What past investment booms say about when funding stops, and what they cannot say about AI

Across five selected episodes from the 1840s to 2022, we found none in which a revenue shortfall clearly came first; in two, funding conditions clearly turned first. This is an order of dates in chosen cases, it does not show that funding ends booms, and the studies we read give no reliable end date.

This is not investment advice. It is not a forecast, and it offers no view on any company's shares.

Every few years a new technology draws a wave of money, and someone asks when it stops. The question put to us: given the history of booms, the dot-com years among them, when does investment in a new technology stop, and where is AI now, when the way companies raise money differs so much from one era to the next?

We split the work in three: five past episodes, where AI money comes from today, and what the research literature says about seeing an ending coming. Each part was pre-registered, meaning the method was committed before we looked at results. The short answer: history gives a pattern worth knowing, the present is changing, and the literature gives no date.

What five past booms show

The five episodes are the British railway mania of the 1840s, American railroad building in the 1870s, the American electrical and radio boom of the 1920s, the 1990s telecom and dot-com boom, and the 2021 run in growth technology, special-purpose acquisition companies (SPACs, listed shell companies that raise money to buy a business) and crypto, which unwound in 2022. For each we fixed a rule in advance for when new money stopped: the first period the funding series fell to half its peak or lower and did not climb back to 80% of the peak within two years. Then we asked which came first on the calendar, a change in funding conditions such as a rate rise or credit squeeze, or a shortfall in the revenue or technology the boom was built on.

In none of the five did a revenue or technology shortfall clearly come first. In the British railways and the 1920s, funding conditions turned more than fifteen months earlier (the 1920s case has no revenue stop in the boom sector itself, so it tests the order less strongly than the railways do). The 1873 American case reads as simultaneous or funding-first depending on whether short blips count, the dot-com case leans funding-first weakly (its count of new listings peaked in 1996, so that series dates the stop to 1998, two years before the Nasdaq peak, while the venture-capital series stops in 2001), and the 2021 case can come out any of three ways, so it is not robust. Some of these rules were set after we had seen the data, which the full report lists.

This is not the same as saying funding ends booms. Rate rises and credit squeezes also happen in the middle of booms that carry on. We did not count the times funding tightened and the money kept flowing, so the false positives are missing. All five were also picked because they are famous bubbles, so no base rate can be drawn. The order we found is an order of dates in selected cases, not a cause and not a frequency.

Two other findings matter more for the present. First, the lead time. From the first sign of tightening to the stop, the gap was roughly one to two years across the five, with readings ranging from about 5 months to about 26 months. That range depends on conventions: annual observations, which date counts as the stop, and, for 2021, whether only the 2022 rate rises count (the 5-month end, the reading under which the 2021 case becomes revenue-first). It is not a measured constant. The events later called decisive, such as the Jay Cooke bankruptcy in September 1873 or the crash of October 1929, fell inside the stopping period or just before it for most of the series we used. Almost all the signals were public at the time, but we picked them knowing how each episode ended.

Second, money does not stop all at once. In 2021 and 2022 the SPAC market stopped first, then initial public offerings, then crypto fundraising, then venture capital, spread over up to five quarters. In the British railway case, Parliament's approvals for new lines stopped in 1847 while construction spending halved three years later, in 1850. "When did investment stop" depends on which route you watch.

Where AI funding stands

We looked at four large American technology companies that are building most of the data centers (Microsoft, Alphabet and Amazon, which sell cloud services, and Meta, which does not), using their filings with the US Securities and Exchange Commission, and measured capital spending (cash paid for property and equipment, not counting finance leases) against operating cash flow.

From 2023 to 2025 that ratio stayed between 36% and 69%. In 2026 it rose to 86% in the first quarter, 96% in the second and 115% in the third. In the third quarter of 2026 the four spent about 201.9 billion US dollars on capital against 175.5 billion in operating cash flow, and their combined free cash flow was minus 26.4 billion dollars. It was the first quarter, of the fifteen we measured, in which spending exceeded operating cash flow.

