The AI Concentration Window: Why the Productivity Paradox May Shape Who Controls the Next Economy

Mark Howard

— 17 min read
Share

The most important question in the AI race may not be whether AI is powerful.

It is where we are in the cycle.

Are we still in the early phase of introduction? Are we in the moment when a new general-purpose technology is spreading faster than institutions can adapt? Are we in the period when expectations are outrunning measurable productivity? Are we in the window when power begins to consolidate around the firms, states, and institutions that control the missing complements?

My working answer is yes.

AI appears to be in the early introduction phase of a general-purpose technology. That does not mean its future is predetermined. It does not mean concentration is inevitable. But it does mean we should expect a period of instability between technological promise and broad social benefit.

That period matters.

It is not just a waiting room before productivity arrives. It is a political and economic window. During that window, the actors with capital, compute, data, cloud infrastructure, distribution channels, implementation expertise, and government relationships may be able to consolidate power before the broader economy, public institutions, and civic life fully catch up.

That is the AI concentration window.

The concept starts with three ideas.

First, information industries often move through cycles of openness, concentration, enclosure, and partial reopening.

Second, general-purpose technologies often produce a productivity paradox: technical capability and investment grow before broad measurable productivity gains appear.

Third, the productivity lag is not neutral. It can create pressure for control. Investors demand returns. Executives demand savings. Governments demand modernization. Institutions demand proof that the technology works. Under those pressures, powerful actors may centralize the systems needed to make AI productive.

That is the basic claim:

AI is not simply moving toward concentration by default. It is entering a concentration-prone window. The outcome will depend on whether public policy, public opinion, open technical alternatives, and civic institutions can build dispersion mechanisms before the early structure hardens.

This paper stops there. It does not yet offer a full policy agenda. Later papers can ask what public policy should do at different points in the cycle, how public opinion shapes whether concentration is accepted or resisted, and how tools like the Separations Principle can help preserve openness.

For now, the first task is to define the window.

1. What is the cycle?

Tim Wu’s The Master Switch offers a useful starting point. Wu argues that information industries tend to move through a recurring cycle: open experimentation gives way to commercialization; commercialization gives way to consolidation; consolidation becomes enclosure; enclosure eventually provokes pressure for reopening, disruption, antitrust action, or structural separation.

This is not a law of nature. It is not a clock. It does not say every communications technology must end in monopoly and then be reborn as an open system.

It is better understood as a pattern.

Information industries are prone to this pattern because they deal in networks, standards, access points, audiences, infrastructure, and control over communication. Once a firm controls enough of those layers, it can shape not only what it sells, but who can connect, who can build, who can reach an audience, and who gets heard.

Wu’s AT&T case is especially useful. AT&T was not a cartoon-like simple monopoly. For much of its history, its concentration was publicly legitimated. It was treated as a public-interest monopoly: reliable, orderly, universalizing, technically competent, and useful to the state. The Bell System produced real innovation. Bell Labs mattered. The system worked in many visible ways.

And yet it also suppressed innovation outside its own controlled structure. It controlled access. It controlled compatibility. It decided which innovations would move and which would wait. Wu’s AT&T case shows that concentration can be publicly justified, government-supported, and genuinely innovative while still becoming dangerous to democratic openness.

That lesson matters for AI.

If AI concentration develops, it may not look like crude private greed. It may be justified as safety. Or national competitiveness. Or reliability. Or the only way to build frontier systems. Or the only way to deliver AI at scale. Some of those claims may be partly true. That is what makes the problem hard.

The point is not that concentration is always fake or always harmful. The point is that concentration can become legitimate before people fully understand what has been lost.

Wu’s cycle also forces a useful correction to a simple “default concentration” story. If information industries move in cycles, then concentration is not necessarily the final state. It is a phase. It may be long. It may be durable. It may do real damage. But it can be challenged, reopened, or redirected.

