The AI Bubble

The AI Bubble

November, 2026

The greatest capital expenditure frenzy in technology history is colliding with an uncomfortable economic reality: artificial intelligence may be revolutionary, but that does not mean the companies building it are worth what investors are paying.

There is an old saying on Wall Street: "Markets can remain irrational longer than you can remain solvent." The observation, commonly attributed to John Maynard Keynes, captures the extraordinary spectacle unfolding across the global artificial intelligence industry.

Hundreds of billions of dollars are being committed to GPUs, data centers, electricity infrastructure, foundational models, and AI startups. Trillions of dollars of market capitalization have become dependent on the assumption that artificial intelligence will transform virtually every industry and create unprecedented economic value.

Yet beneath the spectacular announcements, soaring valuations, and breathless predictions of artificial general intelligence lies a far more uncomfortable question.

Where is the money?

Not the investment money. There is plenty of that. Not the revenue circulating between cloud providers, model developers, and infrastructure companies. There is plenty of that too.

Where is the sustainable, high-margin, independently generated economic value sufficient to justify the enormous capital already committed?

That question is becoming increasingly difficult to answer.

The AI industry has constructed an extraordinary financial machine in which companies invest in one another, purchase computing infrastructure from one another, rent computing capacity to one another, and report the resulting transactions as evidence of explosive demand. Meanwhile, the ultimate customers—the businesses and consumers who must eventually pay for all this infrastructure—have yet to demonstrate willingness to spend at anything approaching the scale required.

The distinction is crucial. A technological revolution and an investment bubble can coexist. The internet was revolutionary in 1999. The railway was revolutionary in the nineteenth century. Neither prevented investors from losing fortunes when financial expectations outran commercial reality.

Artificial intelligence is following the same historical script.

And the reckoning may be much closer than the market believes.

1. The $600 Billion Question Nobody Wants to Answer

In June 2024, Sequoia Capital partner David Cahn published an influential analysis titled AI's $600B Question. His argument was deceptively simple: the extraordinary spending on AI infrastructure required an equally extraordinary amount of revenue from customers to justify it.

The industry had built an enormous supply of computational capacity without establishing a corresponding market of paying end users.

Cahn's analysis expanded an earlier $200 billion question into a $600 billion problem as infrastructure investment accelerated.

The underlying arithmetic remains devastating.

A company purchasing billions of dollars of GPUs cannot recover that investment merely by announcing a powerful new model. The hardware must be deployed, utilized, maintained, powered, cooled, and eventually replaced. The resulting services must generate sufficient revenue to cover depreciation, operating costs, financing, research, and a reasonable return on invested capital.

And GPUs do not last forever.

Unlike a railway bridge or a hydroelectric dam, an AI accelerator can become economically obsolete within a few years. New architectures deliver greater performance per watt, newer models demand different hardware configurations, and competitive pressure forces infrastructure providers to continually reinvest.

This creates a vicious cycle: enormous upfront expenditure followed by relentless pressure to spend even more.

Goldman Sachs raised similar concerns in its June 2024 report Gen AI: Too Much Spend, Too Little Benefit? Jim Covello, the firm's head of global equity research, questioned whether generative AI could solve sufficiently valuable problems to justify its enormous infrastructure costs.

His argument struck at the heart of the investment thesis: transformative technologies typically deliver dramatic improvements in cost or capability. Yet generative AI was being deployed at extraordinary expense into many tasks that were already being performed adequately and cheaply.

The industry's answer has largely been to promise that future models will create the necessary demand.

But promises do not pay electricity bills.

The fundamental economic problem is that infrastructure spending is real, immediate, and measurable, while the anticipated productivity revolution remains uncertain, uneven, and frequently difficult to monetize.

An investment cycle built on this imbalance cannot expand indefinitely.

2. The Great AI Revenue Illusion

One of the most revealing features of the AI boom is the circularity of its financial relationships.

Consider a simplified example.

A large technology company invests billions in an AI model developer. The model developer spends substantial portions of that investment purchasing cloud computing services from the same technology company. The cloud provider reports growing AI-related revenue, which investors interpret as validation of the original investment.

On paper, everyone appears to be growing.

But the critical question remains: how much new money entered this ecosystem from independent customers purchasing economically valuable services?

Capital circulating between investors, cloud providers, and AI developers is not equivalent to sustainable end-market demand.

The pattern resembles the telecommunications boom of the late 1990s, when equipment vendors, network operators, and financiers supported one another's expansion while assuming that future internet traffic would eventually justify the infrastructure.

