Tokens, Frontiers, and Bubbles: Why AI’s Economic Value Will Lag Behind Its Capital Expenditure

I want to offer a preliminary framework for understanding AI tokens, frontier-model economics, and the bubbles that tend to accompany technological revolutions. This is not a definitive theory. It is an attempt to connect several observations that are often discussed separately: the commoditization of intelligence, the deteriorating pricing power of frontier laboratories, the migration from tokens to workflows and operating systems, and the structural tendency for capital expenditure to outrun realized economic value.

My central argument is straightforward: AI may ultimately become one of the most consequential technologies in human history, but that does not mean today’s investment can be justified by tomorrow’s distant productivity gains. Capital markets routinely compress decades of technological diffusion into a few years of financial expectations. Because technological capabilities can improve faster than organizations can absorb them, expenditure will initially rise much faster than monetizable value. That temporal mismatch makes an investment bubble not merely possible, but highly probable.

The Token as a Graded Cognitive Commodity

I think of a token as a unit of cognitive production. It is not intelligence itself, but a standardized input into machine-generated reasoning, interpretation, writing, coding, planning, and decision-making.

However, tokens are not homogeneous. They should be understood as graded commodities.

The analogy I find useful is crude oil. Light, medium, and heavy crude are all oil, but they are not economically identical. Products within the same grade are relatively interchangeable, while substantial barriers exist between grades. Buyers may not care greatly about the precise origin of a standardized barrel of oil, but they care enormously about its composition, processing requirements, and suitability for a particular refinery.

AI tokens have a similar structure. Within a particular capability tier, users often care less about which laboratory produced the token than about its price, latency, reliability, context length, and compatibility with their software. A company may use several models simultaneously, routing different requests to different providers. Once multiple suppliers can perform the same task with comparable quality, the token increasingly behaves like a commodity.

I therefore divide cognitive tokens into two broad categories: commoditized cognition and frontier cognition.

Commoditized cognition includes tasks that many models can perform adequately. Examples include summarizing documents, translating routine text, generating standard marketing copy, classifying customer-service requests, extracting structured data, and writing conventional software components. When several vendors can supply approximately equivalent outputs, competition drives the price toward marginal cost.

This process is not gradual in the conventional industrial sense. It can be extraordinarily rapid because software capabilities diffuse faster than physical production techniques. Once an architectural improvement becomes widely understood, it may be replicated, distilled, optimized, or approximated by competitors. Hardware becomes more efficient, inference software improves, and older frontier models become cheaper to operate. A cognitive capability that was scarce and expensive can become abundant within a few product cycles.

Frontier cognition is different. It refers to the highest level of capability that only a small number of laboratories can provide at a given moment. Frontier models may command a substantial premium because they can solve tasks that cheaper models cannot solve reliably. Their advantage may appear in advanced coding, scientific reasoning, long-horizon planning, multimodal interpretation, autonomous tool use, or performance under ambiguous conditions.

But the crucial point is that the valuable asset is not cognition in a static sense. It is temporary frontier status.

A laboratory does not possess durable pricing power merely because it created an intelligent model. It possesses pricing power because its current model remains meaningfully ahead of competing alternatives. The moment that capability is reproduced, approximated, or rendered unnecessary, it falls from frontier cognition into commoditized cognition.

In other words, the laboratory is not really selling intelligence. It is selling perhaps six months of cognitive superiority.

Frontier Intelligence Is a Rapidly Depreciating Asset

This framework explains why leading AI laboratories must continue spending aggressively. Their existing products depreciate at an unusual rate — not necessarily because the models become technically worse, but because the competitive frontier moves forward.

A powerful model can remain just as capable as before while losing most of its economic scarcity. Its absolute intelligence is unchanged; its relative position has deteriorated.

This is a peculiar form of depreciation. A factory depreciates because machinery wears out. A frontier model depreciates because competitors catch up and customers discover cheaper substitutes. The economic half-life of the product may therefore be much shorter than its technical lifespan.

