The opening act

Every major technology transition follows a recognizable sequence, though it is rarely recognized in the moment.

First comes the enabling constraint: the scarce resource without which the new technology cannot scale. Capital rushes toward it. Prices reflect the scarcity. Fortunes are made by those who saw it early and positioned before the crowd.

Then the constraint eases. Production catches up. The scarce thing becomes abundant, then cheap. And the capital that flooded in to price the constraint suddenly finds that the story has moved on — up the stack, to whoever built something valuable during the window when the constraint was real.

The railroad required steel. Steel was the opening act. The main event was who used the railroad to build something that could not exist without it. The telegraph required copper wire strung across a continent. The main event was Reuters, the AP, the financial information networks that the wire made possible. The internet required fiber and bandwidth — and the companies that laid it mostly went bankrupt. The ones that built on top of it are now among the most valuable in history.

The pattern is consistent enough to be a framework, not just a metaphor.

The current constraint

Artificial intelligence, in its present phase, has a clear enabling constraint: compute and memory. High-bandwidth memory, advanced packaging, GPU capacity, the data center infrastructure to run it — all are in shortage. Demand is growing faster than supply can follow. The scarcity is real, and it is investable.

Capital has recognized this. Hardware and infrastructure have been repriced to reflect the opportunity. The businesses that supply the constraint are doing exactly what businesses do when demand runs ahead of supply: they earn extraordinary margins, attract capital, and expand.

None of that is wrong. But it describes the opening act.

What resolves

Constraints in hardware tend to resolve. Not quickly — semiconductor cycles are measured in years, not quarters — but they resolve. Capacity gets built. Competition increases. Margins normalize. The businesses that were extraordinary during the scarcity window become ordinary businesses serving a commodity market.

This is not a prediction about any specific company or cycle length. It is an observation about how technology markets work over time. The hardware that was scarce in one era of computing became cheap in the next. The same dynamic has played out in storage, networking, processing, and bandwidth. There is no obvious reason AI compute is structurally different, even if the cycle is longer than usual.

When the constraint eases, the question that matters is: who built something durable during the window?

The main event

The businesses that will define the AI era are not necessarily the ones supplying the current constraint. They are the ones using the constraint period to build positions that will persist after it resolves.

Three types of position tend to survive the rotation up the stack.

The first is ecosystem depth. Software vendors who have rebuilt their products around AI — not bolted on a feature, but genuinely reorganized their architecture, their data relationships, and their go-to-market around what AI makes possible — are creating a structural advantage that does not depend on hardware scarcity. When model access is cheap and widely available, the companies with the deepest integration and the most entrenched workflows will earn the margin. The model is a commodity; the ecosystem is not.

The second is data position. The businesses sitting at the data tier — those whose core product generates the proprietary, structured data that makes AI applications genuinely better — have a structural advantage that strengthens as model quality improves. A better model applied to richer data produces better outcomes. The data owner captures that value disproportionately.

The third is workflow ownership. Application-layer businesses that have successfully inserted AI into decisions their customers can no longer make as well without it have created a new form of switching cost. The question is not whether their AI feature is impressive. It is whether the customer's operation has reorganized around it.

Below the stack

The physical layer deserves separate treatment, because it is not simply the opening act of one technology cycle. It is a structural claim on something more durable than any single wave of compute demand.

The cost of intelligence is ultimately a cost of energy. Every model trained, every inference served, every data center humming through the night is consuming power. That demand is not going away when the current hardware cycle matures. It will grow as more of the economy becomes computational. And the supply of reliable, dispatchable power — particularly baseload generation that can be counted on regardless of weather or time of day — has been structurally underinvested for decades.

Energy infrastructure, unlike semiconductor hardware, cannot be scaled quickly. A new power plant takes years to permit, finance, and build. That illiquidity is a feature for the long-duration investor: it means the supply-demand imbalance cannot be arbitraged away in two or three years. The operators with contracted capacity and the ability to build ahead of demand are positioned for sustained pricing power that extends well beyond any single wave of AI spending.

The financial rails

There is a third layer to the AI transition that is less discussed but no less consequential: the financial infrastructure through which the AI economy will settle its transactions.

The internet economy required payment rails, and those rails became enormously valuable. The AI economy — in which software agents execute transactions, negotiate terms, and move value on behalf of humans and other agents — will require settlement infrastructure that is faster, more composable, and more programmable than what exists today. The candidate infrastructure is not a bank. It is the open, programmable financial networks that have been building toward this use case for years.

The networks that will matter are those with genuine liquidity, credible economic design, and the developer infrastructure to support machine-initiated commerce. The transition from human-mediated to machine-mediated finance does not happen overnight. But the infrastructure that supports it is being built now, and the networks that earn that role will be very difficult to displace.

Knowing which act you are in

The investor who prices only the opening act misses the main event. The investor who skips the opening act entirely — impatient for the application layer to emerge before the infrastructure constraint is resolved — arrives before the audience has gathered.

The discipline is not predicting exactly when the constraint resolves. It is maintaining a clear view of which act is currently playing and positioning accordingly: owning the constraint while it is real, watching who is building during the constraint window, and rotating toward the ecosystem builders before the rotation becomes obvious to everyone else.

The opening act is worth owning. It is not worth mistaking for the whole show.

— Shash Hegde