During the dot-com era, Cisco became one of the most valuable companies in the world by selling the equipment needed to build the internet. Its customers were investing aggressively in networks, often on the assumption that demand would materialize later.

When capital spending slowed, the entire ecosystem suffered.

The same structural vulnerability exists in today's AI economy.

The hardware supplier records revenue when a GPU is sold. The cloud provider records revenue when computing capacity is rented. The model developer records revenue when its API is called. But the enterprise customer at the end of this chain must ultimately earn or save enough money to justify the combined costs.

If that final economic equation fails, the revenues upstream become unsustainable.

This is the central illusion of the AI boom: the industry has confused the financing of AI infrastructure with the monetization of AI itself.

And those are two very different things.

3. AI Is Not SaaS, and Its Economics Are Far Worse

Traditional software became one of the world's most profitable industries because of a remarkable economic characteristic: once developed, software can be replicated and distributed at almost negligible marginal cost.

A software company could spend millions building a product and then sell the same product to thousands or millions of customers. Revenue could grow dramatically without a corresponding increase in the cost of serving each customer.

This created the extraordinary gross margins that made software-as-a-service businesses so attractive to investors.

Generative AI breaks this economic model.

Every prompt consumes computational resources. Every generated response requires inference. Every additional customer creates additional demand for processors, memory, networking, and electricity.

More customers do not automatically mean dramatically better margins. They can mean more infrastructure expenditure.

The problem becomes especially severe for applications involving complex reasoning, long contexts, multimodal processing, or autonomous agents that repeatedly invoke models.

An ordinary software application might execute a few database queries to complete a transaction. An AI agent might perform dozens or hundreds of model calls to accomplish the same business objective.

The economics are fundamentally different.

Now consider the thousands of startups building products on top of foundational models.

Many of these businesses are effectively interfaces wrapped around APIs controlled by larger companies. Their differentiation often consists of prompts, workflows, user interfaces, and integrations.

These features can be reproduced.

As foundational model providers improve their products, they absorb functionality previously offered by independent application developers. Meanwhile, competing startups use increasingly similar models to offer increasingly similar services.

The result is predictable: falling prices, weakening differentiation, rising customer acquisition costs, and compressed margins.

The businesses that appeared to be the next generation of software monopolies begin to resemble commodity service providers.

And commodity businesses do not deserve extraordinary software valuations.

4. The Enterprise AI Revolution That Has Yet to Pay for Itself

The strongest argument for AI valuations has always been enterprise productivity.

Businesses, we are told, will pay enormous sums because AI will automate knowledge work, replace expensive processes, and fundamentally transform operations.

Yet the evidence of widespread economic returns has been far less impressive than the investment announcements.

In 2024, the RAND Corporation reported that more than 80% of AI projects fail, a figure it described as roughly twice the failure rate of non-AI information technology projects. The study highlighted problems including misunderstanding the business problem, inadequate data, insufficient infrastructure, and misalignment between technology and operational needs.

The finding was not that AI could never work. It was that successfully transforming an AI demonstration into a dependable business system was much harder than the industry's promotional narratives suggested.

A 2025 MIT NANDA report, The GenAI Divide: State of AI in Business 2025, went further, reporting that 95% of the organizations in its studied enterprise generative AI initiatives were seeing no measurable profit-and-loss impact. The figure referred to the report's examined initiatives and methodology, not a universal census of all AI deployments.

Nevertheless, the implication was striking.

Companies were experimenting enthusiastically, but experimentation was not translating reliably into measurable financial returns.

The distinction between a successful demonstration and a profitable production system is enormous.

A chatbot that answers questions during a presentation may look impressive. A system that must operate continuously, securely, accurately, and affordably across thousands of employees is an entirely different engineering and economic proposition.

Production deployments require integration, testing, security, compliance, monitoring, human oversight, maintenance, and ongoing model expenditure.

A technology that saves an employee ten minutes but introduces substantial verification costs may deliver little net value.

A technology that automates 80% of a workflow but requires human intervention for the remaining 20% may not reduce staffing costs proportionately.

A technology that generates plausible answers but occasionally makes expensive mistakes can create liabilities greater than its productivity benefits.

The industry's promotional materials emphasize gross productivity gains.

Corporate finance departments care about net returns.

The gap between those two numbers is where the AI bubble begins to unravel.

