← Jimmy E. Chan

What Is Frontier AI?

Capital That Performs Labor and Produces Ideas

August 2026

There is a recurring attempt to describe frontier artificial intelligence as an economic object. Some people argue it is a utility, like electricity, and others say that it is a platform, like iOS or Windows. Neither analogy is sufficient to understand the future impact of frontier AI in the world.

The utility analogy treats AI as interchangeable capacity that is meterable by the kilowatt-hour. But the value of a model query depends so heavily on context and quality that no two calls produce comparable output. A kilowatt-hour is a kilowatt-hour. A million tokens of frontier inference applied to drug discovery is not the same economic object as a million tokens applied to summarizing meeting notes, even if both cost roughly the same to produce.

The platform analogy doesn’t quite fit either. Bill Gates once defined a platform as “a system where the collective revenue of the participants who build on top exceeds the revenue of the platform itself”[1]. Windows, iOS, and to a large extent AWS all fit that description. The AI economy could eventually pass it too, but there’s heavy debate on where value accrues. Even then, a platform describes how revenue flows around a thing rather than what the thing itself is.

So if both of the most common analogies are wrong, then what kind of economic object are we actually dealing with?

Back to high school economics

Economists have long described the inputs to production as factors. The classical list is land, labor, and capital[2], with entrepreneurship added later[3]. The most significant modern addition is “ideas” (aka “technology” or “designs”), which behave differently from the others[4].

Each factor of production has its own defining characteristic. For instance, land is scarce and yields rent, labor is embodied in people and yields wages, and capital is itself a produced good used for further production, which yields a return. Entrepreneurship combines the classical factors under conditions of uncertainty, and yields profit as the residual after the others are paid. Ideas are “nonrival”, which means one person's use of them does not diminish anyone else's, and yield returns through excludability, via patents, copyrights, or trade secrets. The distinguishing feature of ideas or technology as an input is that “it is neither a conventional good nor a public good; it is a nonrival, partially excludable good”[4]. Nonrivalry turns out to drive most of long-run economic growth.

But frontier AI does not fit into any one of these factors. It sits across three of them and I will expand on each.

Capital

A frontier model passes the economics 101 test for capital. A model is produced, not found in nature. A training run costs a zillion dollars in compute, electricity, training data, and researcher-years, all of which become embodied in a set of weights. Weights are owned by frontier labs and they are durable enough to be used across many production cycles, and depreciate as better generations arrive on a regular basis. The lab earns a return on this capital through inference revenue, against the amortized training investment.

Structurally this is the same economics as a steel mill or a server farm. There’s a heavy upfront expenditure that produces a long-lived productive asset and generates returns over its useful life.

Labor

From a customer's perspective, the experience of using AI is often akin to commissioning cognitive work. A customer hands the AI a contract to review, a function to write, or a research question they want to investigate, and the model returns the finished work. The per-token pricing could map onto hourly billing and the transaction structure resembles hiring professional services.

How that work is then deployed varies widely. Many customers use frontier models to extend the reach of professionals they already have. Others use them to take on work they could not previously afford to do, or to make services available to people for whom they were previously out of reach. Currently, AI is an input that operates at the level of individual job tasks rather than whole jobs, and the economic structure of the transaction is the same in all of these cases. The customer buys a completed task at a per-unit price, whether it extends their existing staff or reaches customers they couldn’t serve before[5].

Ideas

Once a model is trained, its weights are nonrival. Serving a query consumes inference compute, but it does not consume the weights themselves, so a lab can serve the millionth customer from the same asset that served the first. This is the defining economic property of ideas rather than that of capital or labor. A factory cannot make two cars from one ton of steel but a model can answer two questions from one set of weights.

Frontier systems generate outputs that did not previously exist such as novel mathematical proofs (or disproofs), drug candidates, code, designs, and strategies. AI's largest impact may come from "serving as a new general-purpose 'method of invention' that can reshape the nature of the innovation process and the organization of R&D"[6]. A method of invention is effectively a technology for producing new ideas. Frontier AI may be the first such technology that can be invoked on demand, at scale, in domains that previously required years of expert human research.

Considering these three factors, a more precise and interesting statement would be that frontier AI is capital that performs labor and produces ideas. No prior factor of production has done all three from a single artifact. A cookbook cannot cook and generate new recipes while at it. A steel mill does not perform labor as it needs workers to operate it. A lawyer's advice serves one client at a time, and when a lawyer produces an idea, say a legal theory or a published article, the idea leaves the lawyer and becomes an artifact separate from the labor. A patent does not perform tasks. Arguably, chip designs or architectures come closer. Each generation of design is a nonrival recipe, and each builds on the previous one. But the design and the productive asset are separate artifacts done by separate companies. Nvidia and Arm design chips and TSMC manufactures them. The designer designs the chip, which is then produced by a fab, with the resulting chip actually doing the computation. Frontier AI does all three from one set of weights, which is what makes existing economic categories feel inadequate to describe it.

