The AI Concentration Window: Why We May Overreact to AI's First Shocks and Underreact to the Power Structures Forming Beneath Them

New technologies are hardest to understand when they first arrive.

They appear first as products, workplace experiments, and public controversies. Only later do they become infrastructure, routines, and dependencies.

That creates what we might call the AI time horizon problem.

People naturally judge a powerful technology by what they can see. What they see first are the shocks: job disruption, synthetic misinformation, and data centers placing new demands on land, power, and water.

Those concerns are real. They are also incomplete.

The first wave tells us that something important has arrived. It does not tell us where society will end up.

That distinction matters for The Race.

The central question of The Race is whether AI will concentrate power or disperse capability. But that question cannot be answered by looking only at the most visible disruptions of the present. The more consequential issue may be whether today’s temporary advantages are beginning to harden into tomorrow’s infrastructure.

AI may be overhyped in the short run and underestimated in the long run. Early products may disappoint. Productivity gains may be uneven. Many pilots will fail. Some companies will discover that announcing transformation is easier than producing it. This is not exactly a new corporate tradition.

Yet beneath the noise, more durable arrangements may already be forming.

A small number of firms may gain control over the infrastructure needed to build AI, the platforms through which organizations adopt it, and the interfaces through which people use it. Public agencies may build critical services around proprietary systems. Workers may be reorganized around machine-readable workflows. Standards and procurement practices may settle before most people recognize that they were choices.

The period between technological arrival and broad institutional adaptation is not simply a waiting room.

It is a concentration window.

It is also a window of contestability: a period in which the basic arrangements remain fluid enough to be shaped, separated, reopened, or governed differently.

The window may be longer than panic suggests.

It may be shorter than complacency assumes.

Why the First Wave Is So Easy to Misread

The best-known version of the time horizon problem is often called Amara’s Law, usually attributed to futurist Roy Amara: we tend to overestimate the effect of a technology in the short run and underestimate its effect in the long run.

The phrase is useful, though it is better understood as a warning about attention than as a law of nature.

In the short run, new technologies are easy to overstate because they are novel, vivid, and surrounded by hype. Their most dramatic examples travel fastest. A chatbot passes a professional exam. A company announces an AI-related layoff. A fake image goes viral.

These events are concrete. They can be shown, shared, feared, and debated.

Long-term transformation is harder to see. It is slower, more distributed, and often hidden inside institutions. It appears in procurement rules, software defaults, and utility investments.

It is rarely one dramatic event.

It is a thousand quiet decisions.

Public anxiety also has its own logic. Amy Orben describes the recurring pattern as the “Sisyphean cycle of technology panics”. The problem is not concern itself. A democratic society should be concerned about powerful technologies. The problem is concern that becomes too narrow or too immediate to guide useful institutional action.

Risk-perception research helps explain why AI’s first shocks feel so large. Paul Slovic’s work on the perception of risk shows that people respond not only to probability, but also to dread, lack of control, and unfamiliarity.

AI has all three. It is opaque. It can feel uncontrollable. It is being deployed by powerful institutions that many people do not fully trust.

Public anxiety may therefore be registering something important: decisions affecting people’s lives are moving into systems they cannot easily understand, challenge, or control.

The difficulty is that visible harm and structural transformation move on different clocks.

The first clock measures shock.

The second measures the gradual settlement of infrastructure, defaults, and power.

A sound democratic response has to watch both. Panic focuses only on the first. Complacency notices that some early promises disappoint and assumes the second clock has stopped.

It has not.

The Lag Is Where the System Gets Built

Some technologies improve a particular task. General-purpose technologies reshape many forms of activity.

Economists Timothy Bresnahan and Manuel Trajtenberg described general-purpose technologies as pervasive technologies that improve over time and generate complementary innovation across sectors. Elhanan Helpman and Trajtenberg later explored how growth based on general-purpose technologies depends on waves of improvement, diffusion, and adaptation rather than one moment of invention.

AI increasingly fits that description. It can be used across business, government, education, medicine, science, and ordinary personal work.