5. The Law of Diminishing Returns Is Catching Up With Large Language Models

For several years, the foundational assumption behind the AI investment boom was that larger models would inevitably become dramatically more intelligent.

More parameters. More training data. More GPUs. More electricity. More money.

The formula appeared almost magical: increase computational resources and watch intelligence improve.

But exponential growth in expenditure does not guarantee exponential growth in useful capability.

Stanford University's 2025 AI Index documented that training compute for notable AI models was doubling approximately every five months. This extraordinary growth in computational requirements represents an escalating financial commitment by developers seeking to remain competitive.

Stanford HAI+1

The difficulty is that a model which costs ten times as much to develop is not necessarily ten times as useful to customers.

An improvement in a benchmark score does not automatically translate into an equivalent improvement in business productivity. A model that answers 92% of questions correctly rather than 89% may represent an impressive technical achievement, but it does not necessarily create a commercially viable product.

For applications demanding near-perfect reliability, both models may still require human verification.

For applications tolerating occasional mistakes, the cheaper model may already be sufficient.

This creates an increasingly uncomfortable economic relationship between model developers and their customers.

Developers must spend billions to produce incremental improvements. Customers increasingly expect those improvements to be included in existing subscriptions or delivered at lower prices.

The research community has also identified another fundamental problem: the supply and quality of training data.

A 2024 paper published in Nature, titled AI Models Collapse When Trained on Recursively Generated Data, demonstrated that indiscriminately training successive generations of models on synthetic content produced by earlier models can cause a deterioration in the learned data distribution.

The researchers described a process in which models progressively lose information about less common events and patterns, ultimately producing increasingly distorted representations of the original data.

This creates an extraordinary paradox.

The more AI-generated content floods the internet, the more difficult it becomes to distinguish authentic human-generated information from machine-generated material.

The industry risks contaminating the very resource on which its next generation of models depends.

Better training techniques, curated datasets, and synthetic-data filtering require additional investment. These costs further complicate the economics of scaling.

The central financial assumption—that unlimited expenditure will produce unlimited improvements in intelligence—is becoming increasingly difficult to defend.

And if model capability improvements become more expensive while their commercial value grows more slowly, the investment thesis begins to collapse.

6. Open-Source AI Is Destroying the Pricing Power of Proprietary Models

Perhaps the greatest threat to the valuations of proprietary AI companies is not another expensive model.

It is a sufficiently capable model that customers can obtain cheaply, customize, and operate themselves.

The history of computing is filled with examples of technologies that began as expensive proprietary products and eventually became inexpensive commodities.

Linux transformed the economics of operating systems. Open-source databases challenged proprietary database vendors. Open web technologies reduced dependence on expensive development platforms.

AI is experiencing a similar transformation at extraordinary speed.

Open-weight model ecosystems, including Meta's Llama family and numerous competing architectures, have made powerful AI capabilities accessible to developers and enterprises without requiring them to build foundational models from scratch.

The consequences for pricing are profound.

Stanford's 2025 AI Index found that the cost of querying a model with approximately GPT-3.5-level performance on the MMLU benchmark fell from $20 per million tokens in November 2022 to $0.07 by October 2024.

That represents a reduction of more than 280 times in approximately eighteen months.

For consumers and enterprises, this is excellent news.

For investors who financed proprietary model developers at extraordinary valuations, it is a warning.

Imagine investing billions of dollars to create a product whose selling price falls by more than 99% before the original investment has been recovered.

The product may become enormously popular. Usage may explode. The technology may transform industries.

And the original investors may still lose money.

This is the distinction between technological adoption and financial returns that the AI market has consistently failed to appreciate.

Competition does not need to produce a better model to destroy a premium-priced business. It only needs to produce a model that is good enough.

Once AI capabilities become widely available, customers begin choosing on price, reliability, privacy, deployment flexibility, and integration rather than brand prestige.

The economic value migrates away from the model itself.

The irony is devastating: the very improvements in AI technology celebrated by investors are steadily undermining the scarcity on which their valuations depend.

7. The Dot-Com Bubble Is Repeating Itself

In March 2000, Cisco Systems briefly became the world's most valuable company.

The internet was transforming communications, commerce, entertainment, and business. Demand for networking equipment appeared limitless.

Investors believed the infrastructure providers enabling the internet revolution would enjoy decades of uninterrupted growth.

Then the bubble burst.

Cisco's share price fell by roughly 80% from its peak during the subsequent technology crash.

The internet did not disappear.

Internet usage continued expanding. Online commerce grew. Digital communications eventually transformed almost every industry.

