AI and the return of physical scarcity

Scarcity is the beginning of economics.

Capital is scarce. Talent is scarce. Energy is scarce. Time is scarce. Attention is scarce. Land is scarce. Trust is scarce. The investor's task is not merely to identify what is growing, but to identify what remains scarce as growth unfolds.

Every technological revolution appears, at first, to abolish a constraint. The railroad collapsed distance. Electricity extended the working day and reorganized production. The semiconductor reduced the cost of computation. The internet lowered the cost of distribution. Cloud computing made infrastructure more elastic. Artificial intelligence now promises to reduce the cost of cognition itself.

But history rarely abolishes scarcity.

It moves it.

When one constraint falls, another becomes more important. When distribution becomes cheap, attention becomes scarce. When software becomes easy to ship, customer trust becomes scarce. When capital floods an industry, disciplined capital allocation becomes scarce. When computation becomes more powerful, the power, chips, land, cooling, and infrastructure required to support that computation may become the new limiting factors.

The central question for investors is therefore not simply who benefits from AI. It is where scarcity moves.

Chokepoints and value

In complex systems, value often migrates toward the chokepoint. A chokepoint is not always the largest market. It is not always the most visible part of the value chain. It is the constraint without which the rest of the system cannot expand.

Sometimes the chokepoint is a physical asset. Sometimes it is a network, distribution, regulation, brand, trust, manufacturing skill, energy access, or accumulated know-how. The form changes. The principle does not.

The history of technology is filled with these shifts. The California gold rush made some miners rich, but the sellers of tools, land, transport, finance, and logistics often held more reliable positions. The automobile revolution did not merely create car companies; it reshaped steel, oil, rubber, roads, suburbs, insurance, and consumer finance. The internet created software winners, but also required fiber networks, data centers, semiconductors, payment rails, logistics, and cloud platforms.

The obvious beneficiary is not always the best investment. Sometimes the better opportunity lies one layer away from the narrative — especially when markets chase the visible winner while underpricing the constraint that allows the winner to scale.

The AI narrative and the physical world

AI is often discussed as though it is weightless. Models are described in terms of intelligence, reasoning, agents, assistants, automation, creativity, and productivity. These are digital outcomes. But beneath them sits a deeply physical stack.

AI requires chips. Chips require advanced semiconductor manufacturing, equipment, materials, packaging, memory, substrates, and precision logistics. Data centers require land, power, cooling, interconnection, transformers, switchgear, backup systems, permits, construction labor, and grid access.

The models may feel immaterial. The infrastructure is not.

This is the paradox of AI: the more intelligence becomes abundant, the more valuable certain forms of physical scarcity may become. The marginal cost of generating an answer may fall. The cost of building the infrastructure that allows billions of answers to be generated may rise. The user may see intelligence on demand. The system behind it may be constrained by megawatts, memory bandwidth, transmission lines, and construction timelines.

The lesson of Jevons

There is an old economic lesson hiding inside the AI boom.

In the nineteenth century, William Stanley Jevons observed that improvements in the efficiency of coal use did not necessarily reduce coal consumption. Greater efficiency made coal more useful, which expanded demand. This became known as the Jevons paradox.

AI may have a similar structure. If models become more efficient, the cost of intelligence falls. But cheaper intelligence may increase the number of use cases. More agents, more workflows, more inference, more personalization, more simulation, more automation, more synthetic data, and more machine-to-machine activity may expand total demand for compute.

Efficiency does not automatically reduce resource use. It can increase it.

That does not mean demand grows without limit. Economics still matters. Capital costs matter. Power prices matter. Model efficiency matters. Regulation matters. Customer willingness to pay matters. But the simple assumption that better chips or better algorithms will make physical constraints disappear may be wrong. Improvement can expand the frontier — and an expanded frontier often demands more of the scarce input.

The capital cycle

Scarcity is not permanent merely because demand is strong. The capital cycle matters.

