A RAAK.work project
Capex WatchThe BIS sets the AI build-out beside four earlier investment booms. This page puts numbers on the comparison. Against the size of the economy, the AI spending has passed the telecom build-out of 2000 and is still far below Britain’s railway boom of the 1840s. It has grown more in two years than dot-com investment did in five. And most of it buys equipment that lasts about five years, where a railway line lasted a century.
The BIS names four earlier booms and one thing they share. Each was built on a real invention, and each drew more money than the invention could repay.
The BIS’s 2026 annual report sets the AI build-out beside the US canal boom of the 1830s, the British railway boom of the 1840s, the electrification boom of the late 1920s and the dot-com boom of the late 1990s. In each, it says, a real technical advance attracted more capital than its commercial returns could justify. Investment then went into reverse, and the reversal brought a recession (Chapter I, pages 22–23). The front page carries this in one sentence, in §5’s fork.
The report compares the booms on one measure: how many times yearly investment multiplied from its low point before each boom (its Graph 11.C). §3 redraws that measure for the two booms whose figures can be checked against official records today. §2 adds a second measure that the report does not use: the size of the investment against the size of the economy. That second measure is this site’s addition, not the BIS’s.
The AI spending has passed the telecom build-out. It is below the dot-com peak on the five’s own figures, and well below the largest railway boom.
Each bar is one year’s investment at a boom’s peak, as a share of what that country’s economy produced in the same year (its GDP). The last two bars are today.
The bars for the United States from 2000 onwards are from official accounts. The two railway bars are historians’ and analysts’ estimates, because nobody measured an economy’s output in the nineteenth century; the figures were reconstructed later.
How each bar was worked out.
| Boom | Investment in the peak year | The economy’s output | Share |
|---|---|---|---|
| British railways, 1847 | £44M | about £600M | 6.7–7.3% |
| US railroads, 1870s and 1880s | no agreed figure | no agreed figure | 2–6% |
| US telecom build-out, 2000 | $142.1B | $10.25T | 1.4% |
| US computers and software, 2000 | $450.6B | $10.25T | 4.4% |
| US computers and software, 2025 | $1,404.0B | $30.86T | 4.5% |
| The five AI builders, 2026 plans | $842.5B | $30.86T (2025) | 2.7% |
British railways. Two historians’ estimates, and they are close. Andrew Odlyzko puts railway investment in 1847 at about £44M, against an economy of about £600M, which is 7.3%. An older estimate by B. R. Mitchell, cited by Gareth Campbell, gives 6.7% for the same year. The bar is drawn solid to the lower figure and shaded to the higher.
US railroads. There is no agreed figure, and the bar is a range for that reason. Goldman Sachs puts the peak of the railroad boom at 2% to 3% of GDP. A figure of 6% for 1870, taken from a 1960 study by Melville Ulmer, is the one most often quoted beside AI. The two differ by a factor of two or three. Read the bar as “somewhere in here”, not as a measurement.
US telecom. From the US national accounts: what businesses spent in 2000 on communications equipment ($123.3B) and on communications structures ($18.8B). The equipment alone is 1.2% of that year’s output, which is the figure usually quoted for the telecom boom.
US computers and software. Also from the national accounts: private investment in what the statisticians call information processing equipment and software. This is the series the BIS uses for the dot-com boom. It is the only bar that can be drawn for both booms on one definition, and on it the United States is now above the dot-com peak: 4.5% of output in 2025 against 4.4% in 2000.
The five AI builders. The $842.5B is what Alphabet, Amazon, Meta, Microsoft and Oracle have said they will spend in 2026, the same total the tracking page counts. It is set against US output in 2025, the last full year.
What the bars cannot settle. Three cautions.
The five spend around the world, and the bar divides by the US economy alone. It also divides this year’s plans by last year’s output. Both make 2.7% a little too high as a share of the United States.
The two bars for today count different things. The official series counts every US business’s computers and software, whether or not they have anything to do with AI, and leaves out buildings. The five’s figure counts chips, buildings and power equipment, in any country. They overlap, and neither contains the other.
A peak year is not the whole boom. Stijn Van Nieuwerburgh’s paper, on the reading page, counts every AI data centre in the United States over eight years and arrives at 3.6% of US GDP a year. The paper sets that above the canal, railway, electrification, highway and telecoms booms. A boom that holds a lower level for longer can outweigh one that touched a higher peak for a single year.
The five’s spending has grown more in two years than dot-com investment did in five.
