Why the AI Boom Isn’t Like Other Bubbles

The Washington Post put a sharper AI bubble argument on Google News on September 7, 2026, challenging the comforting parallels investors often draw with railroads and electricity. Financial columnist Matthew Lynn argues that today’s buildout differs in three critical ways: geographic concentration, extraordinary speed, and its potential effect on skilled employment.

That framing matters because the familiar historical analogy usually carries an optimistic ending. Railroads, electrical grids, and the internet attracted excessive investment, destroyed capital, and still left society with valuable infrastructure. Lynn’s column questions whether that sequence offers meaningful reassurance when one country and a few companies carry much of the exposure.

The concern is no longer limited to inflated technology stocks. The buildout now involves data centers, power systems, semiconductor supply chains, corporate bonds, private credit, and long-term capacity contracts. A pullback would therefore reach beyond AI laboratories or Nvidia shareholders.

The strongest conclusion is not that an AI crash has become inevitable. It is that historical comparisons can conceal the specific mechanisms that would shape one. The current boom combines useful infrastructure with uncertain returns, concentrated financing, fast hardware depreciation, and an unresolved labor transition.

What the Google News AI Bubble Argument Actually Changed

The Washington Post column changes the debate by attacking the reassuring part of the historical analogy, not the existence of an investment boom.

Published on September 7, the AI bubble argument accepts that railroads, electrification, and the early internet offer relevant comparisons. Each attracted speculative capital around technology that eventually became economically important. Each also produced losses when investment exceeded near-term demand.

The disagreement concerns what readers should infer from those precedents. A common interpretation says investors can lose money while society still wins. Failed railroad companies left tracks behind, and failed internet businesses left fiber networks that later supported valuable services.

Lynn argues that three differences make that ending less reliable for AI. First, much of the financial risk is concentrated in the United States, although China remains an important competitor. Earlier infrastructure waves spread construction and financing across more countries.

Second, the AI buildout has accelerated within less than five years. Rail and electrical networks developed over decades, giving operators, banks, and governments more time to identify weak business models. Today’s buyers often commit to chips, power, leases, and construction before demand becomes visible.

Third, generative AI is sold partly as a substitute for cognitive labor. Previous infrastructure waves displaced some occupations, but they also required large workforces to build and operate physical networks. AI vendors instead promise software that can perform parts of writing, programming, analysis, customer support, and administration.

Those differences do not prove that AI infrastructure lacks lasting value. They identify where the analogy stops carrying evidence. Useful technology can still produce poor investments, and an enduring network does not guarantee that its original financiers earn acceptable returns.

The column also shifts attention from stock prices toward the structure of the buildout. A stock-market bubble can deflate without disabling the underlying technology. A debt-supported infrastructure retrenchment can affect construction firms, utilities, private lenders, equipment suppliers, and regional economies.

Google News users encountering the opinion are therefore seeing a wider claim than “AI stocks look expensive.” The claim is that the economy has limited experience with this combination of concentrated capital, short investment cycles, and potential labor substitution.

That is a useful correction. The label “bubble” often compresses several separate questions into one word. Valuations can be excessive while adoption rises. Infrastructure can be necessary while capacity becomes oversupplied. Productivity can improve while individual firms fail to capture the gains.

A better test asks where the returns appear, who finances the assets, and who absorbs losses when forecasts miss. Those questions lead directly to the hyperscalers building the underlying capacity.

Five Hyperscalers Turned AI Into a Macroeconomic Bet

The boom feels different because Amazon, Alphabet, Microsoft, Meta, and Oracle have made corporate infrastructure spending large enough to influence national growth.

Federal Reserve researchers calculated that aggregate capital expenditure by those five companies reached $131 billion during the fourth quarter of 2025. It totaled $412 billion across that year, equal to about 1.31 percent of annualized US gross domestic product.