But investors who had paid extraordinary prices for anticipated future growth suffered enormous losses.

The technology was real.

The valuations were not.

That distinction is central to understanding the AI bubble.

Today's GPU manufacturers occupy a position strikingly similar to the networking equipment suppliers of the late 1990s.

During the dot-com boom, telecommunications companies raced to install fiber-optic networks in anticipation of future demand. Vast amounts of capital were committed to infrastructure before the revenue-generating applications had matured.

When expectations collapsed, the industry was left with substantial excess capacity.

Some of that infrastructure eventually became useful. But its eventual usefulness did not rescue the shareholders and creditors who financed it at inflated valuations.

The same danger now hangs over AI data centers.

A hyperscaler can spend tens of billions of dollars constructing facilities filled with advanced accelerators. Those facilities may eventually support valuable applications.

But if the applications do not generate sufficient revenue before the equipment depreciates, the investment can still be economically disastrous.

The historical parallel extends beyond infrastructure.

During the dot-com bubble, companies frequently justified their valuations using website traffic, user growth, and speculative future market opportunities rather than earnings.

Today's AI companies emphasize token consumption, model benchmarks, active users, compute clusters, and hypothetical productivity gains.

These measurements may demonstrate technological activity.

They do not establish profitability.

The dot-com era taught investors a lesson that appears to have been forgotten: a company can be right about the future and catastrophically wrong about the price of reaching it.

8. The Pickaxe Sellers Cannot Escape the Gold Rush's Collapse

The most celebrated investment strategy during a gold rush is to sell picks and shovels rather than search for gold.

During the AI boom, this has made GPU manufacturers and infrastructure providers extraordinarily attractive.

They receive revenue regardless of whether the AI applications being developed ultimately succeed.

At least initially.

But the strategy has an overlooked weakness.

The pickaxe seller's prosperity depends on miners continuing to purchase pickaxes.

If miners discover that the gold is insufficient to justify their expenditure, demand for equipment eventually collapses.

The AI infrastructure market is similarly dependent on continued capital investment by a relatively concentrated group of enormous technology companies.

These buyers have the financial resources to sustain the boom for longer than most previous speculative industries.

But they are not immune to economic reality.

At some point, shareholders will demand to know whether the hundreds of billions committed to AI infrastructure are generating acceptable returns.

If the answer remains unsatisfactory, capital expenditure will be reduced.

The consequences would extend far beyond GPU manufacturers.

Semiconductor foundries, memory suppliers, networking equipment companies, data center operators, electricity infrastructure providers, cooling manufacturers, construction firms, and cloud infrastructure businesses have all become exposed to expectations of continued AI expansion.

A slowdown in spending by a handful of hyperscalers could therefore propagate across a much larger economic ecosystem.

There is another important problem.

Much of the installed GPU infrastructure represents a depreciating asset rather than a permanent competitive advantage.

New generations of accelerators can deliver better performance and efficiency, reducing the economic attractiveness of older hardware.

The industry could therefore find itself in the extraordinary position of owning enormous amounts of expensive computing equipment while simultaneously facing pressure to replace it.

A gold rush in which the pickaxes become obsolete every few years is an exceptionally dangerous business to finance.

9. The Legal and Regulatory Bill Has Not Yet Arrived

AI valuations also depend on an assumption that receives surprisingly little attention: the industry will be allowed to continue operating under broadly favorable legal and regulatory conditions.

That assumption is far from guaranteed.

Foundational models have been trained using enormous collections of books, articles, images, software code, websites, and other copyrighted material.

Publishers, authors, artists, software developers, and rights holders have challenged aspects of these practices through litigation.

The eventual outcomes will vary across jurisdictions, factual circumstances, and types of use.

But the financial risk is substantial.

If major AI developers must obtain additional licenses for large portions of their training data, compensate rights holders, alter data collection practices, or defend costly litigation, the economics of developing foundational models could deteriorate.

A business model that appears profitable when certain inputs are obtained without licensing payments may look very different when those inputs carry substantial recurring costs.

The same problem applies to competition regulation.

AI development is becoming increasingly concentrated around companies controlling cloud infrastructure, semiconductor supply, distribution channels, foundational models, and enterprise software ecosystems.

Regulators have reasons to examine whether strategic investments, exclusive agreements, and preferential infrastructure arrangements restrict competition.

The AI boom has effectively priced in a future of enormous commercial expansion.

It has not necessarily priced in the full cost of the legal framework within which that expansion must occur.