That makes the central question sharper:

Where are we in the cycle, and what kinds of action matter at this phase?

For AI, the best answer is that we are early. We are in a buildout phase. The technology is being introduced, financed, embedded, marketed, feared, celebrated, and institutionalized all at once. Its long-term productivity effects remain uncertain. Its infrastructure is being built now. Its governance is lagging. Its public meaning is still forming.

That is exactly the kind of phase where concentration can accelerate.

2. Why information industries tend to concentrate

Wu’s cycle is not the only reason to worry about concentration. It connects to a broader body of research on information markets.

Network economics has long emphasized that information markets can tilt toward concentration because of network effects, switching costs, standards, compatibility, and lock-in. Carl Shapiro and Hal Varian’s Information Rules is one of the classic statements of this view. Their central insight is simple: in information markets, the value of a product or system often depends on how many others use it, what it is compatible with, and how costly it is to leave.

That is why standards matter. Interfaces matter. Defaults matter. Switching costs matter.

W. Brian Arthur’s work on increasing returns and path dependence adds another piece. When a technology becomes more valuable as more people adopt it, early advantages can harden. A path that begins as one option among many can become the path everyone must follow.

This is not destiny. Critics of strong lock-in arguments are right to warn that markets can reopen, alternatives can emerge, and not every early advantage becomes permanent. But the basic warning remains important: information markets are often not flat playing fields. They are prone to feedback loops.

AI has many of those features.

A model improves with more users, more data, more feedback, more distribution, more developer attention, more enterprise integration, and more capital. A cloud provider becomes more valuable as more firms build on it. A workplace AI tool becomes harder to leave once it is embedded in documents, calendars, customer records, workflows, prompts, agents, evaluations, and training materials.

Then there is architecture.

Lawrence Lessig’s work on code as a form of regulation reminds us that digital systems are governed not only by law, but by design. Code determines what users can do. Interfaces determine what is easy. Defaults determine what is normal. Technical architecture can open a system or close it.

Jonathan Zittrain makes a related argument about generativity. The early Internet was powerful partly because it allowed unexpected uses by many users and developers. But generative systems can be displaced by tethered appliances and centrally controlled applications. Convenience can come with control. Safety can come with permission. Smooth user experience can come with a narrower field of invention.

That is directly relevant to AI.

AI may be widely available at the surface while becoming more closed underneath. People may use AI tools every day while the deeper system becomes more dependent on a small number of firms that control models, cloud infrastructure, user interfaces, enterprise distribution, evaluation standards, and deployment pathways.

So the concentration question is not simply:

Who has the biggest model?

It is:

Who controls the layers through which AI is built, accessed, integrated, evaluated, governed, and monetized?

That layer question will return throughout this project.

3. What is the productivity paradox?

The productivity paradox is the gap between visible technological progress and delayed measurable productivity gains.

It is not unique to AI.

General-purpose technologies often take time to matter at scale. They are not simple tools that can be dropped into old systems and instantly produce transformation. Their full effects depend on complementary changes: new infrastructure, new skills, new organizational forms, new management practices, new business models, new standards, and new ways of measuring value.

This is the key lesson from Paul David and Gavin Wright’s work on electrification.

Electric power existed before the big productivity surge. Factories did not become dramatically more productive simply because electricity was available. At first, many firms used electric motors inside older factory arrangements. They layered the new technology onto the old system.

The larger gains came later, when factories were redesigned around electric power. Unit-drive motors replaced older shaft-and-belt systems. Factory layouts changed. Materials moved differently. Workflows became more flexible. Capital productivity improved. Labor markets, management practices, and workforce capabilities also mattered.

The lesson is plain:

A general-purpose technology becomes transformative only when the surrounding system is rebuilt around it.

