One half of the market says artificial intelligence is the largest bubble in history. The other half says it is the largest breakthrough since the printing press. They are arguing about the same thing, demand, which is the side of the coin that can be talked up, financed into existence, and counted twice. This note turns the coin over. It reads the supply side, the capital actually being committed, through the framework that has quietly decided every technology mania for two centuries. The question was never whether AI is real. It is who floods in, who drowns, and who inherits.
The loudest argument in markets is whether artificial intelligence is a bubble. It is the wrong argument. Both sides are watching demand, how many people use the tools, how fast revenue grows, whether the models are good. Demand is the side of the coin that can be narrated, and it can be manufactured. The side that has decided the outcome of every technology mania for two hundred years is the other one, supply: the capital being committed to build. This year the largest technology companies are guiding to spend around 725 billion dollars on it. That number, and what history says happens next, is the subject of this note.
The easy read: either AI is a bubble that is about to burst, or it is a revolution that justifies any price. Pick a side and argue about demand.
Our reading: both can be true at once, and usually are. The technology can be genuine and world-changing while the capital financing it is destroyed. That is not a paradox, it is the capital cycle, and it is the normal way a transformational technology gets built. The useful question is not whether there is a bubble. It is where we are in the cycle, who is supplying the capital, and who will still be standing to inherit the assets when the returns on that capital collapse.
Nothing in this note is a price forecast, and nothing here disputes that AI is important. The figures are public: company capital budgets, national accounts, market prices, announced deals, and reported financials that we label as reported. What a research note can add is a framework old enough to have seen this before, and the discipline to apply it to the supply side while everyone else argues about the demand side.
Every coin has two sides. The demand side of a new industry is the one that gets talked about: the users, the growth rates, the total addressable market, the story of how big this could be. It is compelling, and it is exactly the side that can mislead. Demand is a narrative, and a narrative can be exaggerated, believed too early, and, as a later section shows, financed into looking real. The supply side is duller and harder to fake. Building capacity takes capital that has to be committed in advance, and that capital leaves a paper trail. Follow it and you get closer to the truth of what is happening than any adoption chart will take you.
The framework is not new. Its root is in the oldest text in economics. In The Wealth of Nations in 1776, in the chapter on the profits of stock, Adam Smith set out the whole mechanism in a sentence. When the stocks of many rich merchants are turned into the same trade, their mutual competition naturally tends to lower its profit; and when there is a like increase of stock in all the different trades carried on in the same society, the same competition must produce the same effect in them all. Smith's stock is capital. Translated out of the eighteenth century, he is saying that money flows toward wherever returns look high, that the flow brings competition, and that the competition drives those returns back down. High returns contain the seed of their own destruction, because they summon the capital that competes them away.
This idea, modernised by Marathon Asset Management and set out in Edward Chancellor's Capital Returns, is the capital cycle. It ignores the noise of demand and watches the supply side: where capital is flooding in, and where it is fleeing. Its central claim is contrarian and well evidenced. The best returns tend to come from industries capital is abandoning, and the worst from the industries it is rushing into, because supply, not demand, is what sets returns over a full cycle. When you hear that an industry is the future and watch capital pour in to prove it, the cycle says the returns are about to get worse, not better.
This is why serious investors watch capital spending, capacity and free cash flow rather than the size of the story. It is also why the AI debate, conducted almost entirely in the language of demand, is watching the wrong side of the coin. The rest of this note turns it over.
Start with the number, because it is large enough to be worth stating plainly and precisely. For 2026, the four largest hyperscalers, Amazon, Microsoft, Alphabet and Meta, are guiding to combined capital expenditure of around 725 billion dollars, up from roughly 410 billion in 2025. That is a rise of about 77 percent in a single year. The figure is the sum of the companies' own guidance, aggregated by the financial press, and it covers these four firms alone. Amazon is guiding to around 200 billion dollars, Microsoft to about 190 billion, Alphabet to roughly 185 billion, and Meta to some 135 billion. The figure excludes Oracle, Nvidia, CoreWeave and the rest of the field, so it understates the total capital going into the build-out.