Why ideas matter most

Capital and labor have something in common in that both sell their services as “flow”. Flow is a term that economists use for quantities metered over time, like a machine-hour or a person-hour, and what a flow delivers is consumed at the moment of delivery. A factory-hour, a programmer-hour, or an inference call each yields a unit of output that is used up when delivered. Another unit has to be produced for the next customer. The total value such businesses capture per unit of work is anchored to what the customer would otherwise pay to have that unit produced. This is an enormous market in absolute terms. The world spends a vast amount of money every year on professional services, and a generation of AI companies is delivering meaningful value to customers by making that work faster, better, or available to people who could not previously afford it.

Ideas behave differently as an economic object. They accumulate into what economists call “stock”, artifacts that, once produced, persist and can be used again and again rather than being metered out and consumed. "Once the design is created, it can be used as often as desired, in as many productive activities as desired"[7]. A machine is durable too, but it is rival, as a machine-hour sold to one customer cannot be sold to another. An idea can be used by everyone at once without being used up. For example, novel cancer drugs, new manufacturing protocols, verified algorithms, or generative materials designs are artifacts whose value does not get exhausted by use and whose price is not bounded by the cost of the cognitive work that produced them. The price is set by what the resulting artifact is worth to the world, provided the artifact is partially excludable, through patents, trade secrets, or data no one else has. A unit of frontier compute pointed at producing such an artifact can return value many multiples larger than the same unit applied to cognitive work that is consumed when delivered.

There is already evidence of this happening. AI systems have predicted structures for nearly every known protein and discovered faster algorithms for problems mathematicians had worked on for decades. The compute behind those results was spent years ago and the world is still using what it produced. The companies that direct AI compute at producing this kind of artifact are using the AI's time to produce things the world will keep using long after the compute is spent.

What to build

A question for builders thinking about how to produce extraordinary value over the next decade could simplify to something like:

Whose business model uses AI compute to produce lasting artifacts of value, things like discoveries, designs, and recipes, that accumulate into a compounding asset, rather than only delivering cognitive work that is consumed when delivered?

The first wave of frontier AI businesses has reasonably gone after the opportunity closest to the existing economy in the form of cognitive labor that customers are already paying humans to do, where the value proposition is easy to articulate and the unit economics tie to existing budgets. This wave was always expected to happen, and it is generating real value for the customers it serves.

The wave I am significantly more interested in, and that the framework above is trying to identify, consists of companies built specifically to direct AI compute at producing artifacts that did not previously exist. A new drug, material, algorithm, or scientific breakthrough. In those cases the artifact is the company's main product. It is what its compute and effort are directly aimed at producing. These are companies treating frontier AI as a method of invention rather than as a faster means of cognitive labor.

The future

Economies grow because the accumulating stock of ideas produces increasing returns. Each new artifact added to the stock makes the next one easier to reach, as recipes combine with other recipes to produce further recipes, and the value of the stock grows exponentially as more is added. Adding to that stock has historically required human researchers working on long timescales in narrow domains. Frontier AI may be the first widely available input that can be directed at the production of new ideas at substantial scale. And because the input that produces new ideas is itself nonrival and getting better with each generation, the dynamic compounds on itself.

If even a modest fraction of the world's AI compute is directed at this kind of work, the coming decade will likely produce one of the largest expansions of humanity's stock of recipes in the historical record, larger in scale and on a more compressed timeline than any prior wave of invention has produced. The size of that expansion, and the speed of its arrival, is the thing that makes frontier AI fit neither the utility analogy nor the platform analogy, and unlike any of the technological transformations economists have studied so far. In the long run, the rate at which humanity’s stock of ideas grows sets the rate at which the economy grows.

Footnotes

  • [1] By Chamath Palihapitiya's recollection of a 2007 conversation. X post recalling Bill Gates's definition of a platform.
  • [2] Named by Adam Smith in An Inquiry into the Nature and Causes of the Wealth of Nations (1776).
  • [3] Alfred Marshall's Principles of Economics (1890) added a fourth factor he called "organization". The framing of entrepreneurship as bearing uncertainty and earning profit as the residual comes from Frank Knight's Risk, Uncertainty and Profit (1921).
  • [4] Paul Romer, "Endogenous Technological Change" (1990), which treats ideas, what he variously called "technology" or "designs," as a factor with distinct economic properties.
  • [5] Daron Acemoglu and Pascual Restrepo, "The Race between Man and Machine" (2018), which develops the task-based framework referenced here.
  • [6] Iain Cockburn, Rebecca Henderson, and Scott Stern made this observation in their 2018 NBER paper "The Impact of Artificial Intelligence on Innovation"
  • [7] Paul Romer, "Endogenous Technological Change" (1990)