Over the same period several outside-funding channels show growth: bonds, private credit, special-purpose vehicles (separate entities set up to hold a project's debt) and financing from suppliers to customers. Examples, each with its source: a vehicle managed by Blue Owl issued 27.3 billion dollars of secured bonds for a Meta data center campus in October 2025 (the amount is from the Bank of England's December 2025 Financial Stability Report; Meta's own announcement gives about 27 billion dollars as the campus's total development cost, a different figure); the Bank of England's July 2026 report puts AI issuers at 41% of US high-yield issuance that is not refinancing so far in 2026 (its wording is year to date); and an OECD estimate, as reported by the Bank of England, puts private credit's share of AI investment funding at 9% in 2024 and 34% in 2025. Nvidia's own quarterly filing shows 47.9 billion dollars of equity stakes, mostly in privately held companies, carried on its books at the end of July 2026, up from 3.4 billion in January 2025, and guarantees of up to 108.5 billion, of which 105 billion was signed in August 2026 for leases at a data center campus on behalf of an OpenAI affiliate. Sources disagree on totals, so these show direction, not a sum.

This differs from the 1996 to 2001 telecom boom in who the funders were. Telecom and cable companies borrowed in the high-yield bond and bank loan markets, about 192 billion dollars of high-yield debt from 1997 to 2000 (a figure from a data seller's 2002 report). Today's centre is established companies with large cash flows, whose spending has begun to exceed those cash flows while outside money has also grown; we did not measure how much of their spending the outside money paid for (their filings show about 225 billion dollars of gross debt raised by three of them: Alphabet and Amazon for January to September 2026, Meta for January to June only; it is gross, before repayments, and may include refinancing and short-term borrowing, so it is not net outside funding). We could not find telecom operators' operating cash flow, so the same ratio cannot be set side by side. Interest rates and the size of private credit also differ, so a category with the same name is not the same thing.

Three stages, overlapping

A sketch put to us: first anything labelled AI can raise money; then only good themes can; then investors ask about profits. We checked it quarter by quarter. The three can be seen at the same time, and our evidence does not support a clean sequence.

Selectivity cannot be dated to a quarter. Crunchbase's first release for the second quarter of 2024 said AI deals fell 15% from the previous quarter while the amount raised was a record, but Crunchbase later revised that quarter's deal count upward (1,001 to 1,211), the latest quarter is always undercounted when first published, and CB Insights reported the same quarter's AI deal count rising 16%. Dollar shares point the same way over a longer span: the share of AI funding going to foundation-model companies rose from 27% in 2024 to 40% in 2025 (our calculation). Scrutiny of profitability, on evidence from pages we read, shows up from the fourth quarter of 2025, or from the first quarter of 2025 (or even 2024) if search-result excerpts are counted. Because selectivity cannot be dated and scrutiny moves by one to six quarters depending on the evidence allowed, we cannot say which came first. The label premium could not be placed in a quarter, because there is only yearly data from search excerpts, and in that data AI companies were still valued at a premium in 2024 and 2025, alongside selectivity. Scrutiny has not strengthened in one direction either: in the third quarter of 2025 many shares rose after spending increases, and in the third quarter of 2026 reactions split about evenly.

Revenue figures need care

Reported revenue for the large AI model companies is mostly a run-rate, a recent period multiplied up to a year, and in the two cases below it comes from press reports citing unnamed sources, not from audited statements. OpenAI's figure, about 50 billion dollars a year, was reported by the Financial Times in October 2026 (read through Bloomberg Government), lower than some earlier reports of nearly 70 billion because of how sales through cloud providers are counted. Anthropic's figure, above 65 billion, was reported in August 2026 by The Decoder for the end of July 2026. Added together they are about 115 billion dollars a year, a sum of two unaudited figures from different months. Set against the four companies' annualised 2026 capital spending (January to September spending scaled up by four thirds), that is roughly 0.17. We treat 0.17 only as a rough comparison from unaudited press-reported figures of different months and different scopes, an order of magnitude and not a payback test. Published compute commitments by the model companies are 8 to over 100 times their annual revenue, but those are multi-year totals against one year of income.