To make the annual figure legible: 725 billion dollars is roughly 2 billion dollars a day, about 83 million dollars an hour. It is being spent on data centres, the chips inside them, and the power to run them, ahead of the revenue that is meant to justify it. That is the defining feature of the supply side of a boom. The people building the capacity spend first and hope the demand arrives later. They are the ones taking the risk.
The build-out is now large enough to be holding up the wider economy, which is what makes this cycle different from the three that precede it in this note. The Harvard economist Jason Furman, a former chairman of the Council of Economic Advisers, calculated from official national-accounts data that in the first half of 2025 investment in information-processing equipment and software was only about 4 percent of GDP, but accounted for roughly 92 percent of GDP growth. Strip those categories out, on his arithmetic, and the rest of the US economy grew at about a tenth of one percent, an annual rate that rounds to a standstill.
This cuts both ways, and it is worth being fair about it. That estimate is one economist's calculation from a particular slice of the accounts, and others, at JPMorgan and elsewhere, measure the contribution as smaller. Furman himself noted that some of the effect would be offset by the lower interest rates and energy prices a smaller boom would have allowed. But whichever number is right, the point holds: an economy that leans this heavily on a single category of investment has tied its growth to the fortunes of that investment. If the capital stops, the growth stops with it. That is precisely the exposure the capital cycle is built to think about.
The capital cycle moves an industry through four stages, and once you have seen it you will notice it everywhere. It is worth setting out cleanly before we watch it play out three times.
One, the flood. A new technology appears, the returns on offer look extraordinary, and capital pours in to chase them. Two, the boom. Money keeps arriving, competitors pile in faster than demand can grow, and the more everyone builds the more the returns begin to slip. Three, the collapse. The new supply finally crushes the returns that attracted it. Prices fall, companies fail, the industry consolidates, and a great deal of capital is destroyed. Four, the inheritance. This is where the durable money is made. The original builders go bankrupt, and someone walks into the wreckage, buys the assets for a fraction of their cost, and earns the return the builders paid to create. The assets survive. The people who financed them usually do not.
Every turn of the cycle is defended with the same phrase. John Templeton called this time it is different the four most dangerous words in investing, and Howard Marks has spent a career repeating the warning. Carmen Reinhart and Kenneth Rogoff took the phrase for the title of their history of financial crises, This Time Is Different: Eight Centuries of Financial Folly, precisely because the belief that the old rules no longer apply is the one constant across eight centuries of them. The price-to-earnings ratio used to be sixteen, the argument goes, and now it is thirty-two, and that is fine, because the technology has changed the world. The technology usually has. The arithmetic of competition and the cost of capital does not care.
It is worth saying why the technology being genuine does not save the investor. A real, transformational technology is exactly what makes the returns look high enough to summon the flood in the first place. As Carlota Perez documented across two centuries of technological revolutions, and as the record below shows, the more real and important the innovation, the larger the mania it tends to attract. A study of 51 major innovations between 1825 and 2000 found that 37 of them, nearly three quarters, were accompanied by a speculative stock-market bubble, and that the bubbles ran largest for the most radical and most visible technologies. George Soros framed the mechanism through his theory of reflexivity: a bubble does not grow out of thin air, it has a solid basis in reality, and that reality is then distorted by a misconception. The reality here is that AI matters. The misconception is that mattering and paying are the same thing.
Go back to the 1830s and 1840s. The first industrial revolution is in full swing in Britain, and the mines and mills need to move coal and iron faster than horses can carry it. Then the railway arrives and changes everything, and where the excitement is, the money follows. This is the flood. Capital poured into anything with the word railway attached to it, and not only from bankers. Shopkeepers, clergymen and widows put their savings into railway shares because no one wanted to be the person who missed the future.
