Progress in artificial intelligence is usually treated as a software problem. Better algorithms, more data, larger models — these are the main terms of competition. Companies and research institutions compare their standing by model performance, and the center of gravity of the AI industry sits, for the most part, on computation and software.

But if performance rises past a certain level — especially if AGI arrives and begins improving itself — the differences between individual models may come to matter less than they do now.

Recursive self-improvement means that an AI can analyze and revise its own architecture, its training procedure, its use of tools. If several systems repeat that process with enough time and resources, they may converge on similar solutions and similar performance regions even when they start from different places. Much as generative models across a range of fields are now converging on the transformer architecture. Good algorithms will be rediscovered or copied quickly, and differentiation at the level of software will be hard to sustain.

In that case intelligence itself gradually stops being a scarce asset. Just as electricity and telecommunications networks today are less a proprietary capability of any one company than infrastructure common to nearly all economic activity, artificial intelligence above a certain level may become a universal factor of production.

But intelligence becoming universal does not mean power is distributed along with it. Turning intelligence into action still requires physical resources. An AI has to run on computing hardware, and that hardware requires electricity and cooling. Producing new scientific knowledge requires instruments and laboratories; producing physical goods requires factories, raw materials, and logistics. Acting in physical space requires robots, vehicles, drones, sensors.

So even if the intellectual capacities of different AI systems arrive at similar levels, the scale of physical resources available to them can differ enormously. The same AI running on a personal computer and running while connected to a large data center, production facilities, and energy infrastructure do not have the same effective capability. And this difference is more likely to strengthen concentration in the AI industry than to weaken it.

Software, as we have known it, is cheap to copy. One program can be distributed to many machines at little cost. That property gave new companies and individuals a chance to catch up quickly with incumbents’ technology. Open-source software and published research accelerated the trend further.

Semiconductor fabs, data centers, power plants, communication networks, and robot production lines, by contrast, are not easily copied. Building them takes large capital and long stretches of time, and also land, raw materials, specialized labor, and regulatory approval. Physical infrastructure moves less easily than software, and an advantage once established lasts longer.

Recursive self-improvement can widen this gap. Whoever holds more compute can run more improvement experiments. An improved AI can in turn optimize chip design, energy efficiency, manufacturing processes, logistics, financial decisions. The additional resources secured in that process go back into still more computation and still more experiments.

The result can be a cycle. More physical resources make stronger AI systems possible. Stronger AI systems make the acquisition and operation of resources more efficient. Higher efficiency drives further concentration of resources. In this structure the critical asset is not a particular model or algorithm. The critical asset is the gateway an AI must pass through in order to reach reality: semiconductor supply chains, power grids, data centers, sensor networks, logistics facilities, factories, research equipment, legal permits.

Monopoly in the age of AI may therefore take a different form than before. Past monopolies rested largely on asymmetries of information and knowledge — only certain firms held a given algorithm or dataset. But once intelligence and knowledge can be widely copied, the locus of monopoly can shift from information to the capacity to execute. What matters is less what you know than whether you can actually carry out what you know.

The same shift may show up in art and culture. Digital outputs — images, music, writing, video — are produced at ever lower cost. As the number of producible images increases, the scarcity of any individual image decreases. Actual space, material, bodies, time, and place, on the other hand, remain limited. Exhibition space cannot be copied without limit, and an event that occurs in a particular place cannot be repeated identically. A work made of matter has to be transported and stored, and an audience’s bodily experience is not the same as its representation on a screen. The more abundant digital production becomes, the more sharply these physical conditions may function as points of difference.

This does not mean that physical objects are inherently superior to digital ones. What matters is that the location of scarcity moves. What used to be scarce was the skill and capacity to make a complex image. What may become scarcer from here is not the ability to make the image but the ability to place it within some material, some site, some institution.

Something similar may happen in the post-AGI economy. If intellectual capability is cheap enough and widely available, the difference arises not in intelligence itself but in the resources intelligence can draw on. Less who can produce the better answer than who can use more power and more equipment, who controls production and distribution, who holds the authority to act in the world.

Seen this way, AGI does not dismantle existing structures of monopoly. It may instead lower the cost of knowledge and judgment while making the concentration of physical resources and executive authority more visible. When intelligence is scarce, those who have intelligence have power. Once intelligence is universal, power belongs to those who can connect that intelligence to resources, to bodies, to authority.

So the central problem of the age of AI is not only model performance. The more important question is whose infrastructure universal intelligence runs on, and whose purposes it serves when it intervenes in the world. The more abundant intelligence becomes, the less matter disappears. Matter remains, rather, as the condition intelligence must pass through in order to become real. And the power of whoever controls that condition may grow larger than before.