Investors are paying for a future in which AI becomes ubiquitous.

They may discover that making AI ubiquitous is considerably more expensive than anticipated.

10. The Market Is Pricing AI as a Monopoly While Technology Is Turning It Into a Commodity

The contradiction at the center of the AI investment thesis is becoming impossible to ignore.

The market is valuing many AI companies as though they will enjoy the economics of monopolies.

Yet the underlying technology is moving rapidly toward commoditization.

Monopoly valuations depend on durable competitive advantages.

Those advantages may come from exclusive intellectual property, network effects, regulatory barriers, proprietary infrastructure, unique data, or extraordinary switching costs.

But what happens when competing models achieve similar capabilities?

What happens when open-weight models become sufficiently powerful for most enterprise applications?

What happens when customers can reproduce AI applications using readily available development tools?

What happens when the cost of inference continues declining and competition forces providers to pass those savings on to customers?

The result is a market in which AI becomes increasingly valuable to users while becoming progressively less profitable for many suppliers.

This is not an unusual economic outcome.

Air travel transformed global transportation without consistently generating attractive returns for airlines.

Telecommunications transformed human communication while destroying enormous amounts of investor capital during periods of overinvestment.

The internet created extraordinary economic value, but countless internet companies disappeared.

A technology can be indispensable and still be a terrible investment.

AI is approaching precisely this moment.

The fundamental capability is becoming widespread. The barriers to entry for many applications are falling. Customers have growing bargaining power.

Meanwhile, infrastructure providers and model developers must recover investments made under assumptions of extraordinary future profitability.

Those assumptions cannot all be correct.

11. The Moment the Bubble Bursts

Speculative bubbles rarely burst because the underlying technology suddenly stops working.

They burst when expectations change.

For years, AI companies have been rewarded for announcing larger models, greater computing capacity, more ambitious data centers, and increasingly expensive partnerships.

The announcements themselves have been treated as evidence of future success.

But eventually, financial markets begin asking different questions.

Instead of asking how many GPUs a company plans to purchase, investors ask how much profit those GPUs generate.

Instead of celebrating a model's benchmark performance, they ask how much customers are willing to pay for the improvement.

Instead of rewarding announcements of billion-dollar partnerships, they examine the cash flows underlying those arrangements.

Instead of accepting predictions about future productivity, they demand evidence of realized productivity.

The transition from narrative-driven valuation to cash-flow-driven valuation can be brutal.

Consider a hypothetical AI infrastructure business valued on the expectation that its revenue will grow 50% annually for years.

If revenue growth slows to 15%, the company may still be expanding rapidly by ordinary business standards.

But its valuation can collapse because the original price assumed something much greater.

The business does not need to fail for its shareholders to lose enormous amounts of money.

It merely needs to perform less spectacularly than expected.

This is why the AI bubble is so vulnerable.

The market has not simply priced in success.

It has priced in extraordinary success across multiple layers of the industry simultaneously.

GPU manufacturers must sustain enormous demand. Cloud providers must monetize their infrastructure. Model developers must maintain pricing power. Application companies must retain customers. Enterprises must realize measurable productivity gains.

Each layer depends on the commercial success of the next.

If enterprise customers fail to generate sufficient returns, application spending slows. If application spending slows, model revenue disappoints. If model revenue disappoints, demand for cloud computing weakens. If cloud computing demand weakens, infrastructure investment is reduced.

The financial consequences travel backward through the entire chain.

The market has constructed a towering pyramid of expectations on top of an uncertain foundation of end-user economics.

And pyramids of expectations are notoriously unstable.

12. The AI Bubble Is Not a Technology Problem. It Is a Mathematics Problem.

Perhaps the clearest warning came from Goldman Sachs' Jim Covello in 2024.

In discussing the enormous infrastructure expenditure required by generative AI, he posed a simple question:

"What $1tn problem will AI solve?"

Goldman Sachs+1

That question cuts through the industry's promotional machinery.

It does not matter how impressive a model appears in a demonstration.

It does not matter how many parameters it contains, how many GPUs trained it, or how enthusiastically technology executives describe its potential.

What matters is whether the economic value created exceeds the economic resources consumed.

If an AI system costs $10 million to build, operate, integrate, and maintain, it must create more than $10 million in risk-adjusted value to justify the investment.

If a data center requires billions in capital, it must generate sufficient cash flows during its useful economic life to recover that capital and produce an acceptable return.