For the leading laboratories, stopping research is equivalent to voluntarily surrendering pricing power. Once their premium capabilities become widely available, they enter a market dominated by marginal-cost competition. Their expensive research infrastructure then has to compete against distilled models, open-source systems, specialized providers, and companies willing to accept lower margins.

The result is a permanent technological arms race. Every laboratory must spend heavily not merely to advance, but to prevent its current revenue base from being commoditized.

This creates a difficult strategic paradox. The better the industry becomes at reproducing intelligence, the less durable any single laboratory’s advantage becomes. AI progress expands the total market, but it can simultaneously destroy the economic moat of individual model providers.

The economics become especially dangerous if frontier-model differentiation begins to narrow. This could happen because scaling becomes prohibitively expensive, because data or energy constraints intensify, because regulation restricts training, or because models approach similar performance ceilings on commercially relevant tasks.

If the frontier compresses, the laboratories do not necessarily become technologically irrelevant. But their pricing power could deteriorate sharply. A market containing several nearly equivalent frontier systems would behave less like a monopoly on advanced cognition and more like a capital-intensive commodity industry.

That would be disastrous for companies whose valuations assume both enormous expenditure and persistent frontier margins.

Is the Economic Value of Each New Model Generation Diminishing?

The first major question is whether the incremental economic value of frontier models is declining.

Suppose one generation of a model is substantially more capable than the previous generation. That does not automatically mean its economic value increases proportionately. Capability and monetizable utility are not the same variable.

A model may become better at dozens of benchmarks without becoming proportionately more valuable to most customers. Many commercial tasks have a sufficiency threshold. Once a model can perform the task reliably enough, additional intelligence produces limited economic benefit.

A customer-service system does not need to solve advanced mathematics. A document-extraction system does not need to formulate original scientific hypotheses. A code-completion tool may not need to autonomously design an entire software architecture. For these applications, the cheapest model that satisfies the operational requirement is usually the economically rational choice.

As open-source models and inexpensive proprietary models become adequate for more tasks, the domain in which customers must purchase frontier intelligence contracts. Frontier models may continue to improve, but the number of tasks that strictly require them may decline as a share of the total AI workload.

This does not mean frontier models become worthless. It means their premium must be justified by an increasingly narrow set of difficult tasks.

The situation resembles a hierarchy of cognitive demand. At the bottom are abundant, repetitive tasks that can be automated by inexpensive models. Above them are moderately complex tasks requiring better reasoning or reliability. At the top are rare tasks where a small improvement in intelligence may create enormous value.

The frontier provider faces an uncomfortable arithmetic. The highest-value tasks may be extraordinarily lucrative, but they may also be relatively scarce. Meanwhile, most token volume migrates toward cheaper systems.

Consequently, the frontier premium may shrink even while frontier capability continues to advance.

Tokens Are Probably the Wrong Final Unit of Sale

The second major question is whether the industry’s unit of measurement will change. I believe it must.

Tokens are an engineering-oriented billing unit. They measure computational consumption, not completed economic value. Customers do not fundamentally want tokens. They want reports, software, diagnoses, designs, decisions, resolved support cases, completed transactions, and higher revenue.

Selling tokens is analogous to a cloud-computing provider selling processor cycles. It is useful as an intermediate pricing mechanism, but it exposes the supplier to substitution. Customers can route workloads across multiple models, optimize consumption, switch providers, or use open-source alternatives.

A token provider grants customers considerable autonomy. The customer decides which model to call, how to construct the workflow, where to store the data, and when to replace the supplier. This weakens the provider’s strategic control.

For that reason, the industry is likely to move through three stages:

selling tokens → selling workflows → selling AI operating systems

A workflow provider does not merely answer a prompt. It completes a multistep task. It retrieves information, uses tools, verifies intermediate results, coordinates specialized models, and produces a finished output. The relevant unit may become a resolved insurance claim, a completed software feature, an executed procurement process, or a qualified sales lead.

