This product has already been purchased.
- AI may be getting cheaper per use, but total spending is rising fast as organisations become moredependent on it.
- Boards need to manage AI as a strategic risk, with growing exposure to vendor lock-in, supplier concentration and future price increases.
- Directors should focus on total AI spend, dependency risks and maintaining flexibility to switch providers before those costs become difficult to control.
AI is fast becoming essential to how organisations
run, yet its cost, concentration and dependency are slipping past the board.
Directors need to treat it as a core governance issue before it is locked in,
writes Stéphane Chatonsky GAICD.
Nearly every board is now encouraging its organisation to adopt AI while exercising little control over how much it is used, what it costs, or what it commits the business to. A reassuring story makes this easy to ignore: per-token prices are collapsing, so AI looks cheaper. The total bill, however, is heading the other way. It is on track to become one of the largest, and least controllable, lines on the P&L, and that is exactly the kind of thing directors are meant to watch: cost, concentration, and dependence on a handful of suppliers.
I have watched this in miniature at home. In under three years, my personal AI spend went from nothing to roughly $2,700 a year, and it keeps climbing. If my provider doubled the price tomorrow, I would pay it. Run that same pattern through a business, across every function now leaning on AI, and the only things that change are the number of zeros and how hard it becomes to unwind.
Why AI spend belongs on the board agenda
Australian directors already sense this. In the AICD’s Director Sentiment Index for the first half of 2026, close to 90 per cent expect business costs to rise, and more than half say AI is changing faster than their organisation can keep up with.
For two years the question has been “is AI worth it?” The better one now is: what happens to margins, supplier risk and your freedom to change course when AI becomes essential, the bill keeps compounding, and you cannot turn it off? The two numbers move in opposite directions. Management can show falling unit prices while the total invoice climbs. Three forces drive that gap, and each of them is a risk directors need to know:
- The build-out has to be paid back. The four largest US hyperscalers plan to spend around US$725 billion on infrastructure in 2026 alone, up roughly 77 per cent on the year before, and Goldman Sachs models some US$7.6 trillion (A$10.8 trillion) of AI capital expenditure through 2031. All of it will need to be recovered through prices.
- Even the sellers are losing money. The leading labs are burning significant cash to run their models, so today’s price is a subsidy funded by investors and debt. When that subsidy thins, the bill resets, and not downward.
- It is becoming non-optional. Frontier models are moving from nice-to-have to table stakes across a whole range of functions. The moment a capability becomes a necessity, pricing power shifts to the seller, as directors have seen before.
Cheaper tokens, bigger bills
Per-token prices really have cratered. GPT-4-level output fell from about $20 to well under $1 per million tokens. Yet Gartner expects total inference costs to rise, because usage grows faster than price falls: reasoning and agent models consume many times more tokens per task than the chatbots before them. Gartner forecasts AI coding costs will overtake the average developer’s salary by 2028.
So, when management shows a chart of falling unit costs as reassurance, ask for the chart that matters: total spend, and where it is heading.
The rise of AI vendor lock-in
The comforting story is that if a provider gets greedy, you switch. But once you have built agents and embedded workflows around a particular model, switching becomes a re-engineering project, not a simple configuration change.
Analysts estimate switching costs at 19-34 per cent of the original build. That creates a familiar governance risk in a new form: vendor lock-in, supplier concentration and third-party dependency. These are risks the AICD and Human Technology Institute highlight in A Director's Guide to AI Governance.
The collapse of AI developer Builder.ai in 2025 illustrates the point. Despite high-profile backing, customers built on its platform were forced to migrate at their own expense when the company failed. Provider failure is not a theoretical risk. It is a potential operational and financial disruption.
Some organisations are responding by designing for optionality from the outset. Walmart and JPMorgan, for example, route workloads across multiple AI providers and their own models so no single vendor has undue influence over pricing or access. Maintaining that flexibility is increasingly an engineering decision boards should understand and monitor.
The cost lever you are used to pulling no longer works
People are the largest controllable cost, and we have decades of muscle for managing it. AI does not behave that way. Automate customer service and it works beautifully, until the provider raises prices or an upgrade burns twice the tokens per ticket. You cannot freeze the spend without switching off a service customers now expect, because you did not add a tool, you replaced the people who did the work. Your most successful automation becomes your least predictable expense, costing the most exactly when business is good.
What directors should do now
None of this is an argument against AI; the organisations that hold back will be left behind. It is an argument for treating AI spend with the seriousness it deserves, while there is still time to shape it. Here are five questions every board should be asking:
- Total spend, not unit price. Are we reporting total AI and inference spend and its growth rate, shown next to the falling per-token cost, rather than only the reassuring chart?
- Concentration. How much of our AI dependency sits with a single provider, and do we manage it as the supplier-concentration risk it is?
- A price shock. What would a doubling of provider prices do to our margins, and what would we actually do about it? If the honest answer is “pay it,” that is the finding.
- Optionality. Where it is affordable, are we favouring designs such as abstraction layers, more than one model and portability, so that switching stays a choice we made rather than one imposed on us?
- Clear ownership. Have we assigned AI cost and dependency to the full board, the audit and risk committee or a technology committee, and reflected it in the risk register?
Token prices may keep falling, but total AI spend is likely to keep rising. The boards that treat this as a governance issue now, weighing cost, concentration, dependency and optionality together, will be the ones still holding some leverage when the sale ends.
Stéphane Chatonsky GAICD is a professional director and an investor in and adviser to Australian AI and technology companies.
Latest news
Already a member?
Login to view this content