In April 2026, South Africa’s ambition to lead on artificial intelligence (AI) governance stumbled. Weeks after Cabinet approved the Draft National AI Policy, Communications Minister Solly Malatsi withdrew it when at least six of its 67 academic references proved fabricated. The draft policy had proposed a national AI commission and an insurance fund for AI harms. Much of its substance was sound; the unverified references buried the stronger ideas along with the flaws. Ironically, the department charged with building guardrails for AI had used generative AI to draft the policy without verifying the results.
This misstep speaks to a debate now running across the Global South about technology and power. A critical element of AI governance is digital sovereignty, which is a fundamental building block for states trying to control critical technologies, data flows and algorithmic systems. ‘Tech colonialism’ is the term used to describe the status quo: the Global South supplies data and demand while ownership of the infrastructure, and the power to set its rules, is based elsewhere.
As recent research on digital dependency in the Global South argues, states adopt data protection acts and national AI strategies while the relevant infrastructure, models and standards remain under external control, meaning that formal sovereignty is intact but functional sovereignty is out of reach. Sovereignty is the capacity to choose your dependencies rather than have them imposed, and authority follows ownership: whoever owns the machines sets the terms for everyone else. Until South Africa, the greater African continent and BRICS countries, China aside, own more of the infrastructure AI runs on, their sovereignty is rented. Pooling resources is the only credible way of building it.
South Africa is a vocal advocate for inclusive global AI governance. During its 2025 G20 presidency, Pretoria argued that developing countries should help write AI rules, not merely implement standards set elsewhere. Yet leadership abroad depends on capability at home. Africa hosts about 1% of the global data centre capacity, and what exists mostly runs models rather than trains them – a divide researchers call ‘Compute North’ and ‘Compute South’. ‘Compute’ is the processing power needed to train and run AI models. Compute North therefore refers to countries that host AI used for AI development, while Compute South refers to those whose AI compute is more relevant for AI deployment.
BRICS countries, however, have greater collective capacity and heft. What the bloc cannot offer is a common rulebook. Its members regulate differently by design, a response to unequal capacity rather than a failure of coordination, which makes collective bargaining, not harmonised rules, a more realistic possibility for BRICS.
To counter the dominance of Western-led standards in AI rule-making, the pattern that has led to some of the accusations of ‘tech colonialism’, BRICS hopes to build an alternative framework. The group is attempting to shift the language from Western-centric ethical guardrails toward state sovereignty and development rights, demanding that AI governance be anchored in the UN rather than OECD-led bodies. The BRICS Data Economy Governance Understanding mapped a Global South data economy outside Global North ecosystems. The India AI Impact Summit 2026 added the New Delhi Declaration, endorsed by over 80 countries, which makes democratising compute an objective. In July 2026, 29 countries, including Russia, Brazil and South Africa, founded the World Artificial Intelligence Cooperation Organisation, a Chinese-initiated body based in Shanghai that positions Beijing as a rule-maker. The US stayed away. Each development has created more space for alternative rules, although none has added a chip fabricator or a megawatt.
The hard infrastructure of AI remains concentrated mostly in the US. The US and China host roughly 86% of the world’s data centre capacity, and the US accounts for around three-quarters of the high-performance computing used to train advanced AI. When the compute is based elsewhere, so is the power to set prices and rules, decide what may run and switch off access altogether, as recent export controls on advanced chips have shown.
The asymmetry is normative too: fewer than one in 10 AI governance frameworks originate in the Global South, leaving Africa, Asia and Latin America as rule-takers.
Global South governments and enterprises are dependent on US hyperscalers such as Amazon Web Services, Microsoft Azure and Google Cloud, which control close to two-thirds of the market. Russia has built its own clouds, sealed off from those hyperscalers. The rest of the BRICS countries, those without China’s chips or Russia’s closed networks, are caught between sovereign aspirations and rented infrastructure. Brazil has a strong data protection law and India mandates local data storage, yet both rely on processing power from proprietary US tech platforms. South Africa, the bloc’s loudest champion of inclusive governance, commands the least compute.
China is the obvious alternative. Its laboratories release their strongest models as open-weight, and those now account for a larger share of global downloads than US ones. Weights already downloaded cannot be switched off from Washington, which is why Malaysia built its national AI programme using DeepSeek’s LLM and Singapore chose Alibaba’s Qwen over Meta’s Llama. For states priced out of training their own, that is a real gain. But it buys one layer only. Chinese firms have begun holding back their best systems; models carry the assumptions of the places that train them; and a downloaded model still runs on someone else’s chips.
Sovereignty operates on many layers, and the attainable ones – data protection laws and declarations – are not enough. A recent Stanford survey of commercial AI sovereignty offerings found that most only reconfigure dependence: national clouds and AI factories may keep data onshore, but the chips and the underlying expertise stay foreign. The decisive layer is that physical base, and there the Global South remains a tenant rather than an owner.
The AU’s Continental AI Strategy emphasises Indigenous solutions, but electricity is proving a hurdle. Kenya suspended a billion-dollar data centre in May 2026 because it would have drawn a third of national generating capacity, and South Africa faces Eskom-projected shortfalls as coal plants retire. When projects do get built, ownership tells the same story. Cassava Technologies’ AI factory runs on Nvidia hardware, as an Nvidia cloud partner, with Nvidia among its shareholders. InkubaLM, an LLM trained in isiZulu, isiXhosa and other African languages, is the rare model built and owned by African teams.
The goal is not to match the hyperscalers but to choose where sovereignty matters most: pooling compute regionally, backing small open-source models and negotiating better access. South Africa can build on a state-owned high-performance computing centre already shared across the region, and treat its withdrawn draft policy as a lesson rather than a debacle. For BRICS, hosting its summit in New Delhi this September, it means steering New Development Bank finance toward shared compute and collective access to chips. Rule-writing authority accrues only to those who own what the system needs.
A version of this op-ed was published by Business Day.