An AI operating system goes further. It becomes the orchestration layer through which users interact with models, applications, data, agents, permissions, memory, and enterprise processes. At that level, the provider can internalize model routing and reduce the customer’s ability to substitute individual components.

This is where durable power may eventually reside.

A model can be replaced. A deeply embedded workflow is harder to replace. An operating system that governs identity, data access, organizational memory, agent permissions, and application integration is harder still.

The strategic competition in AI may therefore shift away from the question, “Who produces the smartest token?” toward the question, “Who controls the environment in which intelligence is deployed?”

The laboratories expanding beyond tokens are not merely seeking additional revenue. They are attempting to escape commoditization.

Why Technological Revolutions Produce Bubbles

Large financial bubbles usually require two elements: a compelling narrative that is difficult to falsify in the short term, and a mechanism that amplifies capital deployment.

The amplification mechanism can take many forms. It may involve rolling debt, customer prepayments, long-term purchasing agreements, supplier financing, equity issuance, venture-capital recycling, or valuation reflexivity. Rising asset prices can themselves become a form of leverage by lowering the cost of capital and validating further expenditure.

Technological revolutions create especially dangerous bubbles because their central narrative is often true.

AI can improve productivity. Data centers are necessary. Semiconductor demand can rise dramatically. New applications can create large markets. Corporate earnings may genuinely increase. These are not fictional propositions.

The danger emerges when temporary conditions are treated as permanent equilibria.

A company may experience extraordinary margins because supply is temporarily constrained. Investors may then capitalize those margins as though they will persist for a decade. A supplier may obtain long-term contracts during a period of acute scarcity, and the market may interpret those contracts as proof of structural pricing power.

Consider memory semiconductors. Standardized memory is fundamentally a commodity. For equivalent specifications, buyers generally resist paying a permanent 100 percent premium merely because one major manufacturer produced the component rather than another.

During a shortage, however, prices and margins can rise sharply. Manufacturers may sign long-term agreements, and analysts may reach a consensus that supply will remain tight for several years. That consensus can be entirely reasonable.

The error begins when investors extend a shortage expected through 2027 or 2028 into an implicit assumption that scarcity, average selling prices, and profit margins will remain elevated into 2030 or 2035.

Long-term contracts do not completely eliminate this risk. Contracts are often renegotiated when the underlying economic environment changes. A contractual promise is not identical to permanent market power.

The same principle applies to AI infrastructure. Current demand may be real, but investors can still overestimate its duration, profitability, or concentration. A genuine technological revolution can produce an authentic earnings boom and an irrational valuation simultaneously.

Productivity Moves Faster Than Production Relationships

Why does capital expenditure tend to outrun monetization?

Because technical capability can advance much faster than the institutional structures required to exploit it.

A new model can be released overnight. A company cannot redesign its entire organization overnight.

To capture the technology’s full value, firms must change workflows, incentives, responsibilities, software architecture, compliance systems, hiring practices, managerial structures, and sometimes their entire business model. Employees must learn how to supervise AI systems. Organizations must determine which decisions can be delegated and which require human accountability. Legal and regulatory frameworks must adapt.

These are not primarily technical problems. They are coordination problems.

History suggests that the greatest productivity gains from general-purpose technologies often require complementary organizational innovation. Electrification did not transform manufacturing merely because factories replaced steam engines with electric motors. The larger gains emerged when factory layouts, production lines, and managerial processes were redesigned around distributed electric power.

Information technology followed a similar pattern. Installing computers was much easier than restructuring firms to exploit networked information, software automation, and digital coordination.

AI may diffuse faster than electricity or early computing, but organizational inertia has not disappeared. Large institutions remain constrained by procurement cycles, legacy systems, internal politics, regulatory obligations, security concerns, and human resistance.

This produces a recurrent pattern: the technology offers three floors of present economic value while capital markets construct ten floors of expectations above it.

Eventually, the technological foundation may support the full structure. But the financial market does not patiently wait for the building to be completed. It prices the finished skyscraper before the institutional foundation has hardened.

