Every enterprise AI budget built on today's flat-fee pricing is built on a number that was never real. The subsidy that made AI feel cheap is ending everywhere. In South Africa, it collides with a second, entirely local problem: the intelligence is billed in dollars, and the revenue it needs to justify is earned in rands.
I keep meeting South African technology leaders who are budgeting AI the way they budgeted their first cloud migration a decade ago: take this month's invoice, multiply by expected growth, submit to the CFO.
That approach survived cloud computing reasonably well. It will not survive AI, and the reason has almost nothing to do with the technology getting worse. It is getting better, and cheaper, by almost every published measure.
The bill is going up anyway. Understanding why and what to actually do about it, is now board-level literacy, not an engineering concern to be managed downstream.
The subsidy was never the price
Think back to the early days of ride-hailing, when a trip across town cost a fraction of what the driver, the fuel and the vehicle actually required. That price was never real. It was venture capital buying market share and building habit, and it held only until the reckoning arrived.
The first widely adopted price for frontier AI − roughly 20 US dollars a month − followed the identical playbook. It was a calculated introduction, not a market price, and the economics behind it make that obvious: the world's largest hyperscalers put an estimated $410 billion into AI infrastructure in 2025 alone, with projections reaching $650 billion in 2026, while OpenAI reportedly spent $1.35 for every dollar of revenue it earned that year.
No business model sustains that gap indefinitely. The subsidy is not a permanent feature. It is a phase, and the phase is ending.
Three phases, one blind spot
Industry analysts generally describe generative AI's short life in three technology phases, each with a different economic signature.
The first was simple availability, ChatGPT, followed by open-source models, proving the technology worked and was worth exploring.
The second introduced reasoning models capable of working through multi-step problems rather than producing a single fast answer, and with it came the first real price competition as lower-cost distilled models entered the market.
The third phase, the one enterprises are living through now, is agentic: systems that plan, call tools and chain multiple steps together autonomously. This is where the blind spot sits.
Inference prices genuinely have been falling somewhere between 75% and 90% a year. Token volume − the actual number of tokens an agentic workflow consumes to complete the same task a chatbot once answered in one exchange − has been growing 500% to 1 000% over the same period.
Price and invoice are moving in opposite directions, and the gap between them is the entire story.
The rand problem no global report mentions
Every one of those figures is quoted in US dollars, because that is the currency AI is priced and settled in almost everywhere.
South African enterprises earn in rand and are billed in dollars, and that mismatch does not show up in a single global FinOps report, it has to be added on top.
Local market research already shows organisations using hyperscalers exposed to fluctuating operational costs purely from rand-dollar volatility, with many having already experienced real cloud bill shock unrelated to any change in usage.
South Africa's local cloud services market is projected to grow from roughly R49.6 billion in 2025 to R101.5 billion by 2029 on the back of this dynamic.
None of this is a reason for caution about adoption. South Africa already leads the continent, with Microsoft estimating that 23.1% of the working-age population used a generative AI product in the first quarter of 2026.
It is a reason to budget in a currency the organisation does not control, for a consumption pattern the organisation cannot yet forecast, which is a materially harder problem than the one most global AI cost guidance is written for.
The agentic multiplier
The single most important number in AI cost planning for 2026 is not the price per token. It is the multiplier that autonomous agents apply on top of it.
Agentic workflows can consume 50 to 500 times more tokens than the equivalent simple chat interaction for the same underlying task, and three forces compound that: agentic volume multiplication itself, hidden background consumption from monitoring agents, document watchers and compliance surveillance systems that run continuously whether or not a human triggered them, and retrieval overhead, the additional tokens consumed every time an AI system searches internal documents before answering, a cost step almost no early enterprise budget modelled.
Every interaction is measured in tokens, roughly three-quarters of an English word each, and in agentic systems the model also generates internal reasoning tokens that are never shown to the user but are billed exactly the same as visible output.
The cost of intelligence is falling. The cost of deploying intelligence is not. For a rand-denominated business paying a dollar-denominated bill, that gap compounds twice over.
Build vs buy: Token consumer or token provider
As prices rise and consumption grows, enterprises face a genuine strategic fork: remain a token consumer, renting intelligence from frontier providers, or become a token provider, running models on owned or dedicated infrastructure.
The tipping point where owning becomes cheaper than renting depends on usage volume, workload type and technical maturity, and it is not a decision to make once and forget.
There is currently roughly a 4 500-times pricing spread between the cheapest and most expensive frontier models. Enterprises that default every task to the best available model, including the large share of tasks a materially cheaper model would handle just as well, are quietly paying a premium they never approved.
For most South African enterprises, the realistic answer is not a single choice but a hybrid one: predictable, high-volume workloads routed to smaller or locally hosted models, and frontier capability reserved for the tasks that genuinely need it.
Five actions for South African enterprise leaders
Treat AI spend as its own governed line item, not a cloud sub-category: The FinOps discipline built for infrastructure spend does not map cleanly onto token-based, agent-driven consumption. It needs its own owner, its own dashboard, and its own forecasting model built in rand, against a dollar-denominated cost base.
Model-route by task, not by habit: Map which workloads genuinely require frontier reasoning and which are being over-served by the most expensive model available out of convenience. Given the pricing spread between models, this single discipline is often the fastest lever available.
Instrument agentic workloads before scaling them: Background consumption, retrieval overhead and internal reasoning tokens are invisible until they are measured. Pilot agentic workflows with usage telemetry switched on from day one, not added retroactively once the invoice arrives.
Model the rand-dollar exposure explicitly: Budget AI spend the way treasury teams budget any material dollar liability, with a stated view on currency sensitivity, rather than folding it into a flat annual IT number that assumes today's exchange rate holds.
Negotiate for volume predictability, not just unit price: A lower per-token rate is worth little if consumption is growing 500% to 1 000% year-on-year. Contract structures that cap or smooth agentic-era volume growth protect the budget more than a marginally better headline price.
My question for every South African finance and technology leader reading this: if your AI consumption doubled next quarter, purely from agentic workflows doing what they were designed to do, would anyone in the room be able to explain why within hours, or would it take a currency shock and a shareholder question to surface it?
The technology is already ahead of most budgets. Left unmanaged, in a business earning rand against a dollar-priced input, that gap closes at a cost the balance sheet feels twice.
* Eugene Perumal is a strategy and architecture principal with over 20 years' experience in enterprise technology across telecoms and financial services, including senior roles at Vodacom Group and Absa Group. He holds Master’s degrees and certifications in enterprise architecture, AI governance, cloud and analytics. He writes on enterprise AI strategy, ROI measurement and the shift to agentic AI deployment.
Share