But a general-purpose technology does not transform society merely by being invented or widely available. Its larger effects appear as the surrounding system changes with it: infrastructure is built, organizations redesign work, and laws and standards adapt.

Electrification is the classic example. Electric motors existed before factories realized large productivity gains. Many firms initially installed them inside factories still organized around steam-era layouts. The larger gains came later, when production itself was reorganized around electric power.

Paul David’s comparison between the electric dynamo and the computer shows why transformative technologies can appear long before their effects become visible in economic statistics. The breakthrough exists, but the surrounding system has not caught up.

The computer age produced the same puzzle. Computers spread through business and government for years before the expected gains appeared at scale. The Federal Reserve Bank of Richmond’s discussion of the productivity paradox revisits Robert Solow’s famous observation that computers seemed to be everywhere except in the productivity statistics.

Erik Brynjolfsson, Daniel Rock, and Chad Syverson argue that AI may be following this pattern. Their work on the Productivity J-Curve explains why the investments needed to reorganize institutions can initially look like costs while measurable gains appear later.

The conventional lesson is patience. Powerful technologies take time.

The institutional lesson is to examine what fills that time.

Firms choose vendors. Agencies sign contracts. Workers are reorganized. These are not preliminary details surrounding the transformation.

They are the transformation taking institutional form.

Three Ways Concentration Forms

The complementary changes required to make AI productive are expensive, difficult, and unevenly distributed. At the same time, pressure to show results grows.

Investors expect returns. Executives promise transformation. Governments announce modernization.

Integrated systems and powerful vendors offer an attractive answer. They can supply infrastructure, expertise, and implementation support that many organizations cannot readily provide for themselves.

That may accelerate adoption. It may also consolidate control before the benefits of AI are broadly distributed.

Three mechanisms are especially important.

1. Control of scarce inputs

Frontier AI depends on capital, compute, energy, and specialized talent.

Actors that can afford a long period of experimentation gain time to learn. Those that cannot may become customers, acquisition targets, or dependent users.

Control over scarce inputs can therefore shape which forms of AI are built, by whom, and under what terms.

Scale is not inherently undemocratic. Some undertakings genuinely require it. The issue is whether infrastructure becomes a platform others can enter and bargain with—or a gate through which every later actor must pass.

2. Control of adoption pathways

Most organizations will not build AI systems from the ground up. They will adopt them through existing cloud providers, workplace platforms, and implementation partners.

That is often rational. It can also deepen dependency.

An organization may purchase an AI tool while lacking the expertise to evaluate it, integrate it, or bargain effectively with the vendor. The provider may then control the model, the interface, and the terms of exit.

Information markets are especially prone to these dynamics because of network effects, switching costs, and lock-in. Carl Shapiro and Hal Varian’s Information Rules remains a useful guide to these forces. W. Brian Arthur’s work on increasing returns and path dependence explains how an early advantage can reinforce itself as users and complementary services gather around one system.

The first familiar doorway can become the operating layer.

Systems often become durable not because anyone chose them as permanent infrastructure, but because everyone chose them as the easiest next step.

3. Control inside organizations

AI can give workers and institutions powerful new capabilities. It can also make work more explicit, standardized, and measurable.

Tasks become data. Judgment becomes workflow. Performance becomes easier to monitor and rank.

The worker may become more productive while losing discretion. A public agency may improve service while becoming dependent on a vendor it cannot independently evaluate. A smaller firm may gain access to advanced tools while becoming unable to leave the platform that supplies them.

This is where the Separations Principle becomes useful.

The central concern is not whether every part of the AI stack is controlled by a different actor. It is whether control over infrastructure, applications, and evaluation becomes bundled tightly enough that one actor can determine access, terms, and exit.

Taken together, these mechanisms show how a productivity lag can become a concentration window. Organizations are not merely waiting for AI to improve. They are choosing the institutional arrangements through which improvement will be delivered.

Public Value Does Not Prove Structural Openness

Tim Wu’s The Master Switch offers an important caution.

Wu describes a recurring pattern in information industries: open experimentation gives way to commercialization, consolidation, and eventual enclosure.

This is not a law of nature. It is a pattern worth remembering.