The Federal Reserve data excludes leases, so it does not capture every commitment attached to the buildout. That limitation is important because companies can secure data-center capacity through long-term contracts without recording the entire obligation as immediate capital expenditure.

The same research shows how financial expectations have concentrated around the companies supplying AI infrastructure. Between ChatGPT’s release and the end of 2025, Nvidia’s market capitalization rose 975 percent. Broadcom increased 636 percent, while AMD gained 179 percent.

Together, the three chip companies represented 11.2 percent of the S&P 500 at the end of 2025. Their collective share had reached 12.4 percent during October. These figures demonstrate concentration, although they do not establish overvaluation by themselves.

Investment is also flowing into the model developers that create demand for computing capacity. The Federal Reserve reported that OpenAI and Anthropic raised $58 billion and $44 billion, respectively, between 2023 and 2025. Their year-end 2025 valuations stood at $500 billion and $350 billion.

This creates a tightly connected market. Cloud providers invest in AI laboratories, laboratories buy cloud capacity, chipmakers supply the cloud providers, and expected model demand supports further data-center construction. Every transaction can reflect real commercial activity while the network still amplifies optimistic assumptions.

The Bank for International Settlements estimates that the five largest hyperscalers were set to spend more than $1 trillion on AI-related capital expenditure across 2025 and 2026. It also found that these commitments were growing faster than earnings and free cash flow.

That mismatch does not mean the companies face immediate insolvency. The largest hyperscalers own profitable advertising, commerce, software, and cloud businesses. Their existing cash flows give them more room than the telecommunications startups and speculative dot-com companies that depended heavily on external financing.

However, financial strength can intensify the competitive race. Each company fears that spending too little will surrender cloud customers, model leadership, developer loyalty, or consumer distribution. Rational defensive decisions by individual firms can therefore produce excessive capacity across the industry.

The BIS risk assessment describes this as a contest for a market that participants expect a small number of technical leaders to dominate. Under that logic, refusing to invest can look more dangerous than overbuilding.

The pressure target is clear. Chief financial officers must keep approving infrastructure before anyone can cleanly measure its return. Investors must decide whether cloud growth and advertising improvements justify spending that remains bundled inside broader business units.

Suppliers face a related risk. Semiconductor makers, electrical-equipment producers, cooling specialists, utilities, and construction contractors are expanding around hyperscaler forecasts. Their revenues can rise before the ultimate users of AI services demonstrate durable willingness to pay.

This is where the Washington Post thesis gains force. Earlier infrastructure booms also produced supply-chain dependence. Today’s version is unusually concentrated around several corporate budgets whose decisions influence public markets, private credit, power planning, and measured economic growth.

Real AI Adoption Makes the Bubble Question Harder

AI is not a purely speculative story, but genuine adoption cannot answer whether current investment will earn sufficient returns.

The strongest case against a simple bubble narrative is that people and businesses already use the technology. Federal Reserve data shows reported business adoption rising as survey questions and measurement methods evolved. Its revised series placed business adoption near 18 percent at the end of 2025, with planned adoption around 21 percent.

Work-related generative AI use among surveyed workers rose from 33 percent in August 2024 to 41 percent by November 2025. Professional services and finance recorded particularly high usage, while manufacturing showed strong year-over-year growth.

These are not the empty website visits that characterized many dot-com ventures. Workers use models to draft documents, summarize meetings, inspect code, answer support questions, retrieve information, and compare alternatives. Businesses are integrating AI into existing products and internal processes.

Knowledge workers also have more ways to connect models with their own material. A personal knowledge base can ground an assistant in documents, notes, and work history rather than relying only on a general model. That makes AI useful in repeatable workflows instead of isolated prompts.

Yet adoption and profitability are different measurements. A service can spread because vendors subsidize usage, bundle it into existing subscriptions, or accept low margins while competing for market share. High activity does not show whether revenue covers inference, training, depreciation, energy, and customer acquisition.

