Railway King's Dividends and AI Capex · Issue #190

Opening

In 1848, a slim book appeared in London. Its title was The Bubble of the Age; its subtitle was “The Fallacies of Railway Investment, Railway Accounts, and Railway Dividends.” Its author, Arthur Smith, was neither a famous economist nor a politician. He was simply someone who dug all the way through other people’s company accounts.

When people introduce this book today, they usually call it “an 1848 warning about the railway bubble.” I think that is the biggest possible misunderstanding of the book.

Smith never denies railway technology, not once. The trains were actually running, and fare revenue was real. What he would not let go of was exactly 1 thing: which account the costs were entered into.

Here is the conclusion up front. This book was not a bubble warning but an accounting indictment, and the debate now unfolding around AI infrastructure investment is using almost the same sentences as 178 years ago.

🔍 1848, an accountant’s indictment

Smith took apart the accounts of the London and North Western Railway, or LNWR. The company’s main line had already been opened several years earlier: the Liverpool-Manchester section in 1830, and the Manchester-Birmingham section in 1842.

But something was strange. On a line that had already been built, nearly 1,000,000 pounds kept flowing into the capital account¹ every year. In just the most recent 2 years, the figure was 1,983,472 pounds. Of that, 489,589 pounds was spent in the very half-year the company itself described in its report as “a half year of extreme depression and unprecedented difficulty.”

What did that mean? Smith writes:

“Railways can only be worked by a constant addition to the capital account, greater than the dividends declared.” (Railways could keep operating only by continuously adding to the capital account more than the dividends they declared.)

In other words, the costs of wear, tear, and replacement were being pushed not into the revenue account, but into the capital account. On the income statement, that makes the company look profitable. With that profit, it declares a dividend; because a dividend appears, the share price rises; because the share price rises, it can issue new shares; and with the money paid in for those new shares, it pays the next dividend.

Smith compresses the structure into 1 sentence:

“As soon as calls cease to be paid and loans to be made, from that period also cease the payment of dividends.” (The moment additional calls² stop being paid in and borrowing stops, dividends stop too.)

And here is the sentence in the book that held me the longest:

“What information, the single fact, that no Company has closed its capital account, should convey to the cautious.” (How much this 1 fact alone should tell a cautious person: that no railway company has ever closed its capital account.)

This was not an asset you built once and were done with. It was a permanent capital intake valve. Yet the companies were accounting for it as if it were a “one-off construction cost.”

Smith does not stop there. He also digs into transactions among directors. Hull & Selby Railway had a stock-market value of less than 250,000 pounds, yet the North Midland board guaranteed its purchase for 2,000,000 pounds. The Great North of England agreed to buy a company with a market value of 438,900 pounds for 4,000,000 pounds. Smith’s wording is exact:

“Directors in one Company have purchased, guaranteed, and leased their own property in another.” (Directors in 1 company bought, guaranteed, and leased their own assets in another company.)

There is 1 name this book repeats: George Hudson, the man then known as the “Railway King.” Smith singled him out in 1848, and in 1849, the very next year, Hudson collapsed after it was revealed that he had paid dividends out of capital.

📊 2026, the books open again

Now let us come back to the present.

Start with scale. Microsoft, Alphabet, Amazon, and Meta spent roughly $433.9 billion (4,339 × $100 million) on tangible assets over the most recent 4 quarters through Q1 2026. Over the same period, they recognized about $149.0 billion (1,490 × $100 million) in depreciation. Their combined capex in Q1 2026 alone was $129.8 billion (1,298 × $100 million), up 80% year on year.

Here is the question Smith would have asked immediately. Why spend $433.9 billion (4,339 × $100 million) while recognizing only $149.0 billion (1,490 × $100 million) of depreciation?

Part of the answer lies in useful life³. Between 2022 and 2023, Amazon, Alphabet, and Microsoft extended the accounting useful life of servers and networking equipment from the previous 3-4 years to 6 years. If you assume a longer asset life, annual depreciation falls, and reported profit rises by the same amount.

How large is the difference? According to a Goldman Sachs sensitivity analysis, cutting GPU useful life from 5 years to 3 years would raise cumulative depreciation from about $3 trillion to $4 trillion between 2026 and 2031. $1 trillion is at stake. Michael Burry has estimated profit overstatement at more than $17.6 billion (1,760 × $100 million) from 2026 to 2028.

But we need to be honest here. These figures are scenarios calculated by changing assumptions, not settled facts. And within the industry, opinions differ on whether the real life of a GPU is 3 years or 6 years. What is interesting is that the direction has started to diverge. Amazon, from January 2025, moved the useful life of some servers back from 6 years to 5 years; around the same time, Meta extended it further. Companies are giving different answers for the same kind of asset.