High returns attract capital. Capital creates capacity. Capacity can erode returns. This is the old pattern in shipping, energy, memory, railroads, telecom, and real estate. A shortage invites investment. Investment eventually creates supply. Supply can turn scarcity into excess.

History reinforces the warning. Railroads transformed economies, yet many railroad investors lost money because too much capital was deployed into competing lines. The internet transformed the world, yet many dot-com companies failed because markets overpaid for possibility before durable economics were proven. Fiber networks were overbuilt in the telecom boom, yet the cheap bandwidth left behind helped enable the next generation of internet businesses. Semiconductor cycles have repeatedly shown that real demand can coexist with brutal supply gluts.

The pattern is familiar. A technology is real. The investment cycle becomes excessive. Many participants disappoint. A smaller number of durable winners emerge.

The AI infrastructure boom will not be immune to this. If too much capital pursues the same layer of the stack, the economics of that layer can deteriorate even while AI adoption continues. Demand can be real and investors can still overpay. Revenue can grow and returns on capital can fall. A theme can be correct and the investment can be poor.

This is why AI is not a single trade. It is a force that may alter where scarcity sits. But each opportunity must still be judged by durability, capital intensity, competitive structure, reinvestment economics, and price.

Where scarcity may reside

The AI landscape contains several potential chokepoints.

The most visible is compute. Advanced accelerators, memory bandwidth, networking, and packaging have become central to the scaling of frontier models. But compute itself is not one thing. The bottleneck may sit in GPUs, high-bandwidth memory, advanced packaging, foundry capacity, networking fabric, or software that improves utilization.

The next chokepoint is power. Data centers are increasingly constrained by the availability, reliability, and timing of electricity. A site with secured power can be more valuable than a site with land alone. Grid interconnection, transmission capacity, transformers, cooling, and backup generation are no longer mundane utilities. They are strategic inputs.

A third chokepoint is time. Permits, grid connections, construction timelines, equipment lead times, and supply-chain coordination all matter. In a fast-moving technology cycle, the ability to bring capacity online sooner can itself be a competitive advantage.

A fourth chokepoint is trust. As AI-generated content expands, trust, verification, provenance, security, and enterprise reliability may become more important. If intelligence becomes abundant, trusted intelligence may remain scarce.

A fifth chokepoint is distribution. Models may become more capable, but access to customers, workflows, data, and decision rights may determine who captures value. The firm already embedded in the customer's process may be better positioned than the technically superior tool that lacks distribution. Software has not disappeared. Customer relationships still matter. In many cases, the most valuable businesses may emerge where AI improves the economics of existing software platforms rather than replacing them.

The scarcity map is broader than chips. The AI stack is physical, digital, institutional, and behavioral at once.

Roundabout capital

Greater productivity often requires indirect investment: tools before output, infrastructure before scale, capital goods before consumer abundance. AI is a roundabout technology. Before intelligence becomes cheap and ubiquitous, capital must be poured into the layers that produce it — chips, fabs, energy systems, data centers, fiber, cooling, and software tools. The visible output may be an answer on a screen. The invisible input is an enormous capital structure.

This creates both opportunity and danger. The opportunity is that certain capital goods may become indispensable to the new production system. The danger is that capital goods industries are often cyclical, competitive, and vulnerable to overbuilding. A roundabout path can produce extraordinary results — but only when the capital is allocated intelligently.

The return of the real

For decades, investors became accustomed to the idea that the best businesses were asset-light. Software scaled without factories. Platforms scaled without inventory. Networks scaled with low marginal cost. The market learned to reward high gross margins, low capital intensity, and global distribution.

Those lessons remain important. But AI complicates them.

The next phase of digital abundance may require a deeper physical foundation than many investors expected. Intelligence may be delivered through software, but it is produced through energy, silicon, land, cooling, grid capacity, and capital expenditure. The world does not become less physical because intelligence becomes digital. In some ways, it becomes more physical. The scarce inputs simply become easier to overlook because the user never sees them.

The question is therefore not whether AI will matter. It almost certainly will. The question is where that importance becomes durable economic value — and at what price the market is asking investors to own it.