On 14 July 2026, Cape Town’s municipal planning tribunal approved two data centres expected to draw around 170 megawatts, one of the largest such developments proposed in South Africa. The approval was granted before the city had established what the facilities would actually consume in water and electricity. A tribunal in a country that has rationed power for a decade committed the load first and will learn the number afterwards. That inversion, obligation before disclosure, is the pattern this entire build-out runs on.
Compute is arriving in Africa faster than at any point in the continent’s digital history. The International Finance Corporation committed $100 million in debt financing to Raxio Group in April 2025 to expand colocation facilities already operating in Côte d’Ivoire, Ethiopia, the Democratic Republic of Congo, Mozambique, and Uganda. Africa’s total operational data-centre capacity nonetheless sits at roughly 307 megawatts, under two percent of the global figure, which is why every new project is received as a correction to a historic deficit rather than as a transaction with terms.
The terms are the story. That foreign firms dominate African cloud services is established and no longer analytically interesting. What has changed is where the decisive control point now sits. It has moved off the corporate balance sheet and onto a government desk in another country, in the form of an export licence that determines whether the accelerators inside an African building may lawfully continue to operate there at all. A company can be negotiated with. A licensing regime is not a counterparty.
The Licence Is the Real Ownership Document
In April 2026 the US Department of Commerce opened a call for proposals under the American AI Exports Program, inviting industry consortia to assemble what the programme calls full-stack packages. A qualifying package bundles AI-optimised hardware, data-centre storage, models, cybersecurity provision, and sector applications into a single exportable offering. Designated consortia receive expedited export-licence review and prioritised access to federal financing, and their package enters a standing menu that the US government promotes to partner states. The submission window ran to 30 June 2026, and the International Trade Administration received 78 applications.
Two features of that design matter more than the money. The first is bundling. A state that accepts the package accepts the hardware, the model layer, the security stack, and the applications as a unit, which means the dependencies are contracted together and expire together. The second is exclusion: the programme is explicitly structured to exclude Chinese companies and technologies. A recipient country is therefore not simply buying compute. It is accepting a position in a technology bloc, and the terms of that position are set by an export-control authority it does not elect and cannot petition.
This is the layer the sovereignty debate has been missing. Data-residency law relocates records. Local incorporation relocates a tax base. Neither touches the question of whether an accelerator installed in Lagos or Nairobi remains lawfully operable when the licensing jurisdiction revises its rules, and that question is answered abroad regardless of who holds the title deed to the building.
A data centre becomes sovereign compute at the moment its operation no longer depends on a permission granted somewhere else. Almost none of Africa’s new capacity meets that test.
Housed Is Not the Same as Governed
The distinction that African procurement rarely draws is between where a machine sits and what governs it. A facility can be built on African soil, connected to an African grid, staffed by African engineers, and still be operated by an entity incorporated elsewhere, running a control plane administered elsewhere, on hardware whose export status is determined elsewhere, under a service agreement whose termination clauses were drafted for a different legal system.
Each of those is a separate lever, and each is separately negotiable at the point of contract. Ownership of the accelerators can be transferred or retained by the vendor. Workload portability can be contractually guaranteed or left unspecified. Suspension rights can be enumerated or left to the operator’s discretion. Continuity obligations in the event of a licensing change can be priced in or ignored. African governments have signed a substantial number of these agreements. Almost none of the resulting terms are public, which means no legislature on the continent can currently answer the question of what happens to a national workload if the licence changes.
The Bill Approved Before It Was Read
The resource side follows the same logic. Cape Town’s approval drew objections from advocacy organisations including Foxglove and Housing Assembly, arguing that the development would press on infrastructure that is already constrained. Equinix has stated that its facilities will use dry air cooling and will not require significant water for server-room cooling. That claim may well be correct. The structural point is that it functions as an assurance rather than an enforceable commitment, because the approval preceded the disclosure that would have allowed it to be tested.
Power is the harder constraint and the more revealing one. A 170 megawatt industrial load in a constrained grid is not a neutral addition; it is an allocation decision with a counterparty on the other side who does not appear in the tribunal record. Whether that allocation is defensible depends entirely on what the facility produces for the economy that supplies the electricity, and that is precisely what the contract terms would show and the public documents do not. A concession is not identified by who signs it. It is identified by an arrangement in which a scarce national input is committed for a long period in exchange for benefits that were never quantified.
The Version Africa Could Actually Negotiate
The asymmetry is not fixed. Grid access, water allocation, land, tax treatment, and public-sector demand are all African assets, and they are the assets these projects require. What is missing is not leverage but a common disclosure standard that would let leverage be exercised deliberately rather than deal by deal.
The smallest instrument that would change the terms is a mandatory sovereign-compute disclosure schedule attached to every public-sector AI or cloud contract. It would state operator jurisdiction, ownership of the hardware, model and software dependencies, data-transfer routes, termination and suspension rights, workload-portability provisions, power and water obligations, and the specific conditions under which service can lawfully be withdrawn. It is a publication requirement, not an ownership rule, which is why it is achievable, and it would arm every finance ministry on the continent with the one thing they currently lack when a delegation arrives: the ability to compare what they are being offered against what a neighbour already signed.
The verdict is that the most consequential document African governments will sign in this decade is not an AI strategy, a memorandum of understanding, or a national digital plan. It is an unpublished commercial contract that determines who can switch the machines off. The continent is being offered the appearance of sovereignty, which is capacity physically located inside its borders, in place of the substance of it, which is capacity that keeps running when someone abroad changes their mind. Those two things can be told apart, but only by reading the terms, and the terms are currently the one part of the transaction nobody is required to show.
