You Approved AI on Falling Costs. Except Costs Are Increasing. Why?

Every AI business case your organisation approved in the last two years rests on a price assumption nobody wrote down: that compute gets cheaper, and that capability gets better at the same price. The premise is load-bearing, it has no owner, and it has now inverted. The exposure is not a valuation event. It is the run cost of the systems you have already approved and are already running.

So the question is not only why the market turned. It is what your compute price is made of, and why nobody in your organisation told you it had moved. An assumption nobody wrote down is an assumption nobody was asked to watch.

On 1 July I published a piece called You've Modelled the AI Upside. You Haven't Priced the Bust. Halfway down it, inside a list about where a correction would show up on a balance sheet, sat this sentence: "Your investment case assumed a trajectory: capability up, cost down, availability stable." I wrote it, used it to make a different point, and never went back to it. That piece priced the downside. This one sets out the cost structure underneath it.

I should have gone back to it. Five weeks on, it is the only line in that article that has already been settled, and it was settled against the case I was making.

Because the framing of that piece was too generous to itself. Pricing the bust still treats the danger as a valuation event, something that happens to a multiple, an index, a supplier's margin, a number the CFO watches on a screen. What has actually turned up is duller than that and a great deal closer to you. A supplier is telling customers that 2027 is sold. A company with a finished, working AI feature has looked at what it costs to run and decided not to ship it. Neither of those is a market story. Both of them are invoices.

So here is the sharper version. Your AI business case has a cost line in it. It almost certainly does not have the assumption underneath the cost line, which is that the unit cost of the work falls, or at worst holds, across the life of the case. That was true for three years, which is precisely why nobody wrote it down. It is carrying your payback period, nobody monitors it, and it has stopped being true.

I run agentic AI in production across my own business, where it does the work of three people, and I pay those bills myself. The unit cost of the work I have automated is the number I watch most closely. It is also the number that has stopped behaving.

If you own an approved AI system, if you signed the case or you now carry the run cost in a budget line, this lands on you before it lands on anybody else. It will not arrive as a headline. It arrives as a figure on a monthly bill that is larger than the figure you put in the paper.

The assumption that was too obvious to write down

Business cases are full of assumptions, and the serious ones get written down. An exchange rate. Wage growth. An electricity tariff. Occupancy. Churn. Each of those gets a number, a source and, in a well-run organisation, a named person who watches it and a threshold that sends the case back for another look.

Compute cost never got that treatment, because for three years it did not appear to need it. The price of a million tokens fell. Models got better at the same money. Every quarter, the thing you approved got cheaper to run while you did nothing at all. When an input behaves like that for long enough it stops being an assumption and turns into background, and nobody writes down the background.

Go and read your own paper. You will find a run cost in the financials, usually a monthly figure or a per-seat licence, with no working behind it. What you will not find is a sentence stating what you assumed that figure would do over thirty-six months. You will not find a name against it either.

Every other input in the case is monitored by somebody. This is the one that moved in your favour for long enough that you stopped treating it as an input at all. So price it: who is buying it, and what it is made of.

Whose demand sets your price, and whose infrastructure you run on

Start with the demand your price competes against. Ben Thompson has been setting out the shape of it at Stratechery. Google Cloud's 82 per cent AI growth is substantially one customer. Its US$514 billion backlog is largely Anthropic. Banks are offloading US$15 billion of the associated debt.

The market reads that as concentration, and it is. Read it as the demand side of your own price, and that order book rests on very few names. Capacity in this industry is rented and then sublet. Your AI vendor may be running on a cloud that is itself reselling capacity it contracted from someone else, financed with debt that has already been sold on to a third party. There can be three or four balance sheets sitting between your invoice and the machine that does the work.

Which makes price certainty close to an illusion. Your supplier cannot promise you more certainty than it holds itself, and if it does not own the capacity it sold you, it does not hold much. APRA named concentration risk in April, flagging entities that lean on a single provider across multiple use cases with thin contingency planning underneath. This is the price half of the same problem. When the price of the underlying capacity moves, it moves through every party in that chain until it arrives at you, and you learn about it at the same moment as everybody else.

Here is the test. Take one AI system you rely on and answer two questions about it. Whose infrastructure does it actually run on, and does that party own the capacity or rent it? Most people cannot answer either question about a system that is already in production. Whoever it is, they buy the same physical inputs as everyone else.

2027 is already sold, and not to you

All of next year's memory production is sold. Hard drives are sold out as well. Those are not forecasts and they are not a market view. They are the present state of the supply chain your AI runs on, and committed supply does not respond to your order.

Memory is not software. You cannot conjure more of it in a quarter to meet demand. A fabrication plant is a multi-year, multi-billion-dollar commitment made long before the product ships, so next year's supply is contracted now, by buyers with vastly more purchasing power than you have. Once the supply is spoken for, the price is set by the people who spoke for it, and you are not in that conversation.

