Productive Bubbles

Why Capital Arrives Before Demand — and Why That Is Not Always a Mistake

Executive Summary

This paper advances a contrarian but historically grounded proposition: the most transformative periods of wealth creation in modern economic history have almost always been preceded by financial excess. Railways, electricity, telecommunications, the internet, and now artificial intelligence (AI) each attracted capital at levels that looked irrational against near-term demand. Investors were routinely burned. Valuations collapsed. Companies disappeared by the hundreds. And yet, in every one of these cycles, society kept the infrastructure — the tracks, the grid, the fiber, the data centres — long after the capital that built it had been written off.

The distinction this paper draws is between productive bubbles, which leave behind durable capability, and extractive bubbles, which mainly reprice assets that already existed or, worse, were never real in the first place. Railways and fiber optic networks are productive. Property speculation, Samuel Insull's leveraged utility pyramids, and Enron-style financial engineering are extractive. The difference is visible only in hindsight, which is precisely why it is so often missed in real time — and why this paper anchors its 2026 analysis on the one institution whose job is to see it early.

In its Annual Economic Report published June 28, 2026, the Bank for International Settlements (BIS) — the central bank for the world's central banks — made this paper's exact argument independently: "The canal mania of the 1830s, the British railway mania in the 1840s, the electrification exuberance of the late 1920s... and the dotcom boom of the late 90s all shared one common trait: a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify." The BIS now applies that same template to artificial intelligence, estimating that the five largest hyperscalers will spend over $1 trillion on AI infrastructure between 2025 and 2026 alone, and naming an AI capex bust as one of the three most significant threats to global financial stability.

The current AI investment cycle — approximately $725 billion in combined 2026 capital expenditure guided by the four largest United States (US) hyperscalers, up 77% from roughly $410 billion in 2025, financed in part by $159 billion of hyperscaler bond issuance in the first five months of the year alone — exhibits every classic marker of productive-bubble behaviour: infrastructure spend running far ahead of monetized demand, intense competitive one-upmanship, opaque circular financing between chipmakers and AI labs, and open institutional debate about whether the economics will ever close. History suggests this pattern does not automatically signal failure. It may instead be the visible signature of a new layer of economic infrastructure being built before the applications that will justify it have even been invented.

For operators, investors, and policymakers, the practical question this paper answers is not "is this a bubble?" — the BIS, among others, now treats that as close to settled — but "what kind of bubble is it, and what should we do differently because of that answer?"

The $725 Billion AI Bet

In 2026, Amazon (~$200B), Microsoft (~$190B), Alphabet ($175–185B) and Meta ($125–145B) are together guiding to approximately $725 billion in capital expenditure — a 77% increase over 2025's already-record $410 billion, in Goldman Sachs credit strategist Amanda Lynam's description, "a staggering 77%" jump — almost entirely directed at data centres, graphics processing units (GPUs), custom silicon, and the power infrastructure needed to run them. Add Oracle's roughly $50 billion, and Goldman Sachs Research's June 2026 "Tracking Trillions" report projects the AI infrastructure build growing from $765 billion this year to $1.6 trillion annually by 2031 — some $7.6 trillion of cumulative capital across compute, data centres and power over that span. Meta itself raised its 2026 guidance mid-year, citing rising component and data-centre costs, and its stock fell roughly 6% on the news — a reminder that even the bulls are surprised by how fast the number keeps moving.

This is not conventional software investment, where a company spends against a defined, forecastable revenue line. It is capital committed on the scale and logic of national infrastructure: railways, electrical grids, and telecommunications backbones — and, tellingly, it is increasingly financed the way infrastructure has always been financed: with debt. The five largest hyperscalers issued about $159 billion in corporate bonds in the first five months of 2026 alone, more than their combined borrowing over the previous five years, according to Dealogic; Meta's $30 billion October 2025 offering was the largest investment-grade deal of the year, alongside a roughly $27 billion off-balance-sheet financing vehicle. Nvidia itself returned to the bond market for the first time since 2021, raising $25 billion. Trailing four-quarter capital spending across the four hyperscalers reached $433.9 billion against only about $149 billion of reported depreciation — assets being bought at nearly three times the rate the income statement currently recognises them, a gap this paper returns to directly in the section on the revenue deficit.

