Op-Ed written by Dr. Tolulope Ajidahun
A few years ago, as a newly qualified doctor, I sat at a computer at Alimosho General Hospital in Lagos and pulled up a patient’s history in under two seconds. No buffering. No “connecting…” No waiting on the network.
The hospital’s electronic medical record (EMR) system ran on a local server, reachable over the hospital’s own internal network. It did not care whether the internet outside was up or down. When the connection lagged or data ran out, nothing happened to the system itself. Vitals kept getting logged. Prescriptions kept getting pulled.
It did not occur to me that there was another way until I finished training and worked in other Lagos hospitals. I have not since used a records system in Nigeria that matched that level of reliability. Most of the digital tools we are building or buying for African health systems assume a constant, invisible connection to a cloud server somewhere, often outside the country entirely. That assumption is one of the biggest design flaws in health technology across much of the continent right now, and it is one we can fix.
This is not an argument against AI in healthcare. It is the opposite: if AI is going to help, it has to survive local conditions rather than assume them away. That means owning the model, the server, and the data, rather than renting all three from elsewhere.
Why Connectivity Is a Clinical Problem, Not an IT Problem
Nigeria’s connectivity is not consistent enough to be a dependency for anything where delay can cause harm. Hospitals and clinics outside Lagos and Abuja regularly face power cuts, patchy mobile data, and internet service providers that go dark for hours without warning. A cloud-dependent diagnostic tool or clinical decision-support system stops being useful the moment connectivity drops, and in much of the country, it drops often. In banking, that is an inconvenience. In a hospital, it can cost a life.
There is also the question of where patient data ends up. Nigeria’s National Health Act already treats patient information as confidential by default, and the Nigeria Data Protection Act of 2023, now reinforced by a 2025 implementation directive, places real obligations on anyone processing health data: registration requirements, data protection officers, and restrictions on moving data outside the country without proper safeguards. Every time a hospital sends patient records to a foreign cloud service for an AI-generated summary, that data — often concerning HIV status, mental health, or fertility treatment — leaves Nigerian jurisdiction, governed by terms most clinicians never read and regulators cannot fully audit.
―
Cost compounds the problem. Cloud AI is billed per query, in dollars, against a naira that does not sit still. A primary healthcare centre on a fixed government allocation cannot budget for that kind of variable, foreign-currency cost and expect it to survive a year.
The Fix Already Exists
None of this requires inventing new technology. Open-source models such as Llama, Mistral, and Gemma can be deployed offline, running entirely on local hardware with no data leaving the building. Researchers evaluating locally run versions of Gemma 2, Mistral Nemo, and Llama 3 for outpatient ear, nose, and throat care found these models could run on modest local hardware while still producing clinically useful suggestions — precisely because they wanted to avoid sending patient data to an outside cloud service. Google’s MedGemma, a model line built for medical text and imaging tasks, has been adapted by other research teams to run locally on hospital infrastructure for the same reason.
–
This pattern is already proving itself in other resource-constrained health systems. Hikma Health, a nonprofit, built a free, open-source electronic health record designed offline-first for clinics serving displaced populations in Lebanon and Nicaragua — places where connectivity is as unreliable as it is in much of Nigeria. A 2026 scoping review of medical AI in low-resource settings concluded that models designed to work offline, with local data caching, are a pragmatic strategy precisely because they let clinical work continue when the internet does not.
None of these projects waited for perfect connectivity. They built around its absence.
What It Actually Takes
Strip away the terminology and the setup is straightforward. A local server, or in a smaller clinic a well-specified desktop, runs an open-source model. Workstations across the facility connect to it over the internal network – fast, contained, unaffected by whatever the internet service provider is doing that day. No patient note leaves the building unless a clinician deliberately decides it should. Model updates can happen in batches whenever connectivity allows, rather than as a constant dependency — a hospital needs a model that works reliably today, not real-time access to the newest release.
Power supply is a related and separate problem that deserves its own treatment, but it is worth noting that the same resource-constrained health systems running offline AI successfully are also the ones pairing it with solar or hybrid power, because the two problems are linked.
Does This Actually Work?
Alimosho’s EMR is the proof I trust most, because I watched it hold up under an ordinary outage on an ordinary day. It is not an isolated case. Hikma Health’s offline-first system has been running in active clinics for over a decade. And the broader pattern — local model, local server, local control — keeps appearing wherever researchers and clinicians have had to solve the connectivity problem directly rather than assume it away.
What is missing is not proof of concept. It is African builders adopting this as a starting design principle, rather than a workaround discovered after a cloud product has already
failed in the field.
The Honest Catch
None of this is free, and pretending otherwise would be dishonest. Training or adapting a model on local clinical data, disease patterns, and terminology takes real effort, and the clean datasets needed for that mostly do not exist yet. Maintaining a system without constant cloud dependency means a hospital needs at least one person who can troubleshoot it locally — a real cost in a sector already short on technical staff. Hardware is a genuine upfront investment, even if it works out cheaper over time than a recurring, dollar- denominated API bill.
Regulation cuts both ways here, too. The same data protection law that makes cloud- dependent systems risky also creates obligations for anyone running a local system: registering as a data controller, appointing a data protection officer, running impact assessments, securing the server itself against theft or breach. Offline does not mean unregulated. It means the regulatory exposure stays within the country’s own jurisdiction,
answerable to its own courts.
None of this is a reason to stay dependent on the cloud. It is simply the work that has to happen either way.
Where This Has to Start
We need to stop renting infrastructure for something this important when a proven alternative is sitting right there, ours to control. African healthcare does not need to wait for permission from Silicon Valley to use AI well. It needs builders who assume the worst case as the default: that the network will fail, the power will go out, and the system still has to work.
Alimosho’s EMR did not need a foreign cloud subscription or a data centre abroad. It needed a server in the building and a network that never left the room. If we are serious about building health technology for African conditions, that is where it starts: offline-first, locally controlled, designed for the country we have rather than the one with perfect bandwidth we keep building for. Own the model, own the server, own the data. Everything else gets easier once that is settled.
Sources
•
Nigeria National Health Act, 2014, Section 26 (patient confidentiality provisions)
Nigeria Data Protection Act, 2023, and the Nigeria Data Protection Commission’s General Application and Implementation Directive, 2025
Brotherton T, et al. (2022). “Development of an Offline, Open-Source, Electronic Health Record System for Refugee Care.” Frontiers in Digital Health 4:847002
Al-Ganad A, et al. (2026). “Deploying medical AI in low-resource settings: a scoping review of challenges and strategies.” Frontiers in Digital Health 8:1743634
Public developer documentation on local deployment of Llama, Mistral, and Gemma model families via frameworks such as Ollama and llama.cpp