If a foundational model developer spends billions improving performance, customers must be willing to pay enough for those improvements to cover the investment.

These are not complicated principles.

They are the foundations of business economics.

Yet the AI boom has repeatedly treated them as secondary considerations.

The implicit argument has been that artificial intelligence is so important that conventional financial discipline can be postponed.

But economic laws do not disappear because a technology is impressive.

Eventually, expenditure must be justified by returns.

Eventually, revenues must exceed costs.

Eventually, valuations must reflect sustainable earnings rather than speculative narratives.

And eventually, someone must pay for all those GPUs.

13. The Coming Great AI Reset

The bursting of the AI bubble will not necessarily resemble a single dramatic event.

It could unfold through a sequence of disappointments.

An enterprise software company reports weaker-than-expected AI subscription renewals.

A model developer raises additional capital on less favorable terms.

A cloud provider acknowledges that infrastructure utilization is below expectations.

A hyperscaler reduces its capital expenditure guidance.

A semiconductor manufacturer reports slower order growth.

Investors begin questioning the valuations of companies throughout the AI ecosystem.

What initially appears to be a temporary slowdown becomes a reassessment of the entire industry's economics.

The process could accelerate as the financial relationships supporting the boom begin to unwind.

Companies that raised money assuming continued access to inexpensive capital may find financing more difficult.

Startups without sustainable revenue may disappear.

Businesses dependent on subsidized inference pricing may discover that their unit economics were never viable.

Infrastructure providers may be forced to compete aggressively on price to maintain utilization.

The market may suddenly recognize that AI computing capacity, once perceived as an extraordinarily scarce asset, has become abundant relative to profitable demand.

This is how investment bubbles end.

Not because the technology has no value, but because the amount of capital committed to capturing that value exceeds what the market can economically support.

The ultimate beneficiaries may be the companies and consumers who acquire AI capabilities after the speculative premium has disappeared.

History provides a powerful precedent.

Following the dot-com crash, the internet continued expanding. Infrastructure became cheaper. Business models matured. Companies that survived or emerged afterward benefited from the investments made during the speculative boom.

The financial destruction helped establish the economic conditions for the next phase of technological development.

The AI industry may follow a similar path.

The technology will remain.

The inflated expectations will not.

Conclusion: The Greatest AI Illusion

The defining illusion of the AI boom is the belief that a revolutionary technology must necessarily produce revolutionary investment returns.

History demonstrates otherwise.

The railway transformed civilization while bankrupting investors.

The internet transformed the global economy while destroying trillions of dollars in speculative market value.

Telecommunications networks became indispensable even as the companies financing their expansion suffered enormous losses.

Artificial intelligence is now confronting the same economic contradiction.

The industry has committed extraordinary capital to building computational infrastructure before demonstrating that end customers will generate sufficient economic returns to support it.

It has built business models around expensive inference while competition drives prices downward.

It has financed increasingly costly foundational models while their capabilities become progressively more accessible.

It has assumed that enterprise adoption will translate into profitability, even as studies reveal the difficulty of converting AI experimentation into measurable financial results.

It has celebrated infrastructure spending as evidence of demand without adequately distinguishing investment activity from sustainable end-user revenue.

And it has attached extraordinary valuations to businesses whose competitive advantages may prove far less durable than investors imagine.

David Cahn's $600 billion question was not merely a question about revenue.

It was a question about whether the AI industry's financial architecture could support its own weight.

Goldman Sachs' trillion-dollar question was not merely about technological capability.

It was about whether the economic benefits could ever justify the cost.

Those questions remain at the heart of the coming reckoning.

The AI bubble does not need artificial intelligence to fail.

It only needs artificial intelligence to generate less profit than investors have already assumed.

And given the enormous disparity between capital expenditure, competitive pricing pressure, infrastructure depreciation, and demonstrated enterprise returns, that is a formidable risk.

The AI revolution may be real. The AI valuations may be fiction.

The world may eventually use artificial intelligence in virtually every business, every industry, and every household.

But that does not mean today's investors will recover the fortunes being spent to make it happen.

The dot-com crash did not mark the end of the internet.

It marked the end of the illusion that every company associated with the internet deserved an extraordinary valuation.

The approaching AI correction threatens to deliver the same lesson.

The technology will survive. The applications will improve. The infrastructure will find uses. The economic value will eventually emerge.

But the financial excess must first be unwound.

The AI bubble is about to burst. And when it does, the most expensive lesson will be the simplest one: technological progress does not repeal the laws of economics.

References