Inference means running a trained model to generate an answer. Every request consumes computing resources, although the cost varies by model, response length, hardware, and utilization. Falling inference costs can expand demand while also reducing the revenue earned from each unit of computation.

That tension makes AI different from physical infrastructure with relatively stable technical standards. A railway route can remain useful for generations. AI accelerators can lose economic value quickly when new chips deliver better performance or energy efficiency.

Microsoft illustrated the issue in its July 2026 earnings disclosures. The company said about two-thirds of its capital spending involved short-lived assets, primarily processors. Those assets require replacement more frequently than buildings, power connections, or land.

Rapid improvement helps users, but it complicates investment economics. A data center completed today can remain valuable while its installed hardware becomes less competitive. Operators must balance running older equipment longer against spending again to reduce operating costs.

Cloud financial statements provide only partial answers. Amazon, Alphabet, Microsoft, and Meta do not separately report the profits generated by AI infrastructure. AI revenue sits inside cloud divisions, advertising operations, productivity software, and other established businesses.

Reported cloud margins have remained healthy in several cases. That supports the view that the leading companies possess real revenue engines. It does not isolate the incremental return on the newest AI capacity.

The distinction matters because existing cloud workloads can make an AI-heavy division appear profitable. Conversely, current depreciation and construction costs can suppress margins before newly installed capacity reaches efficient utilization. Public reporting does not let outsiders fully separate those effects.

An August analysis found that AWS operating margins had climbed to roughly 39 percent, while Google Cloud reached 35.6 percent during the second quarter. Microsoft’s Intelligent Cloud margin remained near 41 percent. Executives still avoided attributing those results directly to AI.

The cloud profit gap remains the central unresolved issue. Analysts can observe capital expenditure, cloud growth, and backlog. They cannot yet calculate a reliable return for the AI-specific assets.

This is why calling the boom either obviously rational or obviously speculative misses the mechanism. Adoption validates the technology. It does not validate every data center, valuation, financing structure, or forecast built around it.

The New Risk Sits in Debt, Contracts, and Customer Concentration

The skeptical case becomes strongest where transparent corporate spending gives way to borrowing, off-balance-sheet commitments, and circular demand.

The first stage of the AI buildout relied heavily on hyperscalers’ operating cash. As spending moved beyond normal investment levels, debt assumed a larger role. The Bank for International Settlements reported that hyperscaler corporate bond issuance exceeded $100 billion in 2025.

Much of that borrowing carried maturities longer than five years. Long maturities can reduce immediate refinancing pressure, but they also lock companies into funding assets whose economic lives remain uncertain.

The financing extends beyond conventional bonds. Hyperscalers increasingly work with private-credit firms and special-purpose entities, which are legally separate vehicles created to own or finance particular assets. These entities can develop data centers and lease capacity back to a technology company.

In a typical structure, the hyperscaler owns a minority stake while committing to long-term leases or purchases of power and computing capacity. Private investors provide equity, and the vehicle raises additional debt.

The infrastructure financing can replace immediate capital spending with years of operating expenses. The BIS calls some of these obligations “shadow borrowing” because they behave economically like debt while remaining outside the hyperscaler’s main balance sheet.

Such arrangements are not inherently improper. They can distribute risk among parties suited to finance long-lived assets. They can also make total exposure harder for shareholders, lenders, and regulators to measure.

Banks may provide credit lines to the special-purpose entities. Insurers and private-credit funds may hold the debt. Hyperscalers may offer guarantees or contractual commitments that support the financing.

That network creates several potential transmission channels. A decline in expected AI demand can lower data-center values, weaken refinancing conditions, reduce private-credit appetite, and activate corporate guarantees. Problems can travel even if the largest technology companies remain solvent.

Customer concentration creates another uncertainty. Cloud companies count model developers among their fastest-growing AI customers. Those developers also receive large investments from the same infrastructure providers whose capacity they purchase.

