The Industrialization of Cognition

Picture a company deciding whether to enter a new market.

A few years ago, that decision might have centered on a small team: one person gathering competitor research, another interviewing customers, someone building the model, someone writing the recommendation, and a manager deciding whether the argument was strong enough to act on.

Now the same work can be split across a chain of tools and review steps. One system gathers competitor pages and customer notes. Another drafts the market map. A third checks the recommendation against the company’s investment criteria. A human edits the argument, reviews the evidence, and approves the final call.

The final memo may still look like one person’s analysis. But the work of thinking through the decision has been broken apart: finding evidence, comparing options, drafting claims, testing the argument, checking the risks, and deciding what deserves human judgment.

Those pieces now move through models, data sources, routing rules, evaluation systems, permissions, budgets, and human approval gates.

The company may still make the decision. But the capacity to research, synthesize, test, and justify that decision now depends on systems someone has to own, rent, govern, or trust.

That is the industrialization of cognition: thought work becoming less like a private act of analysis and more like a production process run through infrastructure.

My hypothesis is that AI is not just another productivity tool. It is becoming a new production system for work that used to depend mostly on human judgment, writing, coordination, and analysis.

Some parts of that system are becoming cheap and widely accessible. Others behave more like productive capital: assets that can be accumulated, controlled, and used to generate output. Once that happens, the central issue is not only what AI can do. It is who owns the systems through which AI does work, and who captures the value those systems help create.

That is why this is a political-economic question, not only a technical one. AI is turning cognition into something closer to industrial infrastructure, and infrastructure is never only about usefulness. It sets access, dependency, bargaining power, and the conditions under which work gets done.

Do we value human skill anymore?

When I read things online, I skim. Rarely do I stop and really take in what I am reading. I almost expect every single piece of text online to be gen-slop. We live in a time where the value of a word feels lower than it should. Seems like everybody and their mother has a Substack producing AI-generated thought pieces on contemporary society.

My little cousin loves drawing, painting, sculpture, and who knows what else. But she would never pursue art professionally, or otherwise, beyond a mere recreational pastime. She told me “art does not pay,” so as she thinks ahead to college applications, she “knows [she has] to major in STEM.”

Those are the words of someone who spends a significant portion of her time self-expressing through art. I wonder what her attitude tells us about other 16 year olds across the country.

Do we value craftsmanship? The intuition of a good teacher? The well-worn emotional center of a therapist’s mind after they have talked to people about their feelings for 30 years?

It seems like we are confused about which skills our economy actually rewards. Classical industrial automation reduced the economy’s reliance on some forms of physical labor and helped expand the managerial class. As that happened, more work moved one step away from making the thing itself and toward coordinating, modeling, selling, packaging, and presenting it. Some of that work is real and valuable. But it can also make the craft feel distant from the reward. The anxiety comes from that distance: when value is created through layers of abstraction, everyone has to keep proving that their layer still matters.

My read is that AI intensifies this crisis because it changes the underlying mechanics of our relationship to production. It does not only change which tasks are easy or hard. It changes who controls the assets that make work possible. If the assets that make a worker productive sit inside someone else’s model, cloud account, memory layer, evaluation system, or distribution platform, then the worker may become more productive while more of the power over production moves to whoever owns the infrastructure.

I am not arguing for some romantic return to artisan life or industrial labor. The point is not that painters’ studios were economically fair, that factories were morally pure, or that old industrial jobs are coming back. The point is that productivity can rise while the human role in production shrinks or changes. Manufacturing productivity data already shows how output per worker can become detached from the number of people directly involved in making things.¹

To understand why, it helps to start with a simple economic model.

Production functions, in plain language

A production function is a way of describing how inputs become output. For any given set of inputs, there is some calculable maximum output a firm can produce.