Why Narrative Weakness Alone Does Not Burst the Bubble

A bubble usually breaks when two events coincide: the narrative weakens, and the amplification mechanism fails.

Narrative deterioration by itself is insufficient. Investors with strong conviction can tolerate disappointing quarters. They may believe that monetization has merely been delayed. They may correctly argue that the long-term technological thesis remains intact.

This is particularly true when household and corporate balance sheets remain relatively healthy. As long as liquidity is abundant and financing remains accessible, capital can continue supporting a story whose short-term economics are deteriorating.

The decisive event is often the failure of the financing mechanism.

That failure can be triggered by tighter monetary policy, rising credit spreads, a major corporate default, canceled procurement commitments, collapsing collateral values, regulatory intervention, or a sudden recognition that expected cash flows cannot support existing capital structures.

Historically, central banks have frequently held the needle, even when they did not create the bubble’s underlying technological narrative. A change in the cost of capital can transform tolerable uncertainty into forced liquidation.

For AI, early signs of narrative weakening may include greater scrutiny of AI-related revenue, reduced access to frontier models, lower-than-expected utilization, customer resistance to premium pricing, or evidence that inexpensive models can satisfy most enterprise demand.

The economic effects of reduced model usage may appear with a lag. A company can cut internal AI consumption today while infrastructure suppliers continue reporting strong results for another quarter because of existing commitments and delivery schedules.

However, weaker usage alone does not necessarily mark the final peak. A bubble can continue after its narrative becomes visibly fragile. As long as financing remains available, participants may interpret every disappointment as temporary.

The bubble ends when patience becomes financially impossible.

The Nonlinear-Value Objection

A serious objection to my framework is that it may be too linear.

Ordinary intelligence can be treated as a priced commodity because incremental performance improvements often produce incremental competitive advantages. But sufficiently advanced intelligence may generate discontinuous outcomes.

A model that is 10 percent better on a benchmark may be worth only slightly more. Yet a model that crosses a critical threshold — discovering a commercially valuable molecule, autonomously solving a strategic engineering problem, or coordinating a complex military operation — could produce value many orders of magnitude greater.

This argument is valid. Technological value is not always linear.

A system that transforms an impossible task into a possible one cannot be valued through ordinary marginal analysis. The first reliable treatment for a fatal disease is not merely 20 percent more valuable than an ineffective treatment. A military system that enables a previously impossible operational capability may alter the balance of power rather than produce a modest efficiency gain.

Threshold effects matter. Intelligence may display increasing returns in domains where a minimum capability is required before any economic value can be produced.

But nonlinear potential should not be confused with realized value.

Claims about AI identifying elusive military targets, doubling human lifespan, or autonomously producing major scientific breakthroughs require strong evidence. It is easy to take a plausible future capability and treat it as a current commercial fact. That is precisely how technological narratives become financially dangerous.

I therefore distinguish among three categories:

First, demonstrated capability: the system has repeatedly produced the result under real operating conditions.

Second, plausible but unverified capability: the system may be able to produce the result, but the evidence is limited, private, or difficult to reproduce.

Third, speculative capability: the result is theoretically imaginable but has not yet been credibly demonstrated.

Capital markets frequently price the second category as though it were the first and occasionally price the third as though commercialization were imminent.

The possibility of nonlinear breakthroughs does not invalidate a bubble thesis. It can actually intensify the bubble because extreme outcomes are difficult to disprove. When the potential payoff is almost unlimited, investors can justify almost any present valuation by assigning a small probability to a transformative future.

Linear Reasoning Is Usually Rational — Until It Is Not

Human beings often rely on linear extrapolation because it works reasonably well in most ordinary environments. Most companies do not suddenly increase their productivity by a factor of one hundred. Most scientific projects do not generate civilization-altering discoveries. Most technological improvements are incremental.

Linear reasoning is therefore not inherently foolish. It is a useful default model.