Demand-side signals are genuinely strong, not manufactured: Microsoft's commercial remaining performance obligations reached $627 billion, nearly double the prior year, and its AI business alone crossed a $37 billion annualised run rate, up 123% year-on-year; Google Cloud grew 63% in the first quarter of 2026. Contracts tied to Anthropic and OpenAI now make up roughly half of the combined ~$2 trillion revenue backlog sitting across Amazon, Microsoft, Google and Oracle. But a backlog is a contractual promise, not delivered cash, and the central question for the sector remains the one every infrastructure financier before it has faced: will usage arrive, at the pace being underwritten, before the debt comes due?

The Revenue Deficit Paradox

Every general-purpose technology revolution passes through a period in which infrastructure investment outruns realized revenue. Railways faced it in the 1840s, when average dividends on ordinary shares came in at 1.83% by 1850 against promoter promises north of 10%. Samuel Insull's electrification empire faced it in the early 1930s, when banker-driven disputes over depreciation accounting made his solvent utilities look, on paper, like they were losing money. Telecom carriers faced it violently at the turn of the millennium, when roughly 95% of installed fiber sat dark by 2001. AI is facing a version of the same paradox now: trailing four-quarter capital spending across the four largest hyperscalers reached $433.9 billion against only about $149 billion of reported depreciation, and investor Michael Burry has argued publicly that extending the assumed useful life of GPUs — hardware that turns over on a two-to-three-year product cycle — could understate the industry's true depreciation expense by roughly $176 billion between 2026 and 2028.

On the demand side, the picture is more mixed than either bulls or bears prefer to admit. A Massachusetts Institute of Technology (MIT) Project NANDA study found that only about 5% of integrated generative-AI pilots were extracting measurable value at the profit-and-loss line as of mid-2025, against $30–40 billion in enterprise spending; S&P Global found 42% of companies abandoned most of their AI initiatives in 2025, more than double the prior year's rate. Set against that, Anthropic's revenue run rate moved from roughly $9 billion at the end of 2025 to approximately $47 billion by mid-2026 — a fivefold increase in about five months — and AI-related investment is estimated to have driven roughly three-quarters of US gross domestic product (GDP) growth in the first quarter of 2026. Both data sets are real. The paradox is that a technology can be simultaneously failing most of the organisations piloting it and driving the majority of a Group of Seven (G7) economy's headline growth, because the value is concentrated in a small number of applications riding on infrastructure built for many more.

The instinctive response to a revenue deficit of this size is to treat it as proof of a mistake — to assume that infrastructure should only be built once demand for it is already visible and monetizable. History argues the opposite, and the Bank for International Settlements now argues it too: general-purpose technologies are, almost without exception, built ahead of the use cases that will eventually justify them, because the use cases themselves depend on the infrastructure existing first. A revenue deficit is uncomfortable for the companies financing it and can be fatal for individual investors caught in the eventual repricing. It is not, on its own, evidence that the underlying technology lacks long-run economic value — but nor is it, on its own, proof that this time the deficit will close before the debt comes due.

Railways: The Original Productive Bubble

Britain's Railway Mania of the 1840s remains the clearest template for what a productive bubble looks like from the inside. Parliament authorized thousands of miles of track backed by capital that, at its peak, represented a share of national income comparable to some of the largest infrastructure booms in history. Much of the projected traffic never materialized on the routes as originally planned. Speculative promoters floated lines with no realistic prospect of profitability, share prices collapsed through 1847 and 1848, and a large share of investors lost money outright.

