South Africa Buys Its Intelligence In Dollars And Sells Its Output In Rands

Currency exposure, rising cloud fees and unviable local GPU economics make efficiency a condition of entry rather than a competitive edge.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/south-africa-buys-its-intelligence?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

Something has shifted in how the technology industry keeps score. The question that dominated boardroom conversation for three years, being which model is the most capable, is quietly giving way to a harder one about what that capability actually costs to use. For South African organisations, that question arrives with a complication their peers in New York and London do not face.

CONTEXT AND BACKGROUND

A second measure is now moving as quickly as raw capability, being how much useful intelligence an organisation can extract for each dollar it spends. The clearest signal comes from Amazon, which has been routing more Alexa+ traffic through its own weaker models rather than through stronger but costlier alternatives, with the aim of reducing cloud computing costs as the platform scales. What makes that choice defensible is how far the cheaper tier has come. Stanford's AI Index found that the cost of querying a model performing at GPT-3.5 level on the MMLU benchmark fell from twenty dollars per million tokens in November 2022 to seven cents by October 2024, a reduction of more than 280 times in under two years.

Capability that once carried a premium now costs almost nothing, which is why the weaker model is no longer the obviously worse answer. What has not fallen is the bill. Gartner expects worldwide end-user spending on AI models and platforms to reach 64 billion dollars in 2026, up 63.4 percent on the previous year, while noting that enterprise budgets face greater scrutiny centred on usage efficiency, cost control and measurable outcomes.

INSIGHT AND ANALYSIS

The difficulty for South African organisations sits in the denominator. Intelligence is priced in dollars and earned in rand. Local market research has found that enterprises using hyperscalers are exposed to fluctuating operational costs because of the volatile Rand cost against the United States dollar, that many organisations have already experienced cloud bill shock, and that the local cloud services market is expected to grow from an estimated R49.6 billion in 2025 to R101.5 billion in 2029. The same research notes that running large language models on-premises is unfeasible for most local enterprises because of the cost of deploying local GPU infrastructure. The usual escape from a supplier’s pricing power is therefore closed. Falling unit prices offer limited relief, because the shape of the work has changed. EY has documented a customer service interaction that once cost four cents becoming a $1.20 orchestration once tools, planning and subagents are involved, roughly thirty times higher. Cheaper tokens meeting heavier workloads produce larger invoices, denominated in a currency the business does not earn.

I have previously written about this in an article examining what happens when the finance function arrives to interrogate the AI budget, arguing that constrained capital, high operating costs and intense margin pressure leave South African organisations with no comfortable buffer for investments that underperform.

IMPLICATIONS

None of this argues for hesitation. South Africa leads the continent in adoption, with Microsoft estimating that 23.1 percent of the working-age population used a generative AI product in the first quarter of 2026, placing the country 46th of 147 economies measured, though adoption in the Global North is growing at more than twice the rate of the Global South. The task is to participate on terms the balance sheet can sustain. Three disciplines follow. Business cases should be denominated in Rand with the exchange rate assumption stated openly and stress-tested, rather than buried inside a dollar figure nobody revisits.

Measurement should move from cost per token to cost per completed task, since a cheap model that must attempt the work three times is not cheap. Workloads should be classified by risk, so that the least expensive adequate model handles routine volume while stronger models are reserved for decisions carrying regulatory or reputational consequence.

CLOSING TAKEAWAY

Efficiency is often treated as a refinement, something to attend to once capability has been secured. For South African organisations, it is the entry condition. Currency exposure, rising local hosting costs and the impracticality of self-hosting mean that extracting value from every unit of compute is not an optimisation exercise but the basis on which participation is possible at all. That constraint carries an unexpected advantage. Organisations forced to think early about cost per completed task will build a discipline that better-resourced markets are only now discovering they need. The question for boards is no longer which model is best. It is what a unit of intelligence costs the business, in rand, to do work worth doing.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net