Erik Brynjolfsson, Daniel Rock, and Chad Syverson apply this logic to AI. They argue that AI may be a general-purpose technology whose effects are not yet fully visible in productivity statistics because the necessary complements have not diffused widely. Their point is not that AI is fake. It is that AI’s full value requires complementary innovation: organizational redesign, training, new business processes, new datasets, new skills, adjustment costs, and intangible capital.

That idea is crucial.

It means access is not capacity.

A school can have access to an AI tool without knowing how to redesign teaching, assessment, privacy, procurement, staff training, and accountability around it. A local government can buy AI software without having the data systems, internal expertise, legal review, workflow analysis, or evaluation capacity to use it well. A small business can experiment with AI while still depending on the same large platforms for models, cloud, payments, search, advertising, distribution, and customer access.

AI may spread widely at the tool layer while remaining concentrated at the implementation layer.

That is the productivity paradox as a governance problem.

The standard productivity-paradox story says: be patient. Powerful technologies take time.

The concentration-window story says: be alert. The waiting period is when power may be reorganized.

4. How the productivity paradox creates a concentration window

The productivity lag creates pressure.

Companies spend heavily on AI. Investors expect returns. Executives promise transformation. Governments announce modernization. Consultants sell implementation strategies. Workers are told to adapt. Public agencies are told not to fall behind. Universities, schools, hospitals, courts, nonprofits, and local governments all face the same message: AI is coming, and you need to be ready.

But productivity does not arrive everywhere at once.

The early period is messy. Tools are impressive but unreliable. Pilots work in narrow settings but are hard to scale. Data systems are not ready. Workflows are unclear. Procurement is confusing. Staff need training. Legal and compliance risks appear. Customers may not trust the systems. Existing software does not integrate cleanly. Metrics are uncertain. Nobody quite knows what counts as success.

That is the paradox.

And in that paradox, powerful actors often reach for control.

They centralize data. They standardize workflows. They monitor labor more closely. They buy integrated platforms. They rely on outside vendors. They accept cloud lock-in. They restructure teams. They outsource implementation. They consolidate procurement. They adopt default tools inside already dominant software suites. They turn work into machine-readable processes.

Some of this may be necessary. Some of it may improve productivity. Some of it may even improve service quality.

But it can also concentrate power.

The concentration window works through several channels.

Capital discipline

AI is expensive. Frontier models, chips, data centers, cloud contracts, talent, and integration work all require major investment. When returns are delayed, pressure rises. Firms that can afford the delay gain time to learn. Firms that cannot may become customers, acquisition targets, or dependent users.

This favors actors with capital.

Compute and infrastructure control

AI runs on physical infrastructure: chips, data centers, energy, cooling, networks, and cloud systems. These are not evenly distributed. They are capital-intensive and politically contested. Control over compute can become control over what kinds of AI can be built and by whom.

This favors hyperscalers, frontier labs, infrastructure investors, and states.

Implementation capacity

The deepest gains from AI may depend less on access to a model than on the ability to reorganize around it. That means data readiness, workflow redesign, training, evaluation, procurement, legal review, cybersecurity, management adaptation, and institutional learning.

Those capabilities are unevenly distributed.

Large firms, elite institutions, and sophisticated public agencies may build them. Smaller firms, local governments, schools, clinics, nonprofits, and civic institutions may buy them from vendors. That spreads AI use, but not necessarily AI power.

Platform lock-in

If an organization lacks internal AI capacity, the easiest path is often a turnkey product from an existing vendor. That may be rational. It may also create dependency.

The organization “has AI,” but the vendor controls the model, updates, integrations, interface, pricing, data flows, roadmap, and exit conditions.

This is how diffusion can hide concentration.

Workflow control

AI often works best when work is made explicit, standardized, trackable, and machine-readable. That can create productivity. It can also expand managerial control. Tasks become data. Judgment becomes workflow. Workers become easier to monitor, compare, and replace.

This creates a labor side to the concentration window.

Layer-control bundling

The most important channel may be bundling.