What survived the crash was the network itself. Once built, track could not be un-built; freight and passenger traffic grew into the capacity that speculative capital had provided years ahead of need. Manufacturers gained access to national rather than regional markets, freight costs fell structurally, and the railway network became a platform for the rest of the industrial revolution to run on. The lesson that recurs in every subsequent cycle is visible here first: the people who financed the infrastructure and the people who captured the long-run value it created were, in large part, different people.

Electricity: Building the Grid Before the Appliance

Electrification followed the same sequence in reverse order from how it is usually told. Utilities built generation and transmission capacity before most households or factories had any specific use for it beyond lighting. Samuel Insull's holding-company model financed grid expansion across the American Midwest years ahead of proven consumer demand, betting that abundant, cheap power would create its own appliances, its own factories, and its own industries once it existed. It did — but only after the grid was already in place.

Had electrical utilities insisted on demand fully justifying supply before building a single additional substation, electrification of the industrial economy would have proceeded far more slowly and far more expensively. Instead, the infrastructure came first and abundance did the rest of the work: falling per-unit costs enabled experimentation, experimentation produced the washing machine, the refrigerator, and eventually the entire apparatus of twentieth-century industrial manufacturing. This is the pattern most directly analogous to AI compute today — capacity being built on the belief that cheap, abundant intelligence will generate its own applications once it is available at scale.

Telecoms and the Internet: Capacity Before Demand

The telecom build-out of the late 1990s is the closest historical mirror to the current AI cycle, and the most instructive because of how badly it appeared to fail in the short term. Carriers including WorldCom and Global Crossing laid hundreds of thousands of miles of fiber optic cable across oceans and continents, financed by debt markets that assumed internet traffic would grow fast enough to fill it. It did not — not on that timeline. By the early 2000s, a large share of that fiber sat dark and unused, WorldCom's accounting fraud became one of the largest corporate collapses in US history, and telecom-sector losses in the crash ran into the hundreds of billions of dollars.

The fiber itself did not disappear. Over the following decade, the same "dark" capacity that critics cited as proof of waste was lit up, sold cheaply, and became the physical substrate for search, e-commerce, social media, video streaming, and eventually cloud computing. The companies that built the capacity were, in most cases, not the companies that ultimately profited from it. The capital that was destroyed was financial; the capital that survived was physical, and it proved to be worth far more than the balance sheets that had financed it ever recovered.

Netflix and the Infrastructure-First Adoption Model

Netflix is frequently told as a story of visionary strategy — a DVD-by-mail company that correctly anticipated streaming and pivoted ahead of its competitors. The more accurate version of the story is an infrastructure-first one. Telecom carriers did not lay fiber in the 1990s because they foresaw Netflix; they laid it because they believed internet traffic in general would grow. Netflix's 2007 streaming launch was only viable because a decade of speculative bandwidth overbuild had already collapsed the cost of moving video data close to zero.

Netflix, in other words, was a beneficiary of infrastructure financed by someone else's bubble, not the cause of it. This is the pattern strategists most consistently underweight: the company that ends up capturing the value of a technology wave is very often not the company — or even the industry — that financed the infrastructure underneath it. For any business watching the AI infrastructure build-out today and wondering where the applications will come from, Netflix is the reminder that the applications typically arrive after the infrastructure is already cheap, not before.

Why Usage Often Follows Infrastructure Rather Than Precedes It

Conventional business planning assumes that demand should be validated before supply is built — that a company should not lay capacity it cannot presently sell. General-purpose technologies routinely violate this assumption, and the sequence that replaces it is consistent across every cycle examined in this paper: new infrastructure lowers the marginal cost of using a capability, lower cost enables experimentation that would previously have been uneconomical, experimentation surfaces applications nobody had planned for, and those applications generate the demand that retroactively justifies the original investment.

This sequence explains railways, electrification, the commercial internet, and cloud computing. It plausibly explains AI as well. The most consequential AI applications of the early 2030s are unlikely to be visible today, precisely because they depend on a cost structure — cheap, abundant, reliable inference at scale — that does not yet exist and is only being built now. Judging the current infrastructure spend by today's visible use cases risks repeating the exact analytical error telecom critics made in 2001, when the internet's eventual applications were still five to ten years from being invented.

