The cloud has a postcode: reflections on the CNBC Africa AI Summit 2026

In late August, I attended the CNBC Africa AI summit in Johannesburg, South Africa. The conference included eleven sessions with more than thirty speakers. The discussions focused on who would own the systems, data and economic value created through AI. There was considerable optimism about what AI could bring to the continent and what it was already doing. The debate shifted from asking “can we?” to “who owns it?” Most of the answers depended on a policy environment that does not yet exist. 

In the same week, a different set of concerns was being raised in South Africa. Five civil society organisations announced a joint submission to the South African Human Rights Commission about the expansion of data centres. They called for an investigation into the sector’s impacts and urged government to consider a temporary pause on new hyperscale approvals and expansions. Water, electricity and land were central to their concerns. In early September, the Commission told the Associated Press it had received more than 250 submissions from interested groups. 

The two groups (the summit and the campaigners) were talking about the same technology using different tones. This piece is about the gap between them. The summit’s themes and focus did not ignore the environmental risks but rather framed them in a way that made their costs disappear. I use four lenses to examine this gap: evidence, scale, risk, and governance. 

1. Evidence: Contextualising AI 

“AI” covers different technologies and uses. A chatbot that drafts a response, a tool that recommends a clinical action, and an agent with a specific set of authorisations. Underneath all these tools is the same physical chain. A machine-learning system that works by collecting or accessing data for training. Training is then needed to adjust the model’s parameters so it can identify patterns and produce predictions or responses. Using a trained model also requires computation, so its resource-demands continue after training. 

Processing these data gives off a great deal of heat, which requires energy to cool it off and, in many designs, water is widely used to cool them down. Where water cooling is used, part of it is lost from the immediate environment through evaporation. 

Understanding the data is equally important as understanding the technology. Models learn only what they are shown, which means that their understanding of the world is shaped by the information used to train and operate them. If African languages, institutions and economic conditions are underrepresented, a model can perform well on a general benchmark while being unreliable for a particular local task. This was well reflected through the second session of the conference. 

The evidence of AI’s benefits is growing across different sectors such as education, health, and agriculture, while the extent and distribution of its risks are still debated and argued. The success stories presented at the summit were persuasive, but the evidence available on benefits and costs was uneven. A company can count the conversations handled by its chatbot. However, public information on how much water a data centre uses in South Africa remains limited. Civil society organisations attribute this gap to the lack of mandatory disclosure of facility-level water and electricity consumption, and the absence of a single regulator with a full view of the sector. So, the benefit side of the ledger arrives in decimals, and the cost side arrives as an argument. When uncertainty is substantial, its sources should be made explicit. Understanding AI’s benefits, costs, and risks requires evidence grounded in the conditions and contexts where it is deployed. 

2. Scale: small globally, large locally 

From a global perspective, ambitions to expand computing capacities can be understood as an opportunity to strengthen the continent’s position in the digital economy. From a city’s perspective, however, one large campus can represent a substantial addition to power and water demand. A project that is attractive in a national investment case can still be a difficult question for the substation and reservoir that serve it. Looking at the evidence generated at different scales points us towards different directions. Research on scales warns us about carrying conclusions from one to another without fully understanding the relationship between them. 

At a continental level, the European Union (EU) regulates AI systems through its AI Act and addresses data-centre energy and water reporting through a separate framework under the Energy Efficiency Directive. The African Union (AU) on the other hand addresses the environmental risks directly as part of the Continental AI Strategy and sets a shared direction for member states. However, moving to the next scale by translating the continental ambitions into practice requires national policies and decisions about individual developments. 

The question is whether these arrangements align: Do national investment priorities account for local resource constraints? Do companies report environmental performance in ways that allow comparison across facilities and countries? Can local authorities obtain the information needed to assess a development against the capacity of the systems supplying it? Shared principles are useful, but their value depends on how they inform decisions across the different levels. Alignment should support comparable assessments while allowing requirements to reflect local conditions. 

3. Risk: cost and benefits 

The third lens asks whose values define the risk. The fourth panel’s discussion of cybersecurity, information manipulation and model abuse raised a concern that extends beyond protecting digital infrastructure: whether institutions can trust the systems informing their decisions. That trust requires understanding both whether a system has been compromised and whose priorities it serves when functioning as intended. A model optimising a mine or a fishery must work towards some definition of success. Whether that definition accounts for ecosystem damage, livelihoods or animal welfare depends on choices that are themselves political. For governments, retaining control over AI therefore includes the capacity to question its objectives and the costs it leaves out. Human oversight has substance only when those overseeing a system can recognise and challenge these choices, with those bearing the consequences able to influence them. 

4. Governance: Sovereignty and environmental responsibility 

The last lens is governance. The summit’s emphasis on ownership raised an important question: what would it take for African countries to shape AI around their own priorities? The risk of algorithmic colonisation arises when technologies developed around external values and assumptions are adopted with limited local scrutiny, potentially crowding out local alternatives and deepening dependence on foreign software and infrastructure. A model trained primarily on formal economic activity may overlook how livelihoods operate in an informal economy. Sovereignty therefore requires the capacity to question those assumptions and influence how systems are developed, evaluated and used. Some initiatives demonstrate how African language speakers and researchers can participate in developing and evaluating technologies that reflect their contexts. 

But the sovereignty argument is usually made about data and models and stops there. The same logic applies more forcefully to the ground. If it is a loss of sovereignty to run public services on a model nobody local can inspect, it is a larger loss to allocate a catchment’s water to a facility whose consumption nobody local is allowed to know. Environmental responsibility is not an add-on to sovereignty. It is the part of it that cannot be outsourced, because the river does not move. 

Hosting data centres within a country’s borders can provide greater control over where sensitive data is stored and processed. It does not, by itself, establish control over the technology or ensure that benefits reach the communities hosting it. The decision to expand domestic hosting therefore requires weighing the control it offers against continued dependence on foreign providers and the demands it places on local resources. Sovereignty also concerns the capacity to make that decision on informed terms, including whether to approve a development and under what conditions. 

My takeaway from the summit was that it was optimistic, but conditional. Sovereignty tailored to African contexts begins with a clear understanding of the technology, its benefits and costs, in addition to investment in institutions, researchers and the capacity needed for it. A technology with global reach is grounded somewhere. The cloud has a postcode, and it comes with demands. Africa does not need to own every layer of the technology. It does need to know what each layer takes, from whom, and to hold the power to say no before the meter starts running.