Who Captures the Value of Cheaper Intelligence?

A reduction in the cost of an important input can make an existing business more profitable without changing its product, its prices or its customers. That is the simplest economic possibility created by cheaper inference, and it should be the starting point rather than something explained away by predictions about increasingly elaborate AI products. The more difficult question is how much of the initial saving a business can retain, and what happens when the same opportunity becomes available to its competitors.

The cost trend is substantial, although its measurement needs care. Stanford's 2025 AI Index reported a decline of more than 280-fold between November 2022 and October 2024 in the inference price associated with achieving approximately GPT-3.5-level performance on MMLU. This compares access to a specified benchmark capability over time. It does not measure the total cost of every AI application or establish what happened to provider margins.^[1]

An illustrative service makes the immediate effect clear. Suppose it earns £100, consumes £20 of inference and incurs £30 of other delivery costs. Its contribution before unchanged shared costs is £50. If inference expenditure falls to £5 while everything else remains the same, contribution rises to £65. The customer need not pay more, want more or receive a different product for the supplier to benefit.

That comparison holds the service constant. A company might instead spend the saving on more thorough analysis, a broader feature set or additional checking. Its inference bill could remain unchanged while the product improves. Alternatively, it might serve more customers. These are choices about using the improvement, not evidence that the original cost saving was unreal. Combining them obscures the difference between making existing work cheaper and choosing to do more work.

A useful assessment therefore distinguishes the price of inference from the quantity consumed, the resources needed per completed task and the value of the resulting output. A lower price per token can coexist with a higher monthly bill if usage increases. That alone says nothing decisive about whether the business is better off. The additional work might be highly valuable, or it might cost more than customers are willing to pay for it.

Implementation costs belong in the same calculation. A less expensive model that requires substantial integration, checking or correction may not lower the total cost of delivering the service. Conversely, a model that prevents repeated work can create a saving larger than its advertised token-price reduction. The commercially meaningful comparison concerns equivalent acceptable outcomes, with the costs of achieving them included.

Once an actual saving has been established, market structure helps determine its destination. A supplier with an agreed selling price may retain it initially. Customers able to compare competing offers may subsequently obtain some of it through lower prices or improved service. Neither result follows from the inference cost curve alone. Contracts, differentiation, available capacity and the credibility of switching all affect how quickly a productivity improvement changes commercial terms.

Routing infrastructure can make some alternatives easier to exercise. OpenRouter, for example, offers provider selection by price, latency and throughput, alongside controls for supported parameters and data policies. That creates a mechanism through which buyers can compare and redirect suitable workloads. It does not establish that every endpoint is an adequate substitute or that providers have lost all ability to charge above cost.^[2]

Model developers and inference providers also benefit when their own costs fall. A declining selling price need not imply a declining margin if operating costs fall faster. An efficient provider may share part of its improvement with customers and retain part itself. Higher volume can increase total profit even where profit per request falls, although the investment needed to serve that volume still matters to the return earned on capital.

The same distinction applies further down the chain. An application company does not automatically enjoy a permanent advantage because its model bill becomes smaller. Rivals may receive the same saving and compete more aggressively. But access to similar models does not make the applications identical: distribution, integration, specialist information and the quality of the completed service can still differ. Those differences determine how much competitive pressure reaches the selling price.

Consider a hypothetical manufacturer selling an established product. If AI reduces order-processing costs without changing the product or its price, existing customer relationships can become more profitable. Some previously loss-making relationships might cross into profitability; others might simply lose less. There is no requirement for formerly profitable customers to become unattractive. A lower cost can expand the range of worthwhile business rather than exchange one profitable customer group for another.

That improvement is distinct from an immediate payroll saving. Where employees remain on the same terms, less administrative time may release capacity, avoid a future hire or allow additional orders to be handled. The value depends on whether that capacity can be used. It would be misleading either to count every saved minute as cash or to dismiss the improvement merely because the wage bill has not yet changed.

Employees and suppliers of complementary resources also belong in the distributional picture. In one possible outcome, more productive work supports higher remuneration or a shorter working week. In another, reduced labour demand places pressure on earnings. Scarce production capacity, specialised information or access to customers might capture part of the gain elsewhere. These are potential channels, not predictions that every participant benefits equally.

Consumers can receive value without it appearing as supplier revenue. A Stanford Digital Economy Lab working paper used online choice experiments with US adults in July 2025 and March 2026 to estimate the compensation people would require to give up access to AI chatbots for a month. Its resulting consumer-surplus estimates substantially exceeded estimated revenues from those services. The study concerns access valuation under its methodology, not the isolated effect of falling inference prices, company profits or a measurement of the whole economy.^[3]

The distinction is important because spending is not a complete measure of benefit. A customer might obtain the same useful service for less money, or a more useful service for the same money. A company might find an exchange worth making that previously consumed too many resources. Some transactions will merely move between suppliers; others could represent genuinely additional activity. Revenue growth at one business does not establish which has occurred.

Nor is economic value a fixed pot that every participant simply divides. Better technology can reduce the resources needed for existing output and make new output worthwhile. Competition then influences prices and bargaining positions within that enlarged set of possibilities. Transfers between parties should be distinguished from resource savings, additional useful production and any costs imposed outside the transaction.

The resulting market need not have a single permanent beneficiary. A provider might initially retain an efficiency gain, an application company might later use it to improve its product, and competition might eventually deliver more of the benefit to customers. Different workloads can follow different paths. The timing and distribution of gains are questions to investigate, not answers contained in a token-price chart.

Cheaper inference can therefore produce a straightforward improvement in an unchanged business, and sometimes something more consequential: a wider range of services and customers that can be served economically. Whether the gain remains in margins, supports better work or passes into customers' purchasing power depends on the surrounding market. The cost reduction creates the opportunity; the structure of the business and its competition determine what becomes of it.