At the peak in 1846, Parliament passed 272 separate acts authorising new lines, covering some 9,500 miles of track, and railway investment reached an estimated 7 percent of the entire British economy. Then came the boom in its literal sense. Companies stopped building the lines the country needed and started building whatever might push a share price higher. Routes were duplicated, lines were run to villages that would never fill a train, and the returns came in far below what the promoters had sold. By 1850 the ordinary railway share was paying a dividend under 2 percent, against the roughly 10 percent investors had been promised, and the railway share index had fallen from a peak near 1,985 in 1845 to about 675, a decline of roughly two thirds. Fortunes that had felt permanent were gone.
Then the fourth stage. The original investors were wiped out, but the rails they had bankrupted themselves laying were still in the ground. Over the following decades that network became the backbone of the British industrial economy, and the people who prospered from it were the ones who bought the assets cheaply after the crash, not the ones who financed them. The capital was destroyed. The railway remained. That is the pattern to hold in mind as the technology changes and the centuries pass, because it does not.
Jump to the late 1990s. After the United States rewrote its telecommunications law in 1996, capital flooded into fibre-optic cable on the belief that internet demand would be limitless. Carriers buried more than 500 billion dollars of it, on the order of 80 million miles of fibre, across the country and under the oceans. It was one of the largest build-outs in history, and it was justified by a demand story that turned out to be manufactured.
The story had a number attached. WorldCom told the market that internet traffic was doubling every hundred days, and the industry built for it. The mathematician Andrew Odlyzko, who actually measured the traffic, found it was doubling roughly once a year, a fraction of the claim. The capacity built for the myth had nowhere to go. When the dust settled, around 85 percent of the fibre in the ground sat unused, long-haul bandwidth prices fell about 90 percent, and the Nasdaq Composite fell close to 80 percent from its 2000 peak. Trillions of dollars of market value were destroyed.
The clearest way to feel the rhyme is to follow a single company. Corning, one of the largest fibre makers in the world, ran from about 109 dollars a share in 2000 to under 2 dollars by late 2002, a fall of some 98 percent. It then spent the better part of two decades in the wilderness. Over the past two years the same company has become one of the hottest names in AI, as a supplier of the optical connections that data centres need, and its stock has run to a new high above 250 dollars. The same company, the same shape, two bubbles apart.
The fibre bubble also rehearsed the specific trick this note returns to later. The equipment makers wanted to sell gear to a wave of new carriers that could not always afford it, so they lent the customers the money and booked the sales as revenue. By the end of its 2000 fiscal year Lucent Technologies had committed as much as 8.1 billion dollars of customer financing. When the carriers it had funded failed, Lucent took bad-debt provisions of 2.2 billion and 1.3 billion dollars across 2001 and 2002, and in 2004 the Securities and Exchange Commission charged it in connection with the improper recognition of roughly a billion dollars of revenue. Its rival Nortel followed the same path into one of the largest technology bankruptcies on record, and a leading networking supplier took a single inventory write-down of about 2.25 billion dollars when the cycle turned. As the activist manager Buxton Helmsley put it in a recent note, vendor financing did not merely add risk at the margin, it manufactured the appearance of end demand. Hold that phrase.
The most recent full turn of the cycle needs no century of hindsight. It happened just over a decade ago, in oil. Advances in hydraulic fracturing unlocked vast reserves that had been inaccessible, the returns looked extraordinary, and capital and drillers flooded in at once. US oil production roughly doubled over the shale era, and the United States became the largest oil producer on earth. Then the supply blew straight past demand.
Between 2014 and 2016 the price of oil collapsed from over 100 dollars a barrel to about 30. More than 200 North American oil and gas companies went bankrupt, the lenders who had funded them took heavy losses, and the cheap production capacity they left behind was picked up at a steep discount by the survivors. Railway Mania in the 1840s, fibre in the late 1990s, shale a decade ago: three centuries, three technologies, three entirely different industries, with nothing in common except the one cycle running underneath all of them.
Which brings us to the largest build-out yet, and the one the economy is leaning on. The capital is still flooding in, which places AI in the first two stages of the cycle, the flood and the boom, not yet the collapse. But the cracks that mark the transition are already visible, and they are the same cracks that showed in the fibre bubble: demand that is being manufactured rather than met, and returns that are starting to slip at the very centre of the boom.