The mathematical model version I saw as an undergrad taking Econ 100 looked something like this:

Y = F(K, L, A)

where:

• Y = output
• K = capital
• L = labor
• A = technology/productivity

My hypothesis is that AI changes the production function because it changes all the inputs at once. It can look like a tool to the worker, a machine on the firm’s balance sheet, capital to the economist, and infrastructure to whoever controls the workflow.

If that is right, AI will continue to influence the production function such that:

1. the relationship between inputs and outputs changes
2. the ownership and control of the means of production change
3. the institutional structure of “the economy” itself changes

In the past, when we have talked about the “means of production,” we have meant the productive assets society relies on: land, machinery, factories, logistics systems, energy, and so on. But in 2026, AI forces us to consider a new category of input that encompasses things like models, compute infrastructure, algorithmic decision-making, and autonomous coordination. Recent AI infrastructure research makes the same point from another direction: data, compute, energy, model development cost, and organizational structure are now central to how foundation models are built and controlled.²

I personally believe that today’s production function might eventually need to make the AI production system more explicit:

Y = F(K, L, AI, D, C)

where:

• AI = model capability, orchestration, and evaluation systems
• D = proprietary data
• C = compute infrastructure

These inputs are not perfectly separate economic categories. Compute is already a form of capital, and model capability overlaps with what economists would ordinarily call technology or productivity. The point of the formula is to make the emerging AI production system more visible.

It is also to ask a distributional question. A production function is not only a technical map of output. It is a way of asking which inputs create value, which inputs are scarce, and who can claim the surplus once that value is created. If AI changes the inputs, then it also changes the political economy around the output: who gets paid, who pays rent, who owns the bottleneck, and who has leverage when the productive system works.

That helps explain why some firms may become huge, resilient, and hard to compete with while others remain thin wrappers on top of infrastructure they do not control. A consulting shop, media company, or small software team may look newly powerful because it can produce with fewer people. But if the work runs through rented models, rented memory, rented compute, and rented distribution, then some of that power belongs to the infrastructure owner before the work even begins.

The split is already visible. Compute, training infrastructure, and frontier-model development are capital-intensive, while applications and workflows built on top of them may become cheaper and more widely available. Stanford’s 2026 AI Index Report found that global AI compute capacity had grown 3.3 times per year since 2022, reaching 17.1 million H100-equivalents, a standardized estimate of capacity expressed as the equivalent number of Nvidia H100 chips, with Nvidia accounting for more than 60 percent of the total. At the same time, Stanford’s 2025 report found that the cost of querying a model performing at the level of GPT-3.5 fell more than 280-fold between November 2022 and October 2024.³

Between the scarce infrastructure layer and the cheap application layer sit large datasets, foundation models, memory and context systems, evaluation layers, orchestration tools, and distribution platforms. Each layer has different economics. Compute may reward scale because the fixed costs are high. Data may reward incumbency because the best data often comes from existing users and workflows. Distribution may reward aggregation because fragmented producers often need access to the same concentrated channels for customers, attention, identity, payments, or workflow placement. The customer relationship becomes the choke point.

This is why concentration and diffusion have to be analyzed together. The same technology can let millions of people build useful applications while also making a smaller number of infrastructure owners more powerful. Diffusion at the edge does not cancel concentration at the core. Sometimes it depends on it. When ownership of core assets is highly concentrated, their owners may gain disproportionate leverage relative to everyone else over production. They may also claim a larger share of the value created by other people’s work, because the work has to pass through assets they control.

The concentration of productive assets is nothing new. Look no further than land concentration in feudalism, factory concentration in industrial capitalism, or platform concentration in today’s digital capitalism. The current backdrop makes the problem sharper: historically high wealth inequality, rising concentration at the top, and decades of policy that have often favored capital mobility, privatization, deregulation, and weaker public constraints on private accumulation. Even IMF economists have argued that parts of that neoliberal agenda have carried prominent inequality costs, and the World Inequality Lab has documented how wealth growth has been captured disproportionately by the very top.⁴ AI does not introduce concentration into an equal economy. It arrives inside an economy already organized around asset ownership, bargaining asymmetry, and scale.

