In April, South Africa withdrew its draft national artificial-intelligence policy after the document was found to have cited several fabricated academic references. These sources turned out to be AI ‘hallucinations’. The policy had been partially drafted using AI. The irony was hard to ignore: a framework designed to govern AI was undermined by the very technology it sought to regulate.
But beyond the hallucinated citations lies a deeper issue. The document reflected a broader assumption that is increasingly visible in AI policy worldwide: that global AI systems will be built by replicating and expanding the same computational and infrastructure model that has been pioneered in Silicon Valley for wealthy economies. Modern AI systems were developed under conditions of extraordinary abundance: cheap capital, plentiful energy, vast computing infrastructure and access to land and water for cooling. This has shaped the assumption that building ever larger models and increasing their scale is the way forward. But it is already running into physical constraints.
Nature’s message to South Africa’s next government: talk to your researchers
Data centres are notorious guzzlers of water and electricity. South Africa is a water-scarce country; until a year ago, it resorted to scheduled blackouts (referred to as load shedding) to manage its frequent electricity shortages. As a scientist based in South Africa, building various sizes of language models for emerging economies at Lelapa AI — a resource-efficient language AI company in Johannesburg — I see this disconnect as a fundamental strategy flaw.
High energy use is one barrier. The International Energy Agency, an intergovernmental body in Paris that shapes energy policies, projects that global electricity consumption for data centres will more than double between 2024 and 2030, reaching 945 terawatt-hours — approaching the entire annual electricity consumption of Japan in 2025. Even in infrastructure-rich regions, such as the United States and Ireland, electricity grids are struggling to keep up with the rapid growth of AI, prompting communities to push back against data-centre construction.
There are economic constraints, too. Hardware and running costs are high even in the wealthiest markets, and most AI companies aren’t yet making a profit from sales of AI products. To ‘super-scale’ their infrastructure quickly, some firms are relying on private credit — raising questions about whether the current pace of capital expenditure can be sustained by underlying revenue growth alone.
How South Africa can move on from power cuts
Nations are increasingly putting AI sovereignty — control over crucial AI infrastructure and systems — at the core of their security and growth agendas. Governments such as those of India, Saudi Arabia and countries in the European Union are investing in domestic compute infrastructure, cloud capacity and foundation AI models to achieve AI sovereignty and reduce dependence on foreign technology providers.
