The Economics of Abundant Intelligence

The most interesting question about artificial intelligence may turn out to be an old question about economics: what happens when something useful becomes much less expensive?

The first-order answer is that we use more of it. The harder question is where the benefits go. Lower production costs can mean lower prices, higher profits, new demand, or entirely new kinds of work. Often, they mean several of these at once.

This is the question at the center of Compute & Capital. Not a claim that every part of the transition is already understood, but a framework for deciding what to look at next.

Intelligence has historically been expensive

For much of economic life, getting more cognitive work done has meant finding more people, training them, coordinating their work, and paying for their time. A company that needed more software or analysis generally needed additional skilled labor, better tools, or both.

The distinction matters. A tool can make a worker more productive without replacing the worker. It can also change which tasks the worker performs. Neither result, by itself, tells us what happens to employment, wages, or the value of the resulting product.

So the unit of analysis should not begin and end with a job title. It should begin with the task, the cost of performing it reliably, and the organization around it.

Price is not the same as useful output

A cheaper model call is not automatically cheaper completed work. Useful output also requires verification, integration, access to information, and someone accountable for the result.

An agent that produces ten drafts is not necessarily more valuable than one that produces a single dependable answer. The relevant comparison includes the time spent correcting mistakes and the cost of failures that are not caught.

The economic measure is cost per reliable outcome, not cost per token.

That framing makes the thesis testable. If the surrounding costs refuse to fall, abundant model output may coexist with expensive real-world work. If those costs fall too, the consequences could be much broader.

What becomes scarce?

Some complements are physical: electricity, chips, network connections, buildings, and the ability to bring infrastructure into service. Others are institutional: permission, distribution, trust, proprietary context, and the ability to coordinate a complicated project.

But a bottleneck is not automatically a good investment. High prices can attract new supply. A scarce asset may be expensive to maintain. A business can be strategically necessary and still have weak pricing power.

The useful question is therefore more specific: which constraints persist long enough, and under what ownership and competitive structure, to create durable value?¹

Who captures the value?

The producer of a powerful technology does not necessarily capture all of its economic value. Customers may retain the gains through lower prices. Workers may use the technology to become more productive. Owners of complementary assets may gain bargaining power.

Competition determines how much of a cost reduction becomes a margin improvement and how much is passed through. Contracts, regulation, switching costs, and the speed of new entry all influence that division.

This is why a claim about technical capability is not yet a claim about investment returns. The bridge between the two is a theory of market structure, together with evidence that can distinguish that theory from a persuasive narrative.

The physical world still matters

Software can be reproduced quickly. A physical project has a different sequence: land, approvals, equipment, construction, commissioning, and operations. Each stage carries its own dependencies.

For compute infrastructure, proximity is not the same as availability. A transmission line near a parcel is not a promise of deliverable power. A carrier in a region is not proof of two physically independent fiber routes to a building.

These distinctions are economically consequential. The value of a proposed site depends on the time and cost required to turn an apparent advantage into something usable. The difference between a map and an executable project can be the difference between an attractive story and an investable asset.

What would change this view?

The framework should remain open to contrary evidence. Perhaps reliability costs dominate. Perhaps adoption is slower than expected. Perhaps supply responds so quickly that the apparent bottlenecks disappear. Perhaps human and machine capabilities prove more complementary than substitutable.

The work is to track the transition without confusing its direction with its speed, or its potential with its realized economics. This publication begins with that distinction.