A Hundred and Fifteen Percent
Britain’s railways changed the country, and many of the people who paid for them were disappointed. In 2026 four American companies spent $115 on buildings and equipment, largely for AI, for every $100 their businesses brought in. So who is paying this time?
№ 015November 20269 min readPedrum Shokouhi
Turner’s Rain, Steam and Speed, 1844: the Great Western crossing the Thames at Maidenhead, a year before the crash there.
Maidenhead · Whitehall
30 November 1845Technology can change the world and still disappoint the people who paid to build it. Britain’s railways did exactly that in the 1840s. This essay asks whether the same thing is happening with AI. Who is putting their money at risk to pay for it? And who ends up absorbing the loss if the returns fall short? To see how the railway story went, start with a Sunday night in November 1845, when Britain was in the grip of a railway craze.
It was Sunday the 30th of November 1845. A railway could not be built without an Act of Parliament. The plans for every line proposed for the next session had to reach the Board of Trade, a government office in Whitehall, by midnight. A party of solicitors, as British lawyers are called, hired an engine to carry theirs to London. It broke down at Maidenhead on the Great Western line. A second special train, carrying more lawyers, ran into it and smashed the carriage to pieces. Nobody died. The Railway Chronicle, which could not resist the scene, reported that the survivors shook themselves off, packed up their ruffled plans, collected the stranded attorneys and steamed on to arrive in good time.
Ten express trains ran on the Great Western alone that day. Hiring horses and a carriage cost two guineas a mile (a guinea was a little more than a pound). One group of promoters was refused a train by the established company they meant to compete with, so they loaded their plans and a clerk into a hearse and sent the hearse to London by special. By nine that night the agents were pouring down Whitehall. By eleven the clerks had fallen behind the arrivals. Nearly a quarter of an hour after the deadline, a hired carriage pulled up with three men who had left Harwich that morning. They had lost more than an hour and a half being driven around Pimlico by a driver who could not find the office. Inspector Otway told them it was too late. So they threw their plans through the open door, broke the passage lamp, and had the plans thrown back into the street. The projects lodged that year came to 815, against 248 the year before. Britain had caught railway fever, and almost none of the people racing to London that night were the ones paying for the lines.
Emily Brontë
Haworth, West RidingNone of the men on those trains was paying for any of it. The solicitors make the better story, so the people with the money tend to get crowded out of it. The money for Britain’s railways came from ordinary people making private decisions. One of them was sitting that winter in a parsonage (a vicar’s house) at Haworth in Yorkshire. Her name was Emily Brontë, and she would go on to write Wuthering Heights. She was twenty-seven and ran the sisters’ small investments. What she did on that Sunday nobody recorded, but we know how she worked, because her sister Charlotte described it. She had taught herself the subject by “carefully reading every paragraph & every advertisement in the news-papers” about railways.
Most of the sisters’ money went into the York and North Midland, one of George Hudson’s lines, and it was paying ten percent. Two months after the night in Whitehall, with their money still intact, Charlotte told Margaret Wooler that she wanted “to be wise in time” and sell. Then she added that she would sooner risk the loss than hurt Emily’s feelings. So they held. The shares had been issued at £50 and had reached £120. By the end of 1849 they stood at £20, with the next dividend, the payout to shareholders, in doubt. Emily had been dead for a year. She was exactly the kind of person who paid for the railways: a private investor, putting her own money at risk, with nobody to absorb the loss if the returns fell short.
Now the same question about AI. The number that brings Emily Brontë into this essay sits in a table on my NVIDIA page. Microsoft, Alphabet (Google’s parent), Amazon and Meta (Facebook’s parent) are building the huge data centers that AI runs on. Every quarter they report two numbers that matter here. One is the cash their businesses bring in from operations. The other is what they spend on property and equipment, such as buildings, servers and chips. In the third quarter of 2026 the four together spent $201.9 billion on property and equipment, against $175.5 billion of operating cash flow. That is 115%. Put plainly, for every $100 their businesses brought in, they spent $115 on buildings and machines. It comes to $26.4 billion more than they earned in three months, before paying a dollar of dividends to their shareholders.
Amazon ran at 182%, Alphabet at 121%, Meta at 104%, and Microsoft, the outlier, at 68%. A year earlier the group’s figure was 61%. In the twelve months between, spending rose 108% while cash flow rose 10%. Back to the end of 2016, as far as the four line up quarter by quarter, the ratio had never gone above 100%. Its previous high was 96%, set one quarter earlier. The measure is crude. These are the cash purchases from the four cash flow statements, which leave out equipment they pay for through long-term leases. The arithmetic underneath it is plain all the same, and a gap like that gets filled by somebody.