The AT&T case is useful because the Bell System was not a simple villain. A single integrated telephone system promised reliability, universality, and technical excellence. Bell Labs produced real innovation. Government depended on the system, protected it, and eventually challenged it.

The system worked in visible ways.

It also controlled access and compatibility.

Public value and concentrated control can coexist for a long time.

The internet provides a more recent version of the same ambiguity. Open protocols coexist with dominant platforms. Low-cost publishing coexists with concentrated control over search and discovery. Broad participation coexists with centralized cloud and advertising systems.

AI is likely to follow a similarly mixed pattern.

Millions of people can already use powerful tools. Small teams can build products faster. Workers can perform tasks that once required more time or more people.

That is real diffusion.

But control may still be concentrating at deeper layers: compute, platforms, and standards.

The useful question is therefore not whether AI is open or closed in the abstract.

It is:

At which layers is AI dispersing capability, and at which layers is it concentrating control?

Broad user access is evidence of diffusion. It is not proof of structural openness.

The Window of Contestability

The concentration window is not destiny.

Information industries can reopen. New technical architectures can lower barriers. Public policy can change market structure.

Open-weight models may reopen parts of the stack. Smaller models may reduce dependence on massive centralized infrastructure. Procurement and interoperability rules may make it easier for institutions to change providers.

But these possibilities do not realize themselves.

The window of contestability is the period after a powerful technology emerges but before its dominant platforms, standards, and institutional dependencies become difficult to reverse.

During this window, democratic institutions have more leverage.

Public policy can shape procurement. Workers can gain a voice in implementation. Communities can negotiate infrastructure impacts before commitments harden.

After the window narrows, action remains possible. But it becomes more expensive.

Once systems are embedded, changing them is costly. Once users depend on a platform, exit becomes difficult. Once public services are built around proprietary infrastructure, oversight weakens.

The challenge is not to oppose scale. Large systems will be necessary in many parts of AI.

The challenge is to build contestable scale: systems large enough to perform important functions but open enough to be challenged, reviewed, and redirected.

That requires more than consumer choice.

It requires institutional capacity to contest.

What The Race Should Watch

The Race is not a project about whether AI is simply good or bad.

It is a project about direction.

The early evidence will always be mixed. Some developments will expand access. Others will deepen dependency. Some public uses will build civic capacity. Others will outsource public judgment to private systems.

The task is not to predict the future with false precision.

It is to watch the hardening process while it is still happening.

Three questions matter most:

• Where is AI dispersing usable capability?
• Where is control accumulating across infrastructure, platforms, and institutions?
• Can affected actors review decisions, change providers, or build alternatives?

AI is moving on two clocks. The first measures visible disruption. The second measures the gradual settlement of defaults, dependencies, and power.

The danger is not only that we will fear AI too much.

It is that we will notice the deeper hardening too late.

If the infrastructure and platforms surrounding AI are controlled by a narrow set of firms and state partners, AI may spread widely while power concentrates deeply.

If those systems remain portable, accountable, and open to alternatives, the same period could become a dispersion window.

That is the race.

References and Further Reading

Time horizons, technology panics, and risk

• Patrick Lin, “Amara’s Law and Its Place in the Future of Tech,” IEEE Computer Society.
• Amy Orben, “The Sisyphean Cycle of Technology Panics,” Perspectives on Psychological Science 15, no. 5 (2020): 1143–1157.
• Paul Slovic, “Perception of Risk,” Science 236, no. 4799 (1987): 280–285.

General-purpose technologies and productivity lags

• Paul A. David, “The Dynamo and the Computer,” American Economic Review 80, no. 2 (1990): 355–361.
• Elhanan Helpman and Manuel Trajtenberg, “Growth Based on General Purpose Technologies,” NBER Working Paper No. 4854 (1994).
• Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “The Productivity J-Curve,” American Economic Journal: Macroeconomics 13, no. 1 (2021): 333–372.

Information industries and AI market structure

• Tim Wu, The Master Switch (2010).
• Carl Shapiro and Hal R. Varian, Information Rules (1998).
• OECD, Competition in Artificial Intelligence Infrastructure (2025).