The Part of AI That Has a Bill Attached

Most conversation about artificial intelligence concerns capability: what these systems can do, how quickly that is changing, and what it might mean. Comparatively little attention goes to a plainer question that determines which of those capabilities ever reach a customer — what it costs to run them, every single time.

Two different bills

There are two distinct costs to separate. Training is a large, one-off expenditure to build a model. Inference is the cost of using it: the compute consumed each time it produces an answer. Training dominates headlines because the figures are dramatic. Inference dominates business models because it recurs.

Software taught a generation of businesses that serving one more user costs approximately nothing. Inference quietly revokes that assumption. A product built on model calls has a genuine marginal cost per use, which puts it structurally closer to a logistics or manufacturing business than to classical software.

Cost per useful output, not cost per token

The commonly quoted metric is price per unit of text processed, which has fallen substantially. It is also the wrong denominator for most decisions. What a business actually pays for is a completed, acceptable piece of work — and getting there may involve retries, longer prompts, additional reasoning, and a human reviewing the result.

Measured that way, the economics vary enormously by task. Where an output is cheap to verify and usually acceptable first time, the cost per useful result is low. Where it must be checked carefully by someone qualified, the verification can cost more than the generation, and the saving largely disappears.

The physical constraint underneath

Behind the software sits infrastructure with long lead times: chips, data centres, electrical capacity, cooling, transmission. These are heavy, long-lived commitments made against demand that is still being discovered, which is the classic setup for both durable platform-building and significant overbuild. Prior technology cycles have produced both, sometimes simultaneously.

What this implies for reading AI business claims

• —Ask what the cost is per completed, acceptable unit of work — not per token.
• —Ask who absorbs the verification cost, and whether that cost was counted.
• —Treat a business model that only works if inference costs keep falling as a forecast.
• —Distinguish committed capital spending from demand that has actually materialised.
• —Watch power and capacity as the binding constraint on ambitious timelines.

None of this is an argument that the technology is unimportant. It is an argument that the economics are ordinary, and that ordinary economics is where the interesting questions usually live. Capability determines what is possible; the cost per useful output determines what gets built.