Spending past the cash line
Capital spending (buildings and equipment) as a share of operating cash flow (the cash the businesses earn), Microsoft, Alphabet, Amazon and Meta combined, by calendar quarter
How they are paying
Form 10-Q, September 2026The companies that were buying their own shares a year ago are now selling new ones, and that reversal is most of the answer. Meta bought back $26.2 billion of its own stock in the first nine months of 2025 and none in the same months of 2026. On 26 October it launched an offering of about $35 billion of common and preferred shares. It also said it plans to sell about $60 billion more, gradually, on the open market. That would be in a new kind of share, Class C, that carries no vote. Meta’s own quarterly filing, the 10-Q, says this class could prove dilutive to all stockholders, meaning each existing share would own a smaller slice of the company.
Alphabet made the same turn, from $39.8 billion of buybacks in those months of 2025 to none. It then went further the other way. It sold $42.9 billion of common stock, including a $10 billion private placement with an affiliate of Berkshire Hathaway. It sold $19.1 billion of mandatory convertible preferred, a kind of share that must eventually turn into ordinary shares. On top of all that came $85.2 billion of debt. Amazon borrowed: its long-term debt went from $65.6 billion to $173.1 billion in nine months, and it expects cash capital spending to “meaningfully increase in 2027.” Microsoft stands apart from all of this. It sits at 68%, has $18.6 billion of free cash flow (the cash left after paying for equipment), and has issued no bonds for AI as of this week. On the ratio it looks like the sober one.
Promises to rent
$1.0 trillionMicrosoft’s lease note says otherwise, and it changes what the 115% means. A lease is a promise to rent something for years, here mostly data centers. At 30 September Microsoft had $364.6 billion of leases that had not yet started, a figure already on my Microsoft page. Meta had $355.7 billion, Amazon $164.9 billion and Alphabet $130.5 billion. Added together, loosely, because the four do not disclose these on identical terms, that is about $1.0 trillion. That is five times the quarter’s spending, most of it for capacity that somebody else is building or financing. The ratio counts what the companies pay for. It leaves out what they have promised to rent.
The promises go further than rent. Meta’s Hyperion campus in Louisiana shows how this works. A Blue Owl entity called Beignet Investor sold $27 billion of bonds to investors, and Meta’s rent covers the payments on them. If Meta walks away and the property sells for less than the bonds, Meta makes up the difference. Across its data center ventures that sit outside its own accounts, Meta’s 10-Q puts the maximum on such promises at $41.5 billion.
The same pattern shows up elsewhere. Alphabet discloses $75.7 billion of backstops (promises to step in if a payment falls short) tied to data centers, leases and their energy costs included. Of that, $48.0 billion is signed with data center providers and $27.7 billion is an obligation to put more in place. In August NVIDIA entered into guarantees capped at $105 billion, covering part of the lease and power payments at SB Energy’s campus in Pike County, Ohio. An affiliate of OpenAI is the tenant, and the first phase is expected in fiscal 2029. And Moody’s counted $85.2 billion of AI-linked investments at US life insurers at mid-year, up 24% in six months, while warning that the figure probably understates the exposure. Whoever holds an annuity (a retirement product sold by insurers) from one of those insurers is, in a small way, Emily Brontë without the newspapers.
Andrew Odlyzko
Collective hallucinations and inefficient marketsWhat happened to the funders in the 1840s is well documented, and it is more specific than “they lost money.” Andrew Odlyzko, in a long study of those investors, puts railway investment at 7.3% of British GDP in 1847. The new lines cost about 50% more to build than projected. Working expenses (the cost of running the trains) took about half of revenue where the promoters had expected 40%. And revenue came in 30 to 40% short.
The damage did not fall evenly. Bond and preference investors, who were paid first, did very well in almost all cases, and so did shareholders of the lines the big companies bought or leased. The loss landed hard on ordinary shareholders, the ones at the back of the queue, who had no means to put in more money, the Brontës among them. Hardly any of the lines went bankrupt, and almost all stayed in service. Odlyzko thinks it quite likely that the gains to users outweighed the losses to shareholders. That is a polite way of saying that Britain got its network on the account of whoever stood last in line. The Economist delivered the verdict in 1855: mechanically an honor to the age, and “commercially, they are great failures.”
Run that test on 2026 and the guarantees start to read differently. A bondholder in a structure like Beignet has Meta standing behind the shortfall. So the risk of an empty building has moved off the lender and onto Meta’s owners. The same goes for NVIDIA’s owners and Alphabet’s. Every backstop in those filings is a small transfer of risk onto the ordinary shareholder, which today means nearly anyone with a broad American index fund.
Nor is the preferred share the refuge it was in the 1840s. Meta’s new 6.50% issue looks like the safe seat, but it converts into common stock by force. Below $625 a share, each $1,000 of it becomes 1.6 shares. So at $465, the lowest price in the prospectus’s illustrative table, the holder has $744, not $1,000. Alphabet’s 6.25% issue works the same way. Shareholders of the York and North Midland could at least form a committee of investigation, which they did. The buyer of a non-voting Class C share has no say in how the company is run, and at a company whose founder controls the vote will have far less leverage. Of everything in this comparison, that is the difference I find hardest to argue away.
