The Eiffel State Building

We solved the hard part. That may be the problem.

For most of the history of software, the first question was simple: Can we build it? A serious system required programmers, architects, testers, security specialists, project managers, servers, capital, time and the patience to coordinate all of them. The difficulty of production acted as a filter. Most ideas never became products because building was expensive.

Generative AI changes that equation. A technically experienced person can now describe a system, ask an AI to construct it, test it, inspect it, harden it, extend it and sometimes ask another AI to critique the first. The human may never type the code personally. The job becomes closer to architecture, orchestration, judgment and verification than traditional line-by-line programming.

That is extraordinary. It is also the beginning of a new economic problem. If one person can do this, millions eventually can. The scarcity that protected each builder from millions of competitors begins to disappear.

AI can make production abundant. It cannot manufacture more hours in another human being’s day.

This is the Eiffel State Building problem. Give everyone the power to build a landmark and the skyline fills with landmarks. But the number of visitors does not multiply merely because construction became cheap.

The literature is beginning to describe exactly this.

The central idea is therefore not speculative in the loose sense. Several established theories point in the same direction, and current researchers are now applying them directly to generative AI. What remains uncertain is the magnitude: how fast builder saturation will occur, which markets will be most affected, and what new institutions will emerge in response.

Social networks are the cruelest example

A word processor can be valuable to one person. A private database can be valuable to one company. A social network is different. Its value depends heavily on who else is there. A technically superior network with three people may be less useful than an inferior network containing everyone you know. That is the logic of network effects and the cold-start problem.

So a creator can build something that is, feature for feature, “ten times better than Facebook” and still discover that the active population is the creator, a test account, a daughter and two patient friends. Nothing paradoxical happened. Technical quality and network utility are different variables.

A magnificent empty nightclub can still lose to the shabby club across the street—if everybody’s friends are across the street.

The builder may also stop being a consumer

There is an additional hypothesis worth testing. When people spend more of their leisure time building their own AI-generated systems, songs, games, feeds, videos and communities, that time is no longer available for consuming everybody else’s creations. AI could therefore increase supply while simultaneously consuming some of the demand.

If AI drives the denominator sharply upward while builders spend more time creating and less time consuming, the pressure becomes stronger still. This particular builder-as-consumer effect is a reasonable inference, not yet a verified population-scale conclusion.

How many wheels can we reinvent?

The absurdity becomes visible with a toy calculation. There are more than eight billion people on Earth. We do not need all of them building. Move the slider and imagine only a fraction independently receiving the same perfectly sensible instruction from their own AI: “Build me a simple community platform.”

The point is not that eighty million identical social networks will literally appear. The point is that production capacity can grow far faster than coordination capacity. The same duplication can happen at smaller scales inside companies: ten departments independently building ten AI assistants for the same workflow because each can now do so cheaply.

So how do we stop building eighty million wheels?

The solution is not to take AI away from builders. That would throw away one of the most useful productivity technologies we have developed. The solution is to make coordination as easy as creation.

We already know pieces of the answer. Open-source communities learned to coordinate large numbers of contributors through shared repositories, modular architecture, maintainers, issue trackers, standards and reuse. Research on open-source software has long emphasized that modularity and coordination allow many people to contribute without each recreating the entire system. More recent work on open AI ecosystems likewise describes pooling resources and reuse as ways to avoid duplicated effort.

The scarce resource after abundance

If implementation becomes cheap enough, value migrates upward. The important skills become choosing the right problem, discovering what already exists, establishing trust, attracting collaborators, integrating components, creating standards, building communities and deciding what not to produce.

That does not make human beings less important. It makes the particularly human parts—purpose, taste, judgment, social trust, responsibility and cooperation—more important. AI can produce an enormous number of technically plausible answers. Humans still have to decide which answers deserve a shared future.

The question after “Can everybody build an Eiffel State Building?” is not “How do we stop them?” It is “How do we help them discover that somebody else already poured the foundations—and might welcome a partner?”

There may eventually be AI agents consuming services from other AI agents, which would expand demand beyond human attention. But even that future creates another coordination problem: agents need protocols, identity, trust, discoverability and incentives to cooperate. Scarcity does not vanish. It moves.

A practical rule for the AI age

Before creating anything significant, ask five questions: Does it exist? Can I reuse it? Can I contribute to it? Can I interoperate with it? Who should I build it with? Only then ask the old question: How quickly can AI build it?

That is how we keep the extraordinary democratization of AI without turning the digital world into a skyline of empty Eiffel State Buildings—each sparkling beautifully, each maintained by its owner, and each wondering where all the people, jobs and money went.