Does AI complement or substitute for human labor?

Classical industrial automation reduced the economy’s reliance on some forms of physical labor and helped expand the managerial class. When new technology arises, we create new jobs accordingly.

My friends love to talk about how work has seemingly become meaningless. Through spreadsheets and presentation decks, we create abstractions of value; abstractions that AI can increasingly help create. Right now, specialized humans still often outperform AI when it comes to tasks of cognitive labor, coordination labor, managerial labor, creative labor, and even, sometimes, analytical labor. But AI systems keep improving.

So what happens if AI can do more and more of the tasks of work faster, cheaper, and more effectively than humans can?

In that scenario, the marginal productivity relationship between labor and capital changes.

The results may be similar to what we saw with industrial mechanization during the Industrial Revolution and beyond. Demand for workers who experience increased labor productivity as a result of the new technology might increase, while demand for workers who do not experience increased labor productivity might decrease. We might even observe so-called “superstar” effects where a small number of highly AI-leveraged individuals can outperform entire teams.

The risk I am describing is a polarized, winner-take-most market where many knowledge workers lose bargaining power. In practice, that means fewer workers can credibly say, “you need my full labor to get this output.” They may have less leverage over wages, promotion, autonomy, and pace because a firm can ask a smaller group of AI-leveraged workers to produce what previously required a larger team. This is not just a future worry. BLS researchers have already shown that labor’s share of income declined in many U.S. industries from 1987 to 2015, with especially large declines in some information industries such as software publishing.⁵

But that does not mean human skill disappears. It moves from producing every memo, model, deck, design, or analysis by hand to defining the work, evaluating the machine’s output, and understanding the system well enough to take responsibility for the result.

When cognition becomes infrastructure

At least since the early modern print era, production has depended on some combination of labor, capital, technology, energy, and organization. The printing press did not just make books cheaper. It changed how knowledge could be stored, copied, distributed, and standardized. AI is doing something similar to cognitive work when it is harnessed effectively. The combination changes once parts of intelligence itself become automatable: coordination, cognition, analysis, creativity, and the managerial work of turning ambiguity into action.

For some firms and forms of production, the binding constraint may increasingly move away from headcount alone. It may be compute access, energy, data, organizational agility, distribution, and more. In that world, leaner labor forces may coordinate larger productive systems while capital intensity goes up.

By capital intensity, I mean the degree to which production depends on capital assets rather than labor. In a factory, those assets might be machines, buildings, robotics, inventory systems, and energy infrastructure. In an AI-mediated firm, they might be model access, GPU clusters, proprietary data, workflow software, memory systems, evaluation harnesses, and the cloud infrastructure that ties them together. If AI makes cognitive work more dependent on those assets, then cognition itself starts to look more industrial.

Cognitive work becomes less like one person privately thinking through a task from beginning to end and more like a production line for thought. A messy goal gets broken into smaller steps. Some steps are handed to models. Some are checked by evaluators. Some are routed to humans. The whole process runs through software, logs, permissions, budgets, and review rules.

That is the mechanism I mean by the industrialization of cognition.

The new human skill stack

If cognition becomes more automatable, the scarce human skill may shift from doing every unit of cognitive labor ourselves to specifying, evaluating, governing, and improving systems that perform cognitive labor.

That was not the old skill stack. For most knowledge workers, the previous skill stack was closer to domain knowledge, writing, analysis, taste, communication, project management, and institutional memory. You knew the field, produced the memo, built the model, ran the meeting, made the deck, and carried context in your head. Those skills still matter. But when AI enters the work, they are no longer enough by themselves.

This is where the human role becomes more precise, not less important.

