Adrian Vanzyl on AI’s Rising Price Tag for Business.
Australian businesses are spending more on AI even as the cost of running it falls-and I think, like Adrian Vanzyl, that this contradiction is the most useful thing to understand about where AI spending is headed. Deloitte’s Stu Scotis, who leads the firm’s AI practice, says planning an IT budget has become tougher, not easier, because AI doesn’t behave like the fixed costs businesses are used to. Businesses once planned IT spending as a static line item, but AI has made those costs more variable.
The unit at the centre of that variability is the token – the basic measure of how much computing effort an AI system uses to complete a task, with more complex requests consuming more tokens and simpler ones consuming fewer. On paper, this should be good news for budgets: increased competition among AI providers, more efficient chips, and a historic wave of data centre construction have pushed the price of each individual token sharply lower. But that falling unit price hasn’t translated into falling bills, because businesses keep finding new, more demanding tasks to hand to AI, and each new task tends to be more computationally complex than the last – consuming more tokens per request even as each token itself gets cheaper. Adrian Vanzyl finds this the crux of the puzzle: the price of the ingredient is dropping, but the recipe keeps getting more elaborate.
Scott frames the shift in a way I find genuinely clarifying: human workers earn dollars, while AI systems effectively earn tokens. Businesses now need to evaluate what they get from that spending just as they would evaluate a human employee-by asking what value, outcome, or growth it actually produces. That reframing matters because it moves the conversation away from “Is AI cheap or expensive?” and toward “Are we getting a return?”, which is a harder and more useful question.
AI expert Jon Whittle, formerly a technical lead at NASA and now involved with CSIRO’s AI work and Australia’s National AI Centre, has pointed to a related problem: many businesses have taken what he calls a scattergun approach to deploying AI, adopting it broadly without a clear framework for where it adds value. That approach makes the cost side of the equation harder to model, not easier – particularly given his observation that AI systems themselves are often poor at estimating in advance how many tokens a given task will actually require. In effect, businesses are trying to budget for a cost they can’t reliably forecast, using a tool that can’t reliably forecast it either.
The Real Bottleneck Isn’t the Price Tag
What strikes Adrian Vanzyl most is that none of this is really about token prices at all – it’s about measurement. A business that can’t estimate token consumption in advance, and can’t clearly tie that consumption to a business outcome, is flying somewhat blind regardless of how cheap or expensive tokens happen to be that quarter. The token price is falling; the uncertainty around what any given AI deployment will actually cost, and what it will actually deliver, is not.
Conclusion
I’d treat “AI is getting cheaper” and “AI is costing us more” as two claims that can both be true at once, and not contradictory so much as describing different variables. The unit price is a supply-side story about chips and competition. The total bill is a demand-side story about how ambitiously – and how carefully – a business chooses to use the technology. Getting that second part right looks less like a pricing problem and more like a discipline problem.