There is one important difference from the three precedents, and it is not reassuring. When railways, fibre and shale collapsed, they took down their investors and their industries, painful but contained. The AI build-out, on the evidence of the previous section, is now entangled with the growth of the entire US economy. That raises the stakes of the collapse stage without changing the mechanism that leads to it. A bigger flood does not repeal the cycle. It just makes the inheritance larger and the collapse harder to absorb.
A bigger flood does not repeal the cycle. It enlarges the inheritance and the wreckage alike.
The fibre bubble faked demand with a story, the myth of traffic doubling every hundred days. This one is faking it with money. Over 2025 a web of very large deals was announced among the central names in AI, and if you follow almost any arrow in it, the money loops back to where it started. Nvidia commits to invest in OpenAI. OpenAI commits to buy compute from Oracle and CoreWeave. Those firms buy Nvidia chips. The capital makes its rounds and comes home to Nvidia, and at each step it is counted again as somebody's new demand.
| Announced arrangement | Headline size | Structure, as reported |
|---|---|---|
| Nvidia to OpenAI | Up to $100bn | Letter of intent, 22 Sep 2025, to deploy at least 10 gigawatts of Nvidia systems, paid in progressively |
| OpenAI to Oracle | ~$300bn | Five-year compute purchase reported Sep 2025, part of the Stargate build, beginning 2027 |
| OpenAI and AMD | Up to 6 GW | 6 Oct 2025, with a warrant for up to about 160 million AMD shares, near a tenth of the company, vesting on milestones |
| Microsoft and OpenAI | ~27% stake | Worth about $135bn after the 28 Oct 2025 restructuring, on roughly $13.8bn invested over time |
| CoreWeave and OpenAI | ~$22bn | Multi-year compute contracts; Nvidia owns about 6% of CoreWeave, OpenAI holds about $350m of its equity |
This is the vendor-financing trick from the fibre bubble, in a new costume. It can do a great deal. It can manufacture the appearance of demand, prop up revenue, and keep the chips moving. The one thing it cannot do is repeal Adam Smith. Capital chasing high returns still brings competition, and competition still compresses returns for everyone in the industry, however the capital is routed on its way in. Named analysts at Bernstein and Seaport have flagged the circularity directly; whether it proves benign or not, the accounting effect is real, that money moving in a circle can be counted several times as growth. When the circle stops, so does the growth it was manufacturing.
The place to look for the boom compressing returns is the eye of the storm. Two years ago OpenAI more or less was AI, and ChatGPT had no serious competition. Then everyone arrived, Google, Anthropic, Meta, xAI and a dozen others, all selling more or less the same thing. That is the boom stage on schedule, and its cost shows up in the accounts.
| Line, calendar 2025, $bn | Reported | Note |
|---|---|---|
| Revenue | 13.07 | from about 3.7 |
| Research and development | 19.18 | |
| Sales and marketing | 5.73 | from about 1.11, up over 400% |
| Total costs and expenses | ~34 | |
| Loss from operations | 20.92 | the cleaner measure of burn |
| Net loss attributable to the company | 38.53 | includes a large non-cash conversion charge |
Treat those figures with the caution they deserve, because they are a single leaked source rather than audited accounts. But the shape is what the cycle predicts. When competition arrives, the incumbent has to spend more and more simply to stand still, and the returns that drew everyone in begin to fall, exactly as they fell for railways, for fibre and for shale.
The final tell is how the market reacts to anything less than perfect. In June 2026 Broadcom, a chipmaker worth well over 2 trillion dollars, beat its earnings but merely reaffirmed rather than raised its AI forecast, guiding next-quarter AI revenue a few hundred million dollars below what the most optimistic analysts wanted. For a company of that size the shortfall is a rounding error. The reaction was not: about 286 billion dollars of market value vanished in a single session, among the largest one-day losses for any US company on record. When a rounding error erases a quarter of a trillion dollars in a day, the market is not pricing a healthy business. It is pricing perfection, and perfection is the one thing a company can always fail to deliver.