The new skill stack is not generic “AI fluency.” It is not just knowing how to use chat interfaces. It is the ability to translate messy goals into precise instructions, decompose work into bounded tasks, judge whether outputs are actually correct, notice failure patterns, structure the context a system can use, and decide where human review needs to stay in the loop.

In other words, the most valuable human work may move toward operating judgment: specification precision, evaluation, context architecture, trust boundaries, and cost judgment. These are not soft add-ons to the system. They are part of what makes an AI-supported workflow productive at all.

Take that market memo. Specification is naming the actual decision the memo should support, not just asking for “market research”: entering a market, buying a company, hiring a sales lead, or killing a project. Evaluation means checking whether the claims are sourced, current, and relevant: whether the system used last year’s market size, missed the competitor that launched last month, or treated a founder’s blog post like neutral evidence. Context architecture means deciding which customer notes, financial assumptions, competitor pages, and internal plans the system should see. Trust boundaries mean deciding which parts can be automated and which claims need human review before they leave the building. Cost judgment means knowing when another model call, another search, or another human review pass is worth it. Without those choices, the AI may still produce fluent text, but the workflow is not reliable production. It is just accelerated guessing with a better interface.

That is why the industrialization of cognition is not only a labor-substitution story. It is also a systems-design story. The central question is not simply whether AI can perform a task. It is who defines the task, who checks the result, who owns the context, who pays for the compute, who absorbs the error, and who gets the surplus.

The decline of firms’ transaction costs

Ronald Coase argued that firms exist because markets have transaction costs: the costs of finding people, coordinating work, negotiating agreements, monitoring performance, and enforcing expectations.⁶

If AI brings some of those costs down, including search, scheduling, contracting, knowledge retrieval, and operational monitoring, we may see firms restructure in two major, divergent ways.

One possibility is the hyper-concentrated mega-firm whose AI systems scale with data and compute. Another is the highly decentralized micro-firm that uses AI to do things that used to require larger organizations with more firepower.

I think it is likely that we will see the continuation of concentration at infrastructure layers and decentralization at application layers. These are not contradictory outcomes. AI capabilities can become cheaper and more accessible to use while the infrastructure beneath them remains difficult and expensive to own.

That split matters. A tiny team may be able to build and coordinate more than ever before, but only by renting access to infrastructure it does not own. The open question is whether today’s concentration is a temporary feature of an early market or a lasting structural condition.

A five-person firm can use AI to run customer support, generate sales material, write code, analyze churn, and coordinate contractors. That looks like decentralization, and in one sense it is. But if the firm’s memory, workflows, user data, model access, and distribution all pass through a handful of vendors, then the firm is small in headcount but not independent in production. It has escaped the payroll and moved into the platform bill.

AI and the composition of capital

The composition of capital is a way of asking what production is made of. How much of the productive system is living labor, and how much is accumulated capital: machines, buildings, software, infrastructure, intellectual property, and organizational systems? Marx described capitalism as tending toward rising capital intensity: more production flowing through machines and infrastructure relative to labor.

My concern is that AI may accelerate that tendency. If firms can produce more with smaller labor forces, labor may receive a reduced share of the income resulting from production. Labor share data is already one way economists track how much economic output accrues to workers as compensation.⁵ If AI pushes more production through capital-owned systems, the distribution question gets harder, not easier.

The surplus does not simply appear in the economy. It gets claimed: by workers as wages, by customers as lower prices, by firms as margins, by infrastructure owners as platform fees and cloud bills, by shareholders as equity value, or by the state as public revenue. The claim does not always follow the visible work. If a worker does less of the direct production because more of the work flows through rented models, data, compute, and distribution, then the worker’s share of the surplus may not remain commensurate with the value of the final good or service.

Market structure matters here. In a highly competitive market, some AI-enabled productivity gains may be competed away into lower prices. In a concentrated or monopoly-like market, the owner of a scarce asset can keep more of the surplus as margin, rent, or equity value. Supply also does not perfectly track demand: compute, energy, data access, model capacity, distribution, and organizational trust can all become constraints. That is why the composition of capital is not only a technical question about what production uses. It is a political-economic question about who can control scarce inputs and capture the value created through them.