AI: The Next Infrastructure Layer?

The spending behaviour of Microsoft, Google, Amazon, and Meta increasingly resembles the construction of a utility rather than the launch of a product line, and its executives now describe it in exactly those terms. Satya Nadella has called the buildout "a full-on upgrade of the global tech stack," arguing Microsoft's Azure business today is supply-constrained rather than demand-constrained — the opposite problem from a product nobody wants. Jensen Huang told the Consumer News and Business Channel (CNBC) in February 2026 that the industry's capex trajectory is "justified, appropriate and sustainable" because hyperscaler cash flows are rising in step with it, a genuine point of difference from 1999, when most dot-com infrastructure was financed against businesses that had no cash flow at all.

Whether this framing is ultimately correct remains genuinely uncertain, and this paper does not claim otherwise. What can be said with more confidence is that the spending pattern itself — concentrated capital expenditure on compute, power, and physical facilities, run years ahead of confirmed unit economics, increasingly financed with long-dated debt against short-lived hardware — matches the productive-bubble template of every prior general-purpose technology wave examined here closely enough that the Bank for International Settlements now cites it explicitly alongside canal mania, railway mania and the dot-com boom. Power, not chips, has become the binding constraint on how fast this layer can actually be built: the largest US grid operator, PJM Interconnection (PJM), saw its most recent capacity auction clear at a record $333.44 per megawatt-day, US interconnection queues now run a median of over five years, and all four major hyperscalers have signed nuclear power deals totalling more than 9.8 gigawatts precisely because gas and grid capacity cannot be secured fast enough otherwise. If AI capability does become a pervasive, embedded layer of economic activity in the way electricity and connectivity did, the 2025–2027 capital expenditure cycle — and the power infrastructure it is dragging into existence alongside it — will likely be remembered as the construction phase of that layer, not as a cautionary tale about overinvestment.

The Builders and the Beneficiaries

A pattern repeats across every cycle in this paper closely enough to be treated as a rule rather than a coincidence: the companies that finance transformative infrastructure are rarely the companies that capture the largest share of the value it eventually creates. Railway companies enabled industrial manufacturers who never laid a mile of track. Telecom carriers who nearly collapsed under fiber debt enabled internet platforms built a decade later on capacity those carriers had already written off. Cloud providers who spent the 2010s building hyperscale data centres now compete with, rather than automatically capture, the software ecosystems those data centres made possible.

Applied to AI, this raises an uncomfortable but strategically important possibility for the hyperscalers currently financing the build-out: the largest beneficiaries of today's roughly $725 billion capital cycle may be companies that do not yet exist, built by founders who have not yet started, on top of compute that is cheap only because someone else overpaid for it first. The 2026 evidence already shows both edges of this blade — inference cost per million tokens fell by an estimated 80% between 2023 and 2025, which is exactly the kind of cost collapse that made Netflix possible after the fiber crash, but it is also destroying the margins of thin, wrapper-layer AI startups now failing or being acqui-hired at an estimated rate of 70–90% within eighteen months of founding. For operators and investors alike, this argues for distinguishing between exposure to the infrastructure layer and exposure to the eventual application layer — they are not the same bet, and 2026 is already showing they will not be won by the same companies.

Productive vs Extractive Bubbles

Not all financial excess behaves the same way after the correction, and the distinction is the organizing idea of this paper. Productive bubbles finance the creation of new physical or technical capability — track, grid, fiber, data centres — that continues to generate economic value long after the equity that financed it has been marked to zero. Extractive bubbles, by contrast, mainly bid up the price of assets that already existed, or worse, were never real: land, existing securities, or claims on future cash flow that were never backed by new productive capacity in the first place.

