The New Economics

Your team built a working prototype in an afternoon. What used to take weeks took hours. The feeling was euphoric — until someone asked what it costs to run in production. No one could answer. Not because they lacked competence, but because that question never needed to be asked that way before.

The cost structure has inverted. It used to be expensive to build and cheap to run. Months of development, then a predictable operational cost you could budget in advance. Now it's the opposite. Building costs almost nothing — an agent and an afternoon. But running what gets built has become a variable that scales with every user, every request, every inference. And that variable doesn't appear in any planning document.

Every decision your team makes — which model, how large a context window, how many iterations the agent is allowed before delivering — is an economic decision. It hasn't been before. Architecture decisions used to be about performance, scalability, maintainability. Now they're also about money. Not as a side effect, but as a primary consequence. The engineer choosing between a smaller, faster model and a larger, more capable one is making a procurement decision — they just don't know it, and neither does the person who owns the budget.

But it doesn't show up in the budget. It doesn't show up in sprint planning. It doesn't show up until the invoice arrives. And the person who owns the budget wasn't in the room when the decisions were made. Not because anyone shut them out — but because no one understood that economic decisions were being taken.

This is a structural problem, not a competence problem. The budget model assumes that the build phase is the major investment and that operations is a predictable tail. That model no longer describes reality. Today, a team can build ten prototypes in a week. The question is no longer whether you can afford to build. The question is whether you can afford to run what you've built.

The hidden inversion

Your best engineer now delivers what used to require a small team. But the cost didn't disappear — it moved. The budget still sees one salary, one "FTE". It doesn't see the compute bill that grows with every task that person produces.

The CFO is planning around headcount. The CTO is making architecture choices that carry real financial weight. Both are doing their jobs well — inside models that no longer describe the same reality. One is tracking people. The other is generating compute costs. And no spreadsheet connects the two.

The cost paradox

There's a counterintuitive dimension that makes this harder, not easier, to manage. The cost of inference is falling — rapidly. New hardware generations are pushing the price per computation down at a pace that resembles the early days of cloud computing. The intuitive conclusion is that the problem will solve itself: if inference gets cheap enough, the variable cost becomes negligible.

History suggests the opposite. When a resource becomes dramatically cheaper, organisations don't use the same amount at lower cost — they use dramatically more. Cheaper cloud compute didn't reduce infrastructure spending. It enabled architectures that consumed orders of magnitude more compute than anyone anticipated. The same pattern is already visible with AI: as inference costs drop, teams deploy more agents, use larger context windows, run deeper iteration loops, and apply AI to problems they previously wouldn't have considered. The unit cost goes down. The total spend goes up. And the unpredictability of that spend increases, not decreases.

This means the structural budget problem doesn't go away with better hardware. It intensifies. The gap between what teams can build and what finance can track widens with every generation of cheaper, faster inference. The need for economic visibility into AI-driven operations isn't a temporary challenge that market forces will resolve. It's a permanent feature of this new cost landscape.

The growing gap

The people who build understand this intuitively. The people who pay for it haven't been given the tools to see it yet. And the gap between them grows with every passing week. The question is no longer whether your organisation can afford to use AI. The question is whether it knows what it's spending.