Nonlinear productivity growth

AI is already helping improve the process of creating more, better AI. Anthropic has described Claude as increasingly involved in its own development: reviewing code before it merges, catching defects humans missed, running bounded research experiments, and improving the speed of model-development workflows.⁷ That does not mean the loop is fully closed. Hardware, energy, data, research judgment, and organizational coordination still matter. But the direction is no longer merely speculative.

If AI helps build better AI, the productive effects become recursive in nature. Instead of technology improving only through human labor applied from the outside, technology increasingly contributes directly to its own advancement. That is where nonlinear productivity growth becomes plausible: the tool improves the work, then helps improve the next generation of the tool.

The ownership risk is already legible. One structural outcome might be the birth of new economic regimes, a sort of platform neo-feudalism. I mean the phrase as an analogy for dependency, not a literal return to feudalism. When a few firms own the compute, models, data, identity systems, and economic coordination layers, everybody else risks becoming a tenant, a dependent, in a proprietary ecosystem.

Long live open-source AI and public infrastructure. Not because openness solves every problem, but because the infrastructure of cognition has to remain contestable if people are going to do more than rent their own capacity to think.

The distribution question

Now comes the politics of it all.

If labor receives a smaller share of production income, the AI-generated surplus still has to go somewhere.

If this trajectory holds, ownership structures become the key political issue in our economic restructuring. That is where the state should, and must, step in to have a hand in the path our economy takes. If consumption systems detach from wages, then we may need some sort of universal income, universal job guarantee, public compute, data trusts, procurement rules, labor institutions, or another structure for distributing productive surplus.

Maybe the current market structures will be preserved, and inequality will rise sharply. Maybe new public infrastructure, ownership models, or labor institutions will emerge. I do not pretend to know which path wins.

Who owns the infrastructure of cognition?

All of this leads back to the central issue: who will own the new means of production?

Speaking based on our history as a society, when a new foundational means of production emerges, whether it is agriculture, factories, electricity, or computation, societies and their economies eventually have to reorganize around that productive infrastructure. These technologies did not produce identical ownership patterns. That is precisely why the structure of the AI production system matters.

AI is forcing that question again. If cognition itself becomes industrialized, then ownership of the infrastructure of cognition becomes one of the central economic and political questions of our time.

That infrastructure is broader than the model. Compute determines how much intelligence can be run and at what cost. Data determines what the system can learn from. Memory determines whether work persists across sessions. Context determines what the system knows in the moment. Routing determines which model, tool, or human handles which task. Evaluation determines whether output is good enough to use. Trust systems determine what can happen automatically and what requires approval. Distribution determines who gets access to users and customers. Operating interfaces determine where human work actually happens.

These layers matter because they are the new machinery of cognitive production. They decide not only whether an AI system can answer a prompt, but whether a firm can turn repeated cognitive tasks into reliable output at scale.

Our response cannot just be nostalgia or panic. The more constructive path is not full automation for its own sake. It is a human operator model: humans remain responsible for setting goals, defining acceptable risk, approving consequential actions, and improving the system over time. The operator is not just a prompt writer. The operator is the person who understands what the system is allowed to do, what evidence it used, where it tends to fail, and when a human decision has to interrupt the machine.

That model requires inspectable systems: clear instructions, durable memory, logs, receipts, confidence thresholds, approval gates, and review boundaries. The point is to build trust, not just capability.

If humans remain operators with real agency, those layers need to be inspectable and contestable. If they are opaque and privately controlled, we risk becoming tenants inside systems that increasingly mediate our own capacity to think, make, coordinate, and earn.

AI is already changing work. The political fight is whether the infrastructure of cognition becomes something people can understand, shape, and contest, or something they can only rent.