AI can work while its financing disappoints

A useful technology can arrive on time and still disappoint the people who financed it. Customers can benefit, adoption can grow and the underlying engineering can improve, while particular businesses struggle to earn enough on the capital they have committed. There is no requirement for these outcomes to move together.

AI makes this easy to forget because arguments about its usefulness often turn into arguments about its valuation. A convincing demonstration is treated as evidence that enormous investment will pay off. A strained balance sheet is treated as evidence that the technology was exaggerated. Both conclusions skip several steps.

The distinction was acknowledged quite plainly by Andrew Bailey in remarks published by the Bank of England in October 2025. He described the compatibility of believing AI is a major technology and worrying about stretched valuations and financial stability. That remains a useful way to frame the question without requiring everyone to choose a camp.

There are at least three economic questions inside an AI project. Does the system produce something useful? Will a customer pay enough for that result? Can the business deliver it at a cost, and on a timetable, that supports its commitments? The first answer can be encouraging while the other two remain unresolved.

Imagine a service that cuts the time needed to analyse a large set of documents. Customers may value it highly. If competing services quickly provide similar results, much of the benefit may reach customers through lower prices. The technology succeeds at making an activity cheaper. The original supplier may have difficulty retaining the margin assumed when it raised money.

Falling costs make that possibility particularly interesting. Stanford's 2025 AI Index documented a steep decline in inference prices at a specified benchmark performance level between November 2022 and October 2024. That historical comparison is not a promise about every future workload. It does illustrate how quickly the cost of a capability can change, creating opportunities for users while complicating suppliers' forecasts.

Infrastructure brings another timetable into the picture. A long-lived building, a power agreement, a financing schedule and a generation of computing equipment do not share the same useful life. A business can be broadly right about growing demand and still have the wrong assets available at the wrong price. Paying interest while waiting for a connection is a very practical form of uncertainty.

The Bank's October 2025 Financial Policy Committee record also identified competition, infrastructure constraints and technical breakthroughs that change infrastructure requirements as possible threats to expected earnings. These are different routes to financial disappointment. Some arise because progress is slower than expected; others because progress takes a cheaper or unexpected form.

There is a serious counterpoint. Building capacity before demand is fully proven can be necessary. Waiting for perfect certainty would leave useful infrastructure unbuilt and ambitious research unfunded. Some businesses have the balance sheets, distribution and operating cash flows to absorb a long period of investment. A large capital programme is not, by itself, evidence of recklessness.

The question is how much uncertainty the financing allows. Flexible commitments leave room to adapt. Fixed obligations can turn a manageable delay into an urgent problem. Forecasts should therefore be examined for their dependence on timing, utilisation and pricing, rather than judged only by the size of the market they describe. A very large market can accommodate plenty of unsuccessful companies.

This matters to operators as well as investors. A business building essential workflows on a supplier should understand how dependent it is on continuing subsidies, unusually favourable terms or services that may change. Portability, sensible fallback arrangements and an understanding of full operating costs are useful even when the supplier looks strong. They are ordinary precautions in a young market.

It is possible to be enthusiastic about what AI enables and exacting about how a particular business pays for it. Keeping those judgements separate leads to better questions. What is already useful? Who receives the economic benefit? Which commitments depend on assumptions that have yet to be tested? Those questions will remain worth asking through both cheerful markets and less cheerful ones.