This is the BIS’s own measure. Take the low point before a boom and call that year’s investment 1. Then follow how many times it multiplies. The BIS counts the dot-com boom from 1995 and the AI boom from 2023, and so does the chart.
The open circle is a plan, not a result: the five’s own figures for 2026, on their own definitions. The filled points are cash spent, from each company’s accounts. Neither line is adjusted for inflation.
Dot-com. US investment in computers and software was $253.8B in 1995. It peaked five years later at $450.6B, which is 1.8 times the 1995 level, and then fell for two years.
AI. The five spent $154.2B in 2023, $239.1B in 2024 and $411.5B in 2025. That is 2.7 times in two years. Their plans for 2026 come to $842.5B, or 5.5 times in three. The vertical scale of the BIS’s own graph stops at 4.
What this cannot settle. The comparison is kind to the older boom. Five companies can grow faster than a whole country’s investment, because a country’s total includes every firm that did not join in. The BIS sets the same two series side by side, so the chart follows it. A fairer match would be the dot-com boom’s own five largest spenders, and this page has not built that.
Three things repeat: a real invention, a race, and a forecast stretched from a short record.
A real invention. The railways worked. So did the internet. The BIS’s point is that a boom can end badly for its investors while the technology does everything that was promised. Whether AI works and whether the spending pays are two questions, and history answers them separately.
A race. The BIS says the AI spending may be driven partly by the belief that only a few firms will end up with most of the market. Each firm then spends more than it would alone, for fear of being left out (Chapter I, page 22). The telecom build-out ran the same way. The Federal Reserve Bank of Richmond’s account records that the four established long-distance carriers owned 72% of America’s long-haul fibre in 1996 and 30% in 1999, because new firms had laid their own networks beside them.
A forecast stretched from a short record. Throughout the telecom boom, the Richmond Fed records, carriers’ executives said that internet traffic was doubling every hundred days, and fibre was laid on forecasts of that kind. Odlyzko later worked out what the claim implied: internet use multiplying 17 million times between 1994 and 2000. He finds the same habit in Britain’s railway boom of the 1830s, whose promoters reasoned from a couple of successful early lines. The forecast to watch today is on the reading page: by the account of Apollo’s chief economist, lenders are counting on the cash the five’s businesses bring in rising from $600B in 2025 to $2T in 2030.
Three things do not repeat: the equipment wears out, the spenders began rich, and one earlier boom paid.
The equipment wears out. This is the largest difference, and it is the reason the front page has a revenue floor at all. When the British railway boom ended, shareholders lost heavily. But by Odlyzko’s account hardly any of the lines went bankrupt and almost all stayed in service, and he thinks it quite likely that the country gained more as a user of the railways than its shareholders lost. America had far more long-haul fibre than it needed by late 2000, the Richmond Fed records, but a cable in the ground still works when the traffic arrives years later. An AI chip is out of date by then. The front page spreads the spending over five years because that is roughly how long the equipment stays useful. So a bust in this boom would leave less behind, and the bill comes round again every few years whether or not the revenue has arrived.
The spenders began rich. Britain’s railways were paid for by private investors, and much of America’s long-haul fibre by new firms. The five paid for the first years of the AI build-out from the profits of their existing businesses. That has started to change. The strain chart shows which of them now spend more than their businesses bring in, and the reading page sets out how much they are borrowing.
One earlier boom paid. Britain had a smaller railway boom in the 1830s, with investment of about 2% of GDP a year at its height in 1838 and 1839. Odlyzko finds that its investors were earning more than the market rate within a few years, even though costs ran to twice the plan. So the rule is not that booms fail. An investor is repaid when the revenue arrives before the money runs out, and not otherwise.
The usual ending in history is a version of the front page’s bust: the invention worked, its users gained, and the first investors lost.
That ending is the first branch of §5’s fork, with one detail added. The revenue does not fall short because the technology fails. It falls short because the people who use the technology keep most of what it is worth to them. The front page has a name for this: the capture rate in §4, the share of a customer’s gain that reaches whoever sells the AI. In Britain’s railway boom, on Odlyzko’s reading, the users did well and the shareholders did not.
The short life of the equipment closes the way out that earlier booms had. A railway that took twenty years to fill was still there when the traffic came. Equipment that lasts five years has to earn its cost within five, and then be bought again.
Which ending this boom is heading for is a matter of record, not of history. Tracking the bet keeps that record: the revenue arriving against the revenue required, results by results.