This does not make the revenue fictitious. OpenAI and Anthropic serve real users and businesses. However, it means part of the infrastructure sector’s growth depends on customers that continue to require substantial outside capital.

The BIS has also highlighted circular financing across chipmakers, cloud providers, AI laboratories, and specialized cloud companies. An investor may fund a customer that then commits to buy the investor’s computing services or hardware.

Each agreement can serve a strategic purpose. Taken together, they make it harder to identify demand funded by end customers rather than capital circulating within the AI supply chain.

The risk rises when long-term commitments meet falling computing prices. If several operators build similar capacity, competition can push rental prices downward. Customers benefit, but asset owners may earn less than their financing models assumed.

Power constraints can initially hide that danger. Limited grid connections, transformers, advanced chips, and cooling equipment keep capacity scarce. Companies respond by reserving resources years ahead, which strengthens reported backlogs.

Scarcity does not guarantee lasting pricing power. Bottlenecks can encourage every participant to secure more capacity than it ultimately needs. When supply arrives, the market can shift rapidly from shortage to overcapacity.

The Washington Post argument should not be overstated here. Concentrated debt and contractual exposure do not confirm an approaching crisis. The public data remains incomplete, and the leading companies possess substantial cash-generating businesses.

The more defensible conclusion is narrower. The financial system is becoming exposed to AI through channels that stock valuations alone cannot reveal. Any serious bubble analysis must include bonds, leases, private credit, guarantees, and the quality of underlying customers.

Jobs Make This Boom Unlike Railroads and Electricity

The most consequential difference is that AI investment targets both infrastructure expansion and reductions in the labor needed for cognitive tasks.

Railroads changed employment patterns, damaged competing transport businesses, and created new geographic winners. They also employed large numbers of people to construct tracks, operate trains, maintain equipment, and manage expanding networks.

Electrification produced a similar mix. It displaced some technologies and occupations while supporting factories, appliances, utilities, and entirely new industries. The infrastructure required labor before its productivity benefits spread through the wider economy.

AI data centers also create construction and technical jobs. They require electricians, engineers, cooling specialists, security staff, maintenance workers, and energy projects. Equipment demand benefits industries far beyond software.

The difference lies in the product’s stated purpose. Generative AI is often purchased to complete parts of work previously assigned to writers, programmers, analysts, designers, researchers, and service employees. The infrastructure supports a system whose economic value can depend on lowering labor requirements.

That does not mean every deployment eliminates a job. Many applications assist workers instead. A developer can inspect unfamiliar code faster, while a researcher can search a large document collection without manually reviewing every file.

A product manager might use a weekly AI workflow to organize scattered notes and prepare a status update. The immediate result may be more complete work rather than a smaller team.

However, assistance can become substitution when companies redesign processes around it. If one employee handles more customers or projects, hiring demand can fall even when no current worker receives a termination notice.

This transition makes productivity data difficult to interpret. Higher output per worker can support wages and growth over time. It can also coincide with slower employment growth in occupations most exposed to automation.

The BIS observed that US sectors with greater AI exposure had recorded stronger productivity gains alongside weaker employment growth than other sectors. That relationship does not prove AI caused every difference. It does show why labor outcomes belong inside the investment debate.

Geographic concentration deepens the issue. The United States captures much of the market value associated with leading cloud companies and chip designers. It also carries substantial exposure if AI reduces demand for well-paid service work.

Other countries participate through semiconductor manufacturing, assembly, energy exports, and data-center construction. China competes across models, chips, applications, and infrastructure. Europe is using industrial policy to attract capacity and develop domestic alternatives.

Still, US equity markets carry an unusually large portion of global expectations. The BIS reported that US stocks represented about 64 percent of the MSCI Global index. A repricing could therefore transmit wealth effects well beyond American investors.

Labor displacement would follow a different path from a stock correction. Markets can adjust within days, while organizations adopt new workflows over years. A company might announce AI investment today, test tools next quarter, and change hiring plans much later.