30 October 2026
Nasdaq down 2.86%The market spent October asking its own version of this, in its usual way, and this week it raised its voice. Alphabet fell more than 12% on Thursday, taking half a trillion dollars off its value, on worries about search advertising and the rise in AI spending. On Friday the 30th the Nasdaq dropped 2.86% and the S&P 500 1.82%, each the worst day since June. The Nasdaq closed out its worst month since March 2025. Amazon reported free cash flow of negative $42.4 billion over twelve months, which means it spent $42.4 billion more than its operations brought in. Microsoft, the company with the largest stack of unstarted leases, was down just 4% from its high, which tells you what investors are looking at. I would not read much into any of it. A week of prices is a mood, and the filings will take years to reach a verdict.
The best answer to all this is that the York and North Midland had nothing but its trains. These four generated $175.5 billion of cash in a single quarter from businesses that existed before any of this began. Microsoft reports $677 billion of contracted commercial revenue and AWS, Amazon’s cloud business, $578 billion. The railway promoters had only their traffic forecasts. A group with that much cash can run past 100% for a good while.
I would take that more seriously if the contracts were less tangled. AWS added $100 billion from OpenAI and more than $100 billion from Anthropic this year, and Amazon is an investor in both. The asset is also shorter-lived. A Victorian cutting has outlived everyone who paid for it. Amazon shortened the life of a subset of its servers from six years to five, citing the pace of AI, and Meta uses five and a half. The buildings and the power connections will last far longer than the chips. The model on my Amazon page assumes that depreciation, the accounting charge for equipment wearing out, takes the AWS margin from about 38% to below 26%. That is a guess with decimal places.
Users of this capacity come out ahead on almost any path. A build financed at a loss still leaves the capacity behind, as it did in Britain. The people with something to find out are the funders. At 115% they are no longer only the ones who chose to own these companies. More of the risk now sits with people who hold an annuity or an index fund and will never open a 10-Q. Emily at least read the advertisements. Trains still run out of York on the 1840 line the sisters’ shares were a claim on, and nobody aboard has ever been sent the bill. That went to Charlotte, the last of the sisters, with her shares at 20, and to a good many others whose names nobody kept.
Notes
- The Maidenhead collision, the Great Western expresses, the hearse, the scene in Whitehall and Inspector Otway are from John Francis, A History of the English Railway, vol. II (1851), pp. 246–253, which reprints the Railway Chronicle and other contemporary reports. The count of projects is from Frederick S. Williams, Our Iron Roads (1852), pp. 57–58.
- Charlotte Brontë’s April 1845 letter, her 1849 letter on the York and North Midland shares, and the sisters’ holdings are quoted in Andrew Odlyzko, Collective hallucinations and inefficient markets: The British Railway Mania of the 1840s (preliminary version, January 2010), pp. 4–5, 29, 80 and 148; his note 4 records that her shares recovered to about 55 by mid-1854. Her letter of 30 January 1846 is in Elizabeth Gaskell, The Life of Charlotte Brontë (1857).
- Capital spending and operating cash flow are from each company’s Form 10-Q for the quarter ended 30 September 2026 and SEC XBRL data: Microsoft $39.4B and $58.0B, Amazon $77.3B and $42.4B, Alphabet $51.4B and $42.3B, Meta $33.9B and $32.8B, the last two calculated as nine months less six. Calendar quarters; Microsoft’s is its fiscal first quarter. Percentages are calculated from those figures. The series runs from the fourth quarter of 2016 with one quarter, the second of 2017, missing from the data; the chart starts after that gap.
- Financing figures are from the same 10-Qs. Meta’s offering terms are from its free writing prospectus of 26 October 2026 and prospectus supplements of 28 October 2026; the $465 is from the illustrative table there, which assumes no Class C Trigger Event.
- Leases not yet commenced are from the lease notes and commitments tables of the four 10-Qs, which do not use identical terms; Microsoft says some arrangements depend on conditions being met. Beignet Investor is described in NPR’s The Indicator, 25 November 2025. Alphabet’s backstops are from the liquidity section of its 10-Q. NVIDIA’s guarantees are in its 10-Q for the quarter ended 26 July 2026. Moody’s figures are as reported by Reinsurance News, 23 October 2026.
- Railway costs, revenues and the fate of bond, preference and common holders are from Odlyzko, pp. 25, 34 and 80, which also carries the 1855 Economist line (p. 5). The York and North Midland’s Committee of Investigation is on p. 148.
- Market figures for 29 and 30 October 2026 are closing figures from CNN Business, updated 5:35 p.m. ET on 30 October. That Microsoft had not issued bonds for AI spending is from CNBC, 27 October 2026.
- Server lives are from the 2025 Forms 10-K of Amazon and Meta. The state of the York and North Midland line is from Northern Trains’ timetable N34 (York to Sheffield via Pontefract Baghill, 17 May to 12 December 2026), which shows passenger trains leaving York over the 1840 route.