The Economics of Intelligence: Why AI Gets Cheaper When It Thinks

There is a straightforward story about the economics of artificial intelligence, and it goes like this: intelligence costs money. Every API call to a frontier model consumes compute. Every token generated burns through GPU cycles that someone, somewhere, is paying for. The bigger the model, the more expensive the inference. The more you use it, the more you spend. This is not controversial. It is arithmetic.

The numbers are significant. A single query to a large language model can cost anywhere from fractions of a cent to several dollars, depending on the model size, the length of the prompt, and the complexity of the response. At enterprise scale, organisations running thousands of queries per hour find themselves facing monthly bills that rival their headcount costs. OpenAI reportedly spent over $700 million on compute in 2023 alone, and that figure has only grown.^[1] The economics are clear: more intelligence means more cost.

This model has historical precedent, and it works. We pay for electricity by the kilowatt-hour. We pay for cloud computing by the instance-hour. We pay for water by the gallon. Consumption-based pricing is the natural model for scarce resources, and compute is genuinely scarce. Hal Varian, Google's chief economist, argued persuasively that the fundamental economic contribution of AI is reducing the cost of prediction — turning expensive human judgment into cheap machine inference.^[2] Agrawal, Gans, and Goldfarb formalised this insight in their influential work on prediction machines, demonstrating that as the cost of prediction falls, the value of complementary human judgment rises.^[3]

And the costs are indeed falling. William Nordhaus has documented a remarkable trend: the real cost of computation has declined by a factor of roughly ten every six to eight years for over a century.^[4] Moore's Law, even in its weakened modern form, continues to deliver cheaper transistors. New chip architectures from NVIDIA, AMD, and a growing ecosystem of AI accelerator startups push inference costs downward with each generation. If you simply wait long enough, AI gets cheaper.

So the thesis is optimistic, and it is largely correct on its own terms. AI is expensive, but it is getting cheaper, and the value it delivers is growing faster than the cost. For many applications, the return on investment is already compelling. The question is whether this framing captures what is actually happening — or whether it obscures a structural problem that no amount of hardware improvement will fix.

The Metered Intelligence Trap

Consider what happens when you hire a human expert. On day one, they are expensive and slow. They need to learn your systems, understand your domain, internalise your constraints. By month six, they are dramatically more productive. By year two, they have accumulated institutional knowledge that makes them qualitatively more valuable than they were at the start. The cost of their salary stays roughly flat, but the value they deliver compounds. This is what economists call learning-by-doing, and it is one of the most powerful forces in economics.^[5]

Now consider what happens when you use an LLM. On query one, it costs a certain amount per token. On query one million, it costs the same amount per token. The model has not learned anything from your usage. It has not accumulated any understanding of your domain. It has not built any reusable capability. Every interaction starts from scratch within its context window. You are paying for the same cognitive work over and over again, and the system never gets better at serving you specifically.

This is not a minor inefficiency. It is a fundamental structural defect in how AI economics currently work. Ronald Coase's theory of the firm explains why organisations exist: they reduce the transaction costs of coordinating economic activity.^[6] The current AI pricing model does the opposite. It creates a permanent transaction cost for every unit of intelligence consumed. There are no economies of scale on the demand side. There are no efficiency gains from experience. There is only the meter, ticking.

The incentive structure this creates is perverse. Providers benefit from usage volume, which means they benefit from systems that require repeated inference for the same types of problems. There is no economic incentive to help users need the API less. Lock-in is not a bug — it is the business model. Switching costs are high because the integrations are deep, the prompt engineering is non-transferable, and the fine-tuning is provider-specific. Erik Brynjolfsson and Andrew McAfee have documented how digital platforms tend toward winner-take-all dynamics, and AI APIs are no exception.^[7]

The heaviest users get punished the most. An organisation that integrates AI deeply into its workflows — using it for code review, document analysis, customer support, internal search — faces costs that scale linearly with usage, or worse. Volume discounts exist, but they do not change the fundamental shape of the curve. The marginal cost of the millionth query is not meaningfully different from the marginal cost of the first. There are no compounding returns. Daron Acemoglu has argued that the productivity gains from AI may be more modest than anticipated precisely because the technology is being deployed in ways that do not generate cumulative advantages.^[8]

Compare this to software, which has near-zero marginal cost once written. A spreadsheet formula, once created, can be reused indefinitely. A database query, once optimised, runs efficiently forever. A function in a codebase, once debugged, works every time it is called. Software accumulates value. The current AI model consumes value — it converts money into ephemeral outputs that vanish when the context window closes.

The heaviest users — the ones who should benefit most from AI — are instead the ones most trapped by its economics. They cannot easily switch providers because their workflows are deeply integrated. They cannot reduce costs through experience because the system does not learn. They are, in economic terms, captive consumers of a non-appreciating asset. This is the opposite of how productive technology is supposed to work.

Intelligence That Appreciates

There is an alternative model, and it starts with a simple observation: every time an AI system solves a problem, the solution could be captured, verified, and stored as a reusable algorithm. Not the raw output — the underlying pattern. Not the text of the answer — the method by which the answer was derived. If you do this systematically, something remarkable happens: the cost of intelligence starts declining with usage instead of staying flat.

