomplete data, constrained connectivity, and real accountability for patient and public-health outcomes.