The test is straightforward to apply after the fact and considerably harder to apply in real time: what is physically left over once the financing collapses? Railways left railways. Telecoms left fiber. The AI build-out, whatever else happens to the equity valuations financing it, is leaving behind data centres, nuclear and gas power plants, and chip-manufacturing capacity that will outlast any individual company's balance sheet. But 2026 has already supplied a clean extractive counter-example inside the same sector: Builder.ai, a Microsoft- and Qatar Investment Authority-backed startup once valued near $1.3 billion, collapsed into insolvency after it emerged that its "AI-powered" app-building platform was substantially operated by roughly 700 human engineers, with revenue allegedly overstated by as much as 300%. Productive and extractive bubbles are, in 2026, forming inside the very same investment wave — which is precisely why the distinction has to be drawn security by security, not sector by sector.

Technology vs Commodity vs Property

Technology bubbles, property bubbles, and commodity booms are frequently discussed as though they were variations of the same phenomenon, but they behave differently at the level that matters most for long-run economic outcomes. Technology bubbles tend to expand what an economy is capable of doing — new capacity, new processes, new categories of activity that did not previously exist. Property bubbles primarily redistribute ownership of a fixed or slow-growing stock of assets at inflated prices; when they correct, the buildings remain, but the leverage and ownership structure underneath them does not.

Commodity booms sit closer to property than to technology: they generally reflect temporary scarcity pricing rather than the creation of new productive capability, and the infrastructure built to extract a commodity has value contingent on that commodity's price recovering, which is a narrower bet than infrastructure with general-purpose applicability. AI capital expenditure, for all its excesses, sits firmly in the technology category — it is building general-purpose compute and power capacity whose eventual use cases are not yet fully known, which is precisely the characteristic that has made prior technology bubbles productive rather than merely destructive. The International Monetary Fund's (IMF) Tobias Adrian has flagged the one respect in which this cycle behaves more like property than prior technology waves: a "potential maturity mismatch between the duration of the physical assets and the duration of the debt" now financing them, since GPUs depreciate on a two-to-three-year cycle while the bonds raised to buy them often run ten years or longer.

Enron: A Control Experiment

Enron is the useful negative case in this analysis — a company that positioned itself publicly as an infrastructure and innovation story, attracted enormous investor enthusiasm on that basis, and then collapsed in 2001 in what was, at the time, the largest corporate bankruptcy in US history. The critical difference between Enron and the productive bubbles examined elsewhere in this paper is what was actually left standing afterward. Railways left railways. Telecom carriers left fiber. Enron's collapse revealed that a large share of its reported value had never corresponded to physical assets or productive capacity at all — it was accounting structure, off-balance-sheet entities, and governance failure dressed in the language of infrastructure and innovation. Samuel Insull's electrification empire, though built on genuinely productive assets, failed by the same accounting mechanism a decade earlier: a pyramid of holding companies in which roughly $100 million of his own capital controlled some $2.5 billion in underlying assets, a leverage ratio of about 25 to 1 that made the entire structure collapse the moment confidence in any single layer cracked.

Enron and Insull are a useful discipline for evaluating any current investment cycle, AI included: the question is never simply whether a sector is attracting enthusiastic capital, but whether that capital is visibly converting into durable physical or technical capacity, and how transparently it is financed along the way. Data centres under construction, nuclear power-purchase agreements, and chips being fabricated are independently verifiable in a way that Enron's reported earnings never were. But the Bank for International Settlements' 2026 warning about "circular financing" — chipmakers and hyperscalers taking equity stakes in AI labs that in turn commit to multi-year purchases of that same chipmaker's hardware, on terms the BIS itself describes as "typically poorly disclosed, with risks of the same asset being pledged multiple times" — is close enough to the Insull pattern that it deserves to be watched with the same seriousness, not dismissed as mere market chatter.