That delay creates a measurement trap. Supporters can point to limited current layoffs as evidence that fears are exaggerated. Critics can interpret every hiring slowdown as automation, even when interest rates or weaker demand offer simpler explanations.

Readers should resist both claims. Reliable analysis requires occupation-level hiring, wages, hours, productivity, and task changes. Headline layoff counts alone cannot reveal whether AI complements workers or quietly reduces future openings.

The jobs question also changes the political response. Governments tolerated earlier investment busts partly because the remaining infrastructure supported broad future expansion. Public tolerance may be lower if taxpayers confront financial losses alongside concentrated employment pressure.

That does not guarantee restrictive regulation. It does mean policymakers will examine worker transitions, energy costs, local infrastructure, data rights, and market power together. The AI boom is not only a question of whether a new network gets built.

What Would Confirm or Weaken the AI Bubble Warning

Three signals will determine whether the Google News debate develops into a durable warning or fades as AI revenue catches up with spending.

The first signal is AI-specific revenue disclosure from the largest cloud providers. Investors currently receive capital-expenditure guidance, cloud growth rates, backlogs, and broad operating margins. They receive little direct information about returns on recently installed AI hardware.

Better disclosure would strengthen the optimistic case if revenue, utilization, and margins rise together. It would weaken that case if capital spending remains high while AI capacity runs below expectations or prices fall faster than costs.

The second signal is the financing mix. Hyperscalers can sustain aggressive investment more safely when established operations generate enough cash to cover it. Rising bond issuance, guarantees, leases, and special-purpose financing indicate that exposure is spreading.

The BIS found that borrowing already increased as expenditure outran normal investment. Future reports should clarify the size, maturity, counterparties, and guarantees attached to off-balance-sheet projects.

Transparent, long-duration financing backed by diversified customers would reduce near-term concern. A growing dependence on refinancing, concentrated tenants, or loosely disclosed guarantees would support the Washington Post warning.

The third signal is measured workplace adoption. The important question is no longer whether employees try generative AI. It is whether organizations convert usage into sustained productivity, revenue, better services, or reduced costs.

The global growth case notes that information-processing equipment and software investment grew 16.5 percent year over year during the third quarter of 2025. It also argues that conventional statistics can miss intangible benefits from data, algorithms, and organizational changes.

That measurement problem cuts both ways. GDP can record servers and software as investment before the resulting capacity produces enough valuable services. It can also understate benefits that appear through faster research, better decisions, or new products.

Over the next several quarters, credible productivity gains across many industries would strengthen the case for a lasting general-purpose technology. Gains limited to technology vendors and a few professional sectors would leave the concentration concern unresolved.

Readers should also distinguish between the boom surviving and every participant succeeding. Railroads, electricity, and the internet transformed economies despite bankruptcies and severe market corrections. AI can follow that broad pattern while producing very different winners, losses, and labor effects.

The Washington Post column is persuasive when it treats history as a warning against certainty. It becomes less persuasive if readers interpret difference as proof of collapse. Unprecedented features create uncertainty, not a predetermined ending.

For developers, the practical issue is whether demand for applications continues after subsidized experiments end. For enterprise buyers, it is whether AI systems deliver measurable value after integration, governance, training, and review costs.

Knowledge workers should watch how employers redesign roles, not only how many tools they purchase. Investors should examine cash flow, customer quality, depreciation, and financing commitments rather than relying on broad adoption claims.

Google News surfaced a debate that deserves more precision than the word “bubble” usually receives. The next test is evidence: clearer AI revenue, transparent financing, and productivity that reaches beyond the companies selling the infrastructure.

Ask those three questions whenever another record investment arrives. Who ultimately pays for the capacity, what measurable return does it produce, and how much of the demand depends on continued outside funding? The answers will show whether this buildout resembles a durable network, an expensive overcapacity cycle, or some uncomfortable combination of both.