A firm that controls the model, cloud hosting, enterprise software, user interface, data pipeline, app marketplace, evaluation framework, and implementation services does not merely sell a tool. It controls a stack.

That stack can be convenient. It can be powerful. It can also be hard to leave.

This is where Wu’s Separations Principle becomes useful. The key question is whether control is being combined or separated across layers. In AI, the layers include data, compute, cloud, models, applications, distribution, user interfaces, standards, evaluation, procurement, and governance.

A concentration window opens when the pressure to make AI productive encourages firms and governments to combine those layers under fewer actors.

5. Historical example: AT&T and the public-interest monopoly

AT&T is the cautionary case.

The Bell System shows how concentration can become respectable. AT&T was not merely tolerated. For a long period, it was treated as a public-interest institution. A single integrated telephone system promised reliability, universality, order, and technical excellence. Government did not simply stand outside the system. It helped legitimate it, depended on it, bargained with it, and eventually challenged it.

That matters because AI concentration may follow a similar legitimating path.

AI concentration may be defended as necessary for safety, national security, reliability, and global competitiveness. Governments may rely on dominant firms for defense, cybersecurity, public administration, health care, education, and economic productivity. The same state that may someday regulate concentrated AI may first depend on it.

AT&T also complicates the innovation story.

The Bell System produced real innovation. The problem was not that monopoly created no breakthroughs. The problem was that innovation became permissioned. AT&T could decide what fit the network, what fit the business model, what would be delayed, and what would not be allowed to disturb the system.

That is an important warning for AI.

A concentrated AI system may produce spectacular innovation inside the system while suppressing a broader innovation ecology outside it. We may see what the dominant firms choose to build, release, integrate, and permit. We may not see what would have emerged from a more open environment.

The lesson is not that all concentration is useless. The lesson is that public benefit and concentrated control can coexist for a long time. That coexistence makes the eventual democratic problem harder to see.

6. Historical example: electrification and the delayed payoff

Electrification shows why productivity delays matter.

Electric power was not enough. The productivity gains required redesign. Factories had to change. Workflows had to change. Management had to change. Labor markets and skills had to change. Complementary systems had to emerge.

That helps explain why AI may not immediately produce broad productivity gains even if the technology is real.

But the analogy also raises the power question.

Who can afford the redesign? Who can wait through the messy period? Who can absorb failed experiments? Who can hire the technical staff? Who can reorganize workflows? Who has enough data? Who can coordinate across departments, legal teams, procurement systems, and workers?

If the answer is mostly large firms and well-resourced institutions, then the productivity lag can increase concentration.

The electrification analogy tells us that the lag is normal. The AI concentration-window thesis adds that the lag is also political.

7. Historical example: the Internet’s mixed cycle

The Internet is a more complicated example. That is why it is useful.

It did not simply move from open to closed. It remains open in some layers and highly concentrated in others. Open protocols coexist with dominant platforms. Low-cost publishing coexists with search, social, cloud, app-store, advertising, payment, and device gatekeepers.

That is the mixed nature of the cycle.

The Internet lowered barriers to entry for many creators, firms, and communities. It also produced enormous concentrations of platform power. It dispersed publishing and concentrated discovery. It expanded participation and centralized advertising. It opened communication and created new interface gatekeepers.

This is likely to be true of AI as well.

AI may help small firms, nonprofits, workers, and local governments do things they could not do before. It may also deepen dependence on a few model providers, cloud platforms, software suites, app stores, search systems, payment rails, and distribution channels.

Both can be true.

That is why the race is not simply “AI: concentration or dispersion?” The real question is:

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

The Internet teaches us not to confuse user access with structural openness.

8. Where this leaves AI

AI is early. It is powerful. It is uneven. It is expensive at the frontier and cheap at the interface. It is spreading quickly, but deep implementation is still hard.

That combination is exactly what makes the current period important.