That is the mechanism people keep missing. Software got cheaper for thirty years because software has no marginal cost, and because the hardware underneath it got smaller and cheaper on a reliable cycle. The AI build-out has run into the other kind of input. Memory has a factory and a waiting list. Power has a grid connection and a queue, and the Financial Review has reported a national cabinet showdown over federal energy rules for data centres. Land has an owner who is allowed to say no, and in Kentucky one of them recently turned down US$26 million for a data centre site.

None of this requires a crash, a correction or a bubble to reach you. It only requires the price of the input to stop falling, and it has.

Canva ran the unit economics and declined to ship

The Australian Financial Review reported that Canva cut its revenue forecast because an AI feature was too expensive to run. Not because customers did not want it. Because the cost of serving it did not clear.

Read that as good news about Canva. A business built a feature, measured what it cost to serve, decided the unit economics did not work, and took the revenue hit in public rather than ship something that lost money every time it was used. That is a cost control that actually functions, exercised at the last responsible moment, by a company willing to say out loud why the number changed.

Now ask whether your organisation could have made the same call. Approval gates sit at the front of a project. Business case, funding, sign-off, steering committee. There is nothing at the far end. Once a thing is built and works and the team is proud of it, the default is to ship it. I have never read a position description that includes stopping a working feature on cost grounds, and I have never seen an incentive that rewards it.

So the question is not whether your unit economics are good. It is whether anybody in your organisation is permitted to look at them late, decide they do not clear, and stop. If you cannot name the person who holds that authority, nobody holds it, and the feature ships.

Your own system got more expensive without anyone changing the price

Canva could read its own unit cost. The last component of the structure sits inside your own four walls, and most organisations cannot read it. Hold every vendor price flat and the run cost of an approved AI system still drifts upward on its own, for four reasons that all sit in your build rather than in the market.

Agentic systems make many model calls per task where your pilot made one. Reasoning models spend far more tokens producing an answer than the model you benchmarked eighteen months ago, and you have almost certainly been moved onto one. Context grows as you feed the system more of your own documents, and you pay for context on every call. And usage climbs as people find the thing useful, which was the entire point of approving it.

So your unit cost can hold steady while your bill doubles. Or your bill can hold steady while your unit cost quietly doubles and your volume falls away underneath it. Those are opposite problems needing opposite responses, and neither is visible in a spend report.

Almost nobody is computing cost per unit of work. Cost per resolved ticket, per contract reviewed, per claim assessed, per document produced. Without that number you cannot tell a successful system from an expensive one, and you certainly cannot tell whether the case you approved still holds. That is the cost structure: demand you are queued behind, inputs nobody can make faster, resellers between you and the machine, and your own usage climbing inside it.

Find the assumption. Then make it someone's job.

This is addressed to the person who actually owns an approved AI system. The one who signed the case, or who now carries the run cost in a budget line and will be asked about it in October. Not the board. You. Every one of these uses data you already hold, takes an afternoon, and needs nothing from a consultant to get started.

1. Pull your two largest approved AI business cases and find the cost assumption. Not the cost line, the assumption underneath it: what you assumed a single unit of that work would cost, and what you assumed that price would do over the life of the case.
2. If it is not written down, write it down today, in one sentence, with a number in it. An assumption you cannot state is an assumption nobody can monitor, and this one has been holding up your payback period for two years.
3. Divide last month's actual invoice by the volume of work it covered, and put that figure beside the one you assumed. That single division is the whole diagnostic, and if the two numbers are far apart you have found your gap in twenty minutes using nothing but your own data.
4. Set a re-approval trigger on run cost per unit, with a threshold and a named person who pulls it. Decide now what movement sends the case back for another decision, the way you already do for an exchange rate or an energy tariff, so the call gets made on a schedule instead of in a panic.
5. Ask your vendor in writing whose infrastructure serves you and whether they own that capacity or resell it. The answer tells you what your price protection is actually worth, and a supplier who will not answer has told you something as well.
6. Give one named person the authority to decide not to ship on cost, the way Canva did. Write it into the role, because an authority nobody holds by name is an authority that will not be exercised on the day it is needed.

That is work we do, and it starts smaller than most people expect. We take the two or three AI systems you have already approved and reconstruct the cost assumption sitting under each one, then reconcile it against what you actually paid last month, so you finish holding a unit cost per system rather than a spend report. We trace the chain underneath each system until we can name who owns the capacity your vendor is selling you. Then we write the re-approval trigger into your governance with a threshold, a name and a review date against it, so the next move in run cost arrives as a scheduled decision instead of a surprise in a budget review. We run it as AI Governance as a Service, on a retainer, because this assumption does not break once and then sit still. It moves, and the value is in somebody being on it the month it moves rather than the quarter afterwards.

Nobody made a bad decision here. You approved a case on a premise that was true on the day you signed it, and nothing in your governance was built to notice when it stopped being true.

That is the whole gap. It closes in an afternoon.