"Sovereign AI" has become one of the most expensive phrases in public technology. Around the world, governments are announcing national models, national compute centres and national AI clouds, often with large figures attached. The motivations are serious: dependence on a handful of foreign suppliers is a genuine risk, and languages, laws and public services that the global models handle poorly are a genuine gap.
The difficulty is that the phrase covers both a set of legitimate public goals and a set of purchases that may or may not serve them. A data centre with a flag on it is not the same thing as control. A national model that few departments use is not the same thing as capability. When the goal and the purchase are not separated, procurement ends up buying the most visible option rather than the one that achieves the goal.
This is the third part of After the Hype. Part 1 and Part 2 dealt with value inside organisations. This part applies the same discipline to public spending on AI infrastructure, with a South African lens. The companion article in the Rewiring Government series, Sovereign AI is more than data residency, sets out what control actually consists of. This article asks a narrower question: how should an accounting officer decide whether a proposed sovereign AI purchase is worth the money?
The critique, and what it gets right
The sharpest version of the critique calls national AI programmes vanity projects: subsidies dressed as strategy, funding redundant models and data centres that may be outdated before they are commissioned. A recent Emerald Book commentary makes that case forcefully. Its specific cost figures are not sourced, so we do not repeat them, and its tone leaves little room for the legitimate reasons governments act. But its central warning is sound. Public money spent on AI infrastructure has to be justified by what it achieves for the public, not by what it signals.
In South Africa that warning is not only a matter of opinion. The Public Finance Management Act makes accounting officers responsible for the effective, efficient, economical and transparent use of their institution's resources (section 38(1)(b)), and municipal managers carry an equivalent duty under the Municipal Finance Management Act. A proposal that cannot show why its chosen option was the economical way to meet a stated goal has a governance problem before it has a technology problem.
Five legitimate goals
The first step is to name the goal. There are five that hold up under scrutiny.
Jurisdiction over data. Public data should be stored and processed under South African law, with the state able to enforce its rights. This is real, but it is often met without new infrastructure. Several hyperscalers operate cloud regions in South Africa, and contract terms, customer-managed keys and audit rights do much of the work. CloudNala's articles on where data is allowed to live and on why POPIA does not require your data to stay here explain where the legal line actually sits.
Continuity. A public service should survive the failure, withdrawal or sanctioning of a supplier, and the loss of international connectivity. That calls for multi-supplier design, tested exit plans and a local fallback for critical services. It does not automatically call for a national model.
Language and context. Global models serve some official languages much better than others, and know little about local law, place names and public processes. This is one of the strongest arguments for public investment, but the bottleneck is usually data and evaluation, not compute. Funding curated datasets, evaluation suites and the tuning of open models may achieve more than funding a cluster.
Skills and industry. A country that only consumes AI does not develop the people who can build, operate and govern it. Training, research funding and shared access to compute for universities and small firms build capability. A large facility used mainly by one programme may not.
Bargaining power. A government that cannot leave a supplier cannot negotiate with it. Portable data, open standards and enforceable exit rights preserve bargaining power at a fraction of the cost of building an alternative.
Each of these is a reason. None of them is, on its own, a reason to build. The right-hand column of the diagram is the set of cheaper options a proposal should show it has tested and found insufficient.
Where sovereign spend goes wrong
Four patterns recur when the goal and the purchase are not separated.
Capacity without demand. Compute is bought on the assumption that departments will use it, before any department has a workload ready. Accelerators lose value quickly as newer generations arrive, so idle capacity is not a neutral delay. It is a depreciating asset.
A national model without a use. A model is trained to show that it can be, then struggles to find users because departments need an integrated service, not a model. The value was always going to depend on the narrow, embedded uses described in Part 2.
Residency mistaken for control. A local data centre run on a foreign supplier's control plane, identity system and update channel is local storage, not sovereignty. The companion Rewiring Government article covers the layers that matter.
No operating plan. The capital budget is approved; the operating budget for the engineers, energy, security and refresh cycle is not. Within a few years the facility can be under-maintained and its hardware outdated.
An eight-question procurement test
A proposal for sovereign AI infrastructure should answer eight questions, in order.
- Named goal. Which of the five goals does this serve? "All of them" is not an answer; rank them.
- Failure it prevents. What specific failure or dependency does the goal address? A supplier withdrawing service, a language that global models handle badly, a skills gap in a named area.
- Cheapest option that meets it. Which options were assessed, including the ones in the right-hand column of the first diagram, and why were the cheaper ones insufficient?
- Full lifecycle cost. Capital, operating, energy, staffing, security, hardware refresh and decommissioning, over the realistic life of the asset.
- Evidence of demand. Which departments or institutions have committed workloads, with volumes and dates? Letters of intent are not workloads.
- Skills to operate it. Who will run, secure and maintain it, and where will they come from? What happens if they leave?
- Refresh and exit plan. How will the infrastructure be refreshed or retired, and how could the state move away from it if a better option emerges?
- Public outcome measure. What will be different for citizens, officials or the economy, how will it be measured, and when will the decision be reviewed?
The order matters. Cost cannot be judged without knowing which option it buys, and an option cannot be chosen without a named goal. A proposal that stalls at question 1 should not be costed at all.
The same test, applied to a modest proposal
Consider a hypothetical provincial proposal to build an AI compute centre to "ensure sovereign AI for provincial services". Applying the test:
- The goal turns out, on discussion, to be language and context: citizen services in isiXhosa and Sesotho that global models handle unevenly.
- The failure it prevents is poor service quality for citizens using those languages, not supplier dependency.
- The cheapest option that meets it is a programme to build evaluation sets and curated data for the languages, and to test tuned open models running in an existing local cloud region under provincial contract terms.
- Demand evidence shows two departments with ready workloads, both small.
The proposal is redesigned. The province funds the language data and evaluation work, procures hosted capacity in-country with exit terms, and records that the case for its own facility will be revisited if demand grows past a stated level. The goal survives. The spend shrinks. That is what the test is for.
A note for suppliers and advisers
Suppliers will naturally present the most comprehensive option. Advisers, including CloudNala, should be held to the same test. A recommendation to build should carry its own answers to the eight questions, and a recommendation not to build should explain how the goal will still be met. The transfer test for vendor claims in The Boardroom AI Test series applies to sovereign AI proposals as much as to enterprise ones.
Next: Part 4, The AI claims register. All parts are on the series page.
How CloudNala can help
We help public institutions test AI infrastructure proposals before they are funded. We work through the eight questions with the accounting officer's team, assess the cheaper options against the stated goal, build the lifecycle cost and demand evidence, and set out a decision that can be defended to oversight bodies. This is technology and procurement analysis, not legal advice.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book an AI Value Review or write to us at consult@cloudnala.co.za