This is not a theoretical abstraction. It is how human expertise actually works. A senior engineer does not re-derive sorting algorithms from first principles every time they need to sort a list. They draw on a library of known solutions, accumulated over years of practice and study. The first time they encountered a problem, it was expensive — hours of thought, false starts, debugging. The hundredth time they encounter a similar problem, the cost is nearly zero because the solution pattern is already in their repertoire.

The economic principle at work is what W. Brian Arthur calls increasing returns — the idea that in knowledge-intensive industries, the more you use something, the more valuable it becomes.^[9] This is the opposite of the diminishing returns that govern physical goods. A car depreciates with use. A knowledge base appreciates with use. The question is whether AI systems can be designed to operate on the appreciating side of this divide.

The answer requires a different architecture. Instead of treating each inference as an isolated transaction, the system must treat each inference as a potential contribution to a growing library of verified solutions. When the system encounters a problem it has seen before — or a problem structurally similar to one it has solved — it retrieves the existing solution rather than regenerating it from scratch. The LLM becomes the discovery engine, not the production engine. It finds solutions; the system remembers them.

Consider the economic implications. On day one, the system is expensive — it must use inference for everything because its solution library is empty. By month three, a significant fraction of incoming problems match patterns already in the library. By year one, the majority of work is handled by cached, verified algorithms that cost nothing to execute. The marginal cost of intelligence converges toward zero, not because hardware got cheaper, but because the system got smarter about what it already knows.

This architectural shift changes the fundamental economics in several ways. First, it transforms AI from a consumption good into a capital good. In economic terms, a consumption good is used up when consumed — food, fuel, per-token API calls. A capital good is something that produces other things — a factory, a tool, a knowledge base. When intelligence accumulates rather than evaporates, it becomes capital.^[10]

Second, it aligns the incentives of the system with the interests of its users. In the current model, the provider profits from repeated inference — they want you to keep calling the API. In the accumulation model, the system's goal is to make itself unnecessary for previously solved problems. Every successful caching of a solution pattern is a permanent reduction in cost. This is the economic equivalent of a tool that sharpens itself with use.

Third, it introduces genuine network effects on the demand side. If solutions verified by one user can be shared with others — appropriately anonymised and generalised — then the library grows faster than any single user could fill it. The economic literature on network effects, pioneered by Katz and Shapiro, shows that products with demand-side economies of scale tend to become dramatically more valuable over time.^[11] A shared intelligence library has exactly this property.

The Depreciation Inversion

Physical assets depreciate. A car loses value the moment you drive it off the lot. A computer is obsolete in three years. A building requires constant maintenance just to avoid decay. This is the default behaviour of things in the physical world: they wear out.

Knowledge assets, when properly maintained, do the opposite. A mathematical proof is as valid today as when Euclid first wrote it. A well-tested algorithm works as reliably on its millionth execution as on its first. A verified solution to a recurring problem becomes more valuable over time, not less, because it saves more cumulative cost the longer it operates.

The implication for AI economics is profound. If a system converts expensive inference into verified, reusable solutions, then intelligence appreciates for the user. Each problem solved adds to the library. Each addition reduces future costs. The system becomes more valuable the more it is used — not because the underlying model improved, but because the accumulated solutions expanded.

Simultaneously, intelligence depreciates for the provider, in the sense that the user needs less inference over time. This is the correct economic relationship. A good tool makes its users more capable and more independent, not more dependent. The provider's business model must therefore be built around delivering increasing capability, not extracting increasing rent. Romer's endogenous growth theory describes exactly this dynamic: sustained growth comes from the accumulation of ideas and knowledge, not from the extraction of rents on scarce resources.^[12]

From Renting Intelligence to Owning It

The shift from per-token pricing to accumulated cognitive capital is not merely an efficiency improvement. It is a change in the ownership structure of intelligence itself. In the current model, intelligence is rented. You pay for access to a model you do not own, generating outputs you cannot systematically build upon. In the accumulation model, intelligence is owned. Every interaction produces a durable asset — a verified solution, a tested pattern, a reliable algorithm — that belongs to the user and compounds over time.

This distinction matters enormously for organisations making long-term strategic decisions about AI adoption. An organisation renting intelligence faces perpetually rising costs as it deepens its AI integration. An organisation building its own cognitive capital faces declining costs — and rising capability — over the same timeline. The difference between these two trajectories is the difference between a cost centre and a strategic asset.

The practical requirements for this shift are demanding but achievable. The system needs a verification layer that can distinguish correct solutions from plausible ones. It needs a formalisation engine that can extract reusable patterns from specific instances. It needs a retrieval system sophisticated enough to match incoming problems against stored solutions even when the surface-level framing differs. And it needs rigorous testing infrastructure that ensures cached solutions remain valid as conditions change.

None of these requirements are exotic. Formal verification, pattern matching, and test-driven validation are well-understood engineering disciplines. What is new is applying them systematically to the outputs of AI inference, treating the LLM not as the product but as a discovery mechanism feeding a growing library of proven solutions.

The economic incentive to build such systems is powerful. Any organisation currently spending significant sums on AI inference has a direct financial motivation to convert that expenditure into durable assets. The question is not whether this shift will happen. It is whether it will happen through proprietary platforms that centralise the accumulated intelligence, or through open architectures that allow users to own and control their cognitive capital. The answer to that question will shape the economics of intelligence for decades to come.^[13]

Intelligence should appreciate for the user and depreciate for the provider. Any system that achieves this inversion is not merely cheaper. It is structurally better — and the economics, once inverted, do not flip back.