The error lies in applying it mechanically during a regime change — or, conversely, abandoning it completely whenever someone invokes technological disruption.

There are two symmetrical mistakes.

The first is to assume that progress must remain incremental because it has historically been incremental. This can cause investors and institutions to underestimate genuine technological discontinuities.

The second is to assume that because discontinuities are possible, every new model release will create exponential economic value. This converts possibility into certainty and scientific potential into a valuation premise.

A rigorous framework must accommodate both outcomes. Most improvements will probably generate diminishing marginal value. A small number may cross thresholds and create enormous nonlinear returns.

The central investment problem is that these rare breakthroughs are extremely difficult to identify in advance. Their social value may also be captured by users, governments, complementary businesses, or society broadly rather than by the laboratory that trained the model.

A technology can transform civilization without producing attractive returns for every company financing its infrastructure.

Railways changed national economies, but many railway investors lost money. The internet transformed commerce and communication, but numerous internet companies disappeared. Excessive infrastructure investment can be socially productive while remaining financially destructive for the marginal provider.

Technological significance and investment profitability are separate questions.

Capital Markets Overestimate the Short Term and Underestimate the Long Term

My broader conclusion is that capital markets systematically misprice the timetable of technological change.

They overestimate the short term because current narratives are translated into immediate revenue projections. They underestimate the long term because, after the bubble collapses, investors become excessively skeptical and ignore the applications enabled by abundant, cheap infrastructure.

The cycle is recurrent.

First, a genuine breakthrough appears.

Then, capital floods into infrastructure because the ultimate market seems enormous.

Next, investment grows faster than near-term demand. Valuations assume rapid monetization and persistent scarcity.

When organizational adoption fails to keep pace, revenue expectations weaken. Financing conditions eventually tighten, and the bubble breaks.

Afterward, the market overreacts in the opposite direction. Investors conclude that the technology was overhyped, even though the technological thesis may remain valid.

But the excess infrastructure does not disappear. Its cost falls. Entrepreneurs gain access to cheap computation, networks, factories, or distribution capacity. Applications that were previously uneconomical become viable.

The real transformation often accelerates after the speculative capital has been destroyed.

This is why I can simultaneously believe that AI is a historic technological revolution and that the AI investment cycle will produce a severe bubble. These positions are not contradictory. In fact, they are mutually reinforcing.

The more consequential the technology, the more persuasive the narrative. The more persuasive the narrative, the easier it becomes to finance excessive capacity. The more excessive the capacity, the more painful the eventual financial adjustment. Yet that same overcapacity can reduce costs and support the next generation of applications.

Conclusion: Intelligence Will Become Abundant, but Control May Remain Scarce

I expect ordinary machine intelligence to become progressively cheaper. Most cognitive tokens will be commoditized, and their prices will approach the cost of computation, energy, infrastructure, and distribution.

Frontier cognition will retain temporary premiums, but those premiums will be unstable because frontier status is a rapidly depreciating asset. Leading laboratories will have to keep spending simply to prevent their products from slipping into commodity competition.

As the token itself becomes less defensible, the industry will move upward into workflows, agents, platforms, and AI operating systems. The durable source of power may not be ownership of the smartest model, but control of the environment in which models are selected, coordinated, and embedded into economic activity.

At the same time, capital expenditure will probably continue to exceed directly attributable AI revenue. Technical progress can be rapid, but production relationships — organizations, incentives, laws, workflows, and business models — change more slowly.

This gap creates the conditions for a bubble. The narrative will weaken before the technology fails. The bubble will burst only when narrative weakness intersects with a breakdown in financing or valuation reflexivity.

Afterward, the market will probably make its customary second mistake: it will become excessively pessimistic. It will interpret financial destruction as technological failure. Meanwhile, cheap and abundant infrastructure will enable applications that were impossible during the expensive frontier phase.

The market repeatedly overestimates how quickly a technological revolution will generate profits and underestimates how thoroughly it will eventually transform society.

The timing is usually wrong. The transformation is often real.