So who wins. There is an old line in business that the pioneers get the arrows and the settlers get the land. The first companies into a new market take the risk, the losses and the resistance, and the later arrivals walk in once the ground is cleared and quietly take the prize. You have watched this happen your whole life without naming it.
Take search. In the mid-1990s AltaVista, Lycos and Excite were the pioneers. They built the first crawlers, taught the world what a search engine was, and then buried their pages in banner ads to pay for it and let the product decay. Google walked in later, did not repeat their mistakes, and took the land. Or take ride-hailing. Before Uber and Lyft there was Sidecar, which spent its life being sued by cities and blocked by taxi commissions, fighting the legal battles that made the model possible, and went bankrupt doing it. Uber and Lyft copied the model, raised billions, and inherited the world Sidecar had cleared. The pattern is the capital cycle again: the first movers flood in, prove the market, and pay for it with everything they have, and the assets they built are inherited by whoever is still standing.
The inheritance in this cycle is unusually tangible. It is data centres, power contracts, fibre and chips, and the trained models and talent that outlast the companies that funded them. If the cycle runs its course, much of that capacity will change hands at a fraction of what it cost to build, and the firms that inherit it, some of today's names, some not yet formed, will earn good returns on assets somebody else paid for. That is not a reason to short the technology or to dismiss it. It is a reason to be clear-eyed about the difference between owning the future and paying for the privilege at the top of the flood.
Return to where we began. The whole argument online is about the demand side of the coin, whether the usage is real, whether the revenue justifies the hype. It is the loud side, because a story is what gets clicked, and the financial media will keep the audience staring at it for exactly that reason. But in markets the side that decides the outcome is rarely the loud one. It is the quiet mechanics underneath, the capital flowing through the supply side of the industry, that drive what happens.
Read that side and the online fight looks like the wrong fight. Bubble or not was never the right question, because a transformational technology and a capital bloodbath are not alternatives, they are the normal joint outcome, and the capital cycle is the process that produces both. The right questions are the ones the cycle asks. Who is flooding in. Whose returns are being competed away. Who is manufacturing demand to keep the flood going. And who will still be standing, with the cash and the discipline, to inherit the assets when the returns on all this capital collapse.
The question was never whether AI is real. It is who floods in, who drowns, and who inherits.
The reading behind the argument, on the capital cycle, technological revolutions and financial manias, the historical episodes drawn on, and the figures on the current build-out.
Data as of 8 July 2026. Capital-expenditure figures are company results and guidance aggregated by the financial press; 2026 is guidance. The growth-share figure is an economist's calculation, not an official statistic. OpenAI figures are third-party reported and unaudited. Market series are month-end from public providers. Figures are rounded for readability.
This document has been prepared by Iron Hall Capital for informational and educational purposes. Its content does not constitute personalised investment advice, a recommendation to buy or sell financial instruments, a public offering, or a solicitation to subscribe to any financial product. The opinions and readings reflect Iron Hall Capital's judgement at the date of publication, are based on data considered reliable but not independently audited, and may be revised without notice.
Market figures are drawn from public data. Equity, index and commodity prices are month-end series from public providers. Capital-expenditure figures are company results and guidance, aggregated by the financial press; guidance is a projection, not an actual, and 2026 figures are forward guidance that has been revised repeatedly. The share of United States growth attributed to information-processing investment is an economist's calculation from official national-accounts data, not an official statistic, and estimates of it vary widely with method. OpenAI's 2025 figures are as reported by a third party and verified by the Financial Times; they are leaked and unaudited, and the company has not published them. Announced financing arrangements are frequently described as up to a stated amount, are multi-year and conditional, and should not be read as cash already deployed. Historical episodes are drawn from the cited scholarship and the public record. This note was prompted by a widely circulated video essay making the capital-cycle argument; it takes that framework and tests it against the record. It contains no price forecast. Past behaviour and past statistical relationships do not guarantee future results, and markets can move sharply and without warning.
The author and Iron Hall Capital may hold, have held, or come to hold positions in the instruments referenced. Any reproduction, in whole or in part, requires written authorisation.
Iron Hall Capital · A private investment office · July 2026