What the research literature says

The third part read the main work on how booms end, from Kindleberger and Minsky to studies covering dozens of countries. Two results stand out, and neither is a date.

Probabilities can be raised over years. In 42 countries from 1950 to 2016, when credit and asset prices both surged over three years, the chance of a financial crisis within three years was about 40%, against about 7% in ordinary times. But the signal was a false alarm in 10.8% of non-crisis years, 36% of crises came with no signal, and the United States stayed under the signal for five years before 2007. Among 40 American industries, those that rose more than 100% in two years fell 40% or more within two years in 53% of cases, against 14% unconditionally. So 47% did not crash, and where they did, the peak came about six months after the signal, with a further 30% gain in between.

Dates are not predicted. The empirical studies reviewed here do not give a reliable end date, and we found none that tested whether a venture-capital boom's end can be forecast. Claims from log-periodic models of bubbles, by Didier Sornette and co-authors, who say they can date crashes in advance, remain contested: critics find the record rests on a few forward calls and unstable fits, and we did not evaluate that work. William Goetzmann's 2015 study of stock markets worldwide since the beginning of the twentieth century found that the chance of a crash after a boom is only slightly higher than the chance of a crash in general; in his abstract, the chance that a market gave back its gains following a doubling in value is about 10%. Most tested signals concern bank crises and listed shares. A boom financed by internal cash, bonds, private credit and large private rounds has not been tested against them. Nobody has tested whether "selective funding" is a sign that the end is near.

What we cannot know

We cannot say when AI funding will slow, or whether it will. We cannot say whether signals that worked for banking crises apply to this financing mix, or how often tightening was followed by no stop. Figures for private companies and the model companies are unaudited or come from data sellers who publish different totals for the same quarter.

What the evidence measured

These are the observables the literature's checklist names, and what we found for each. They are not predictions, no threshold here marks an ending, and nothing here is advice on what anyone should do.

Measured here: the four companies' capital spending against operating cash flow and free cash flow, for which we have no tested threshold; the growth of bonds, private credit, special-purpose vehicles and supplier financing, whose share of total spending we did not measure; and Nvidia's stakes and guarantees. Measured in the literature, not for AI: the signal of credit and asset prices both climbing quickly over three years, with its roughly 40% crisis rate and its false alarms, and the finding that among industries whose shares more than doubled in two years about half later crashed and half did not. Seen in the past episodes: funding routes stopping at different times, as in 2021 and 2022. Not measured by us: customer concentration among chip buyers, the working life companies assume for their servers, and audited revenue, since the revenue figures above are run-rates.

How this differs from our earlier note

Our earlier data note on four build-outs (Zenodo, 10.5281/zenodo.23861640) pooled January to September 2026 and found the four companies' capital spending at 0.997 of operating cash flow, just under its threshold of 1, so it did not call it a move to outside funding. Split by quarter, the ratio goes from 0.86 to 0.96 to 1.15, and the pooled figure hid the direction. That note's ratio of AI revenue to capital spending (0.23) used a different numerator and denominator from the 0.17 here, so the two should not be compared. That note found three of its four criteria undecidable and chose famous cases; this one also chose famous cases and gives no verdict on the present.

This article was researched and written by Claude (an AI model by Anthropic) in an automated Glasp research run: the three parts were pre-registered, then the public sources were read, the tables were built and the text was drafted. Company filings and central-bank data are primary where marked in the files; much else is secondary: press reports, data sellers' free summaries (Crunchbase, PitchBook and similar), and central-bank and academic papers read as working papers rather than the journal versions. Search-result excerpts support 117 of 844 funding rows and are marked as such. No Glasp user data was used. An earlier version was reviewed by two AI systems, Claude and OpenAI Codex, and revised in response; not every point was adopted, and the final text was not reviewed again. No human has checked the tabulations against the sources. Anthropic, one of the companies discussed here, makes the model that wrote this article. Nothing here is investment advice.