Two Futures for AI

Two broad scenarios are visible from today's vantage point, and this paper deliberately does not attempt to pick between them — though the institution best positioned to judge has come closer to a view than most. The Bank for International Settlements' June 2026 Annual Economic Report warns explicitly that "disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust, with potential knock-on effects on financial conditions," and points to early warning signs already visible: credit-default-swap spreads on AI-related debt beginning to widen, and outstanding AI-related private credit — a market that produced roughly $3 billion in loans in 2010 and over $40 billion in 2025 alone — now projected to reach $300–600 billion by 2030. In this first scenario, capital expenditure growth slows from its current 77% year-over-year pace, some valuations reset meaningfully, and a number of the ventures currently competing for compute and talent do not survive as independent companies; the June 2026 semiconductor selloff, which erased over $1.4 trillion in chip-sector market value in a matter of weeks, offered a preview of how fast that repricing can move once it starts.

In the second scenario, investment continues at or near current levels, enterprise monetization catches up to infrastructure spend over the following several years — a trajectory Anthropic's run rate growth from roughly $9 billion to $47 billion in five months makes at least plausible for the strongest players even if the median enterprise pilot continues to fail — and the sector avoids a sharp correction entirely. What is notable is that both scenarios converge on a similar physical outcome: an enormous stock of data centre, power, and chip-manufacturing capacity remains in place regardless of which financial path the sector takes. For operators and policymakers, this suggests the more useful planning question is not which scenario will occur — a forecast nobody, including the BIS, claims to be able to make reliably — but how to ensure the infrastructure that survives either scenario is positioned to be useful, accessible, and economically productive once the financing story around it has been resolved one way or the other.

Conclusion: Over-funding the Future

The pattern this paper traces across two centuries — railways, electrification, telecom and internet fiber, and now artificial intelligence — is that societies sometimes become durably wealthier precisely because investors, acting under conditions of excess optimism, fund infrastructure earlier and at greater scale than sober near-term analysis would ever justify. The result each time has been financial waste at the level of the individual investor and acceleration at the level of the economy as a whole. Railways arrived faster than a cautious capital market would have built them. Electrification spread faster. The commercial internet scaled faster. Artificial intelligence, on the evidence assembled in this paper, appears to be following the same path.

The task for operators, investors, and policymakers is not to prevent bubbles from forming — that has never been achieved in any prior cycle and is unlikely to be achieved in this one. It is to correctly distinguish productive bubbles from extractive ones while they are still forming, to position capital and strategy on the infrastructure and application layers accordingly, and to ensure that whatever survives the eventual correction is captured as durable, productive national and organizational capacity rather than written off as simple excess.

Key Frameworks

Framework 1: Productive Bubbles Create New Capacity. — Judge a financial excess by what remains once the capital that financed it is marked to zero — physical and technical capability, not valuation.

Framework 2: Infrastructure Frequently Precedes Demand. — General-purpose technologies are built ahead of the use cases that justify them; the use cases depend on the infrastructure existing first.

Framework 3: Builders and Beneficiaries Are Often Different. — The company that finances a technology wave's infrastructure is rarely the company that captures most of its eventual value — plan exposure to each layer separately.

Framework 4: Corrections May Destroy Valuations While Preserving Assets. — A financial correction and an infrastructure failure are not the same event; distinguish which one you are actually observing before reacting to it.

References

• Perez, Carlota. Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages. Edward Elgar Publishing.
• Schumpeter, Joseph A. Capitalism, Socialism and Democracy. Harper & Brothers.
• McKinsey & Company. Research on AI adoption, enterprise return on investment, and productivity impact.
• Bank for International Settlements. Commentary on AI investment concentration and infrastructure financing risk.
• JPMorgan. Analysis of AI infrastructure economics and hyperscaler revenue requirements.
• Financial Times / hyperscaler Q1 2026 earnings compilations; company capital-expenditure guidance, Amazon, Alphabet, Microsoft, Meta (2026).

Abbreviations

AI Artificial Intelligence

BIS Bank for International Settlements

CNBC Consumer News and Business Channel

GDP Gross Domestic Product

GPU Graphics Processing Unit

G7 Group of Seven

IMF International Monetary Fund

MIT Massachusetts Institute of Technology

PJM PJM Interconnection (largest US grid operator)

US United States