At the surface, AI looks highly dispersed. Millions of people can use chatbots, copilots, image tools, coding assistants, and search interfaces. Small teams can build products faster. Workers can draft, summarize, code, translate, analyze, and automate tasks that used to require more time or more people.

That is real.

But at deeper layers, AI may be concentrating.

The most important layers include:

  • compute;
  • cloud infrastructure;
  • data;
  • foundation models;
  • application ecosystems;
  • distribution channels;
  • user interfaces;
  • operating systems and devices;
  • app stores and agent marketplaces;
  • standards and benchmarks;
  • evaluation and auditing;
  • procurement and payment channels;
  • implementation expertise.

Recent AI market-structure research points in this direction. Work by Jai Vipra and Anton Korinek highlights concentration risks around foundation models, scale, compute, data, and vertical expansion. CEPR/VoxEU’s “cloud-model-data loop” describes how cloud resources, better models, data, and applications can reinforce one another. OECD work on AI infrastructure treats the AI supply chain as a layered competition problem involving chips, data centers, cloud providers, energy, networking, and models.

This is why the productivity paradox matters.

If broad productivity gains require complementary investments, then whoever controls the complements may shape the AI economy before productivity becomes broadly visible. That means the decisive terrain may not be only model capability. It may be implementation capacity.

Who can make AI useful in real institutions? Who can integrate it into workflows? Who can evaluate it? Who can govern it? Who can afford to experiment? Who controls the systems others must use?

That is where concentration may harden.

9. The window is not destiny

The concentration window is not a prediction of inevitable monopoly.

There are real counterforces.

Open-source and open-weight models may reopen parts of the stack. Smaller and more efficient models may reduce dependence on massive centralized infrastructure. On-device AI may move some capability away from cloud systems. Public compute could create alternative pathways. Interoperability rules could reduce lock-in. Procurement rules could protect public agencies from vendor dependence. Independent evaluation could separate the power to build AI from the power to validate it.

Public opinion also matters.

This is one weakness in Wu’s AT&T account and one place this project should improve on it. Public legitimacy can support concentration. The public may accept concentration when it is framed as safety, convenience, public service, national strength, or economic necessity. But public backlash can also change the cycle. Local fights over data centers, labor conflicts over automation, consumer distrust, school concerns, privacy worries, and frustration with opaque systems can all become forces in the cycle.

The question is whether public concern becomes public capacity.

Complaints alone do not disperse power. Protest alone does not create governance. Public opinion matters when it becomes institutions, rules, standards, procurement capacity, civic infrastructure, public-interest alternatives, and political pressure.

That is why the current phase is so important. The future structure of AI is still being built. The window is open, but not forever.

10. What this paper does not yet answer

This paper defines the concentration window. It does not yet tell us how to close it, redirect it, or use it.

That requires another layer of work.

Future papers should ask:

  • What policy tools make sense at different points in the concentration/dispersion cycle?
  • How can public institutions tell whether AI is in an open, consolidating, enclosed, publicly legitimated, backlash, or reopening phase?
  • How does public opinion shape whether concentration is seen as public service or public danger?
  • What would a dispersion-oriented response to the productivity paradox look like?
  • How can the Separations Principle guide procurement, competition policy, public compute, interoperability, and independent evaluation?
  • Which AI layers are most urgent to keep open, portable, contestable, or publicly governed?
  • What signals would show that AI is moving from a concentration window toward a dispersion window?

These are the next questions.

For now, the central claim is enough:

The early AI era is not neutral. It is the period when the complements are being built. If those complements are controlled by a narrow set of firms and state partners, AI may spread widely while power concentrates deeply. If those complements are made public, portable, interoperable, accountable, and broadly available, the same window could become a dispersion window.

That is the race.

Core sources for this paper

Network economics, lock-in, and increasing returns

Architecture, code, and generativity

General-purpose technologies and phased technological change

AI market structure and stack control

Qualifying and countervailing sources

Read more