Second is where the debt sits. Oracle, Meta, xAI, and CoreWeave are estimated to have moved about $120 billion (1,200 × $100 million) of AI infrastructure debt off the balance sheet through SPVs⁴. A representative case is Meta’s Hyperion data center. Meta created an SPV owned 20% by Meta and 80% by Blue Owl Capital, and PIMCO, BlackRock, Apollo, and others lent it about $27 billion (270 × $100 million). That debt does not appear on Meta’s consolidated financial statements.

Third is circular financing⁵. Nvidia invests in OpenAI; OpenAI commits to cloud spending with Oracle; Oracle buys Nvidia GPUs to fulfill that commitment. Analyses in 2026 put the total scale of this kind of circular financing at more than $800 billion (8,000 × $100 million). Nvidia-OpenAI commitments of $100 billion (1,000 × $100 million), AMD commitments of $200 billion (2,000 × $100 million), and Oracle commitments of $300 billion (3,000 × $100 million) are all feeding into one another.

Read Smith’s sentence again: “Directors in 1 company bought, guaranteed, and leased their own assets in another company.”

🧩 The 2 eras are using the same sentences

Reader, by this point, 3 layers of symmetry should be visible.

First, the location of wear-and-tear costs. In 1848, renewal costs were hidden in the capital account to manufacture profit. Today, companies extend useful lives to push depreciation into the future. The account names have changed, but the work being done is the same. The gap between an asset’s real economic life and its accounting life becomes reported profit.

Where Smith wrote that “no company has closed its capital account,” we now have the GPU refresh cycle that returns every 2-3 years. A data center is not an asset you build once and forget. It is an asset you must keep buying again.

Second, self-generated demand. Railway directors bought lines controlled by themselves at 8 times their value and thereby created a share price for those companies. Today, a chip company invests in a model company, and that money flows back into chip purchases. In both cases, what looks like external demand is in fact an internal circulation of funds. The problem is not that this is illegal; it is that in such a structure, no one can answer the question, “How much real end demand is there?”

Third, and this is the real signal: the fight was not about technology but about audit.

In 1848, Lord Monteagle introduced a bill for government audits of railway accounts. Railway directors fought it with everything they had and killed the bill. Glyn, the chairman of LNWR, went so far as to say he would resign if the bill passed. Smith called this “certainly most suspicious.”

Think about it. If the accounts are clean, a government audit should help the share price. Smith made that point too. Yet the directors fought as if their lives depended on it.

The issue was never “are railways real?” It was who had the authority to verify the books.

The same is true now. Useful-life assumptions are management’s discretion. They are disclosed, but the mechanisms for forcing an outside test of whether those assumptions match economic reality are weak. For reference, FASB issued ASU 2024-03⁶ in 2024, a standard requiring expenses to be broken out by line item, including depreciation, and it applies to fiscal years beginning after December 15, 2026. We are standing just before these books open.

⚡ But the differences matter more

If we stop the analogy here, this becomes a banal “history repeats itself” essay. I think another point is much more practically important.

Railway capital was raised through partly paid shares. When you bought a share, you paid only part of its par value; the company could demand the rest whenever it needed money through a call. The remark Smith quotes from one shareholder is brutal:

“For every 1 pound I received in dividends, I paid 5 pounds in calls.”

This structure was designed to bankrupt individual investors systematically. That is why, when more than 200 million pounds vanished from British railway shares within 2 years after the 1845 peak, the shock flowed straight into household bankruptcies and shop failures.

A large share of today’s AI capex is different. It comes from hyperscalers’ operating cash flow. This is not the dot-com era, when data centers were built with debt or new equity issuance. This really is a different balance sheet. That objection deserves to be taken seriously.

But if we stop there, we have seen only half the picture. Leverage has not disappeared; it has migrated to the periphery. The $120 billion (1,200 × $100 million) of SPV debt above, private credit, neocloud vendor financing: the transmission channels for losses have changed, not vanished. And in 2026, reports have begun to say that Big Tech’s free cash flow is converging toward effectively zero. The defense that “operating cash flow can cover it” is itself now being tested.

Oswald’s Lens

When I build GTM strategy, I habitually ask: “Where did this number come from?” I have seen more than once a pipeline that looked good but was really revenue created because a partner bought the product. On paper, the contracts were all normal. Nobody lied. But if you build next quarter’s hiring plan on that revenue, something will inevitably break.

That is exactly what Smith did in 1848. He did not say “railways are a fraud.” He did not say the trains were not running. He followed the accounts line by line and asked where the money came from. And he found that the answer was: “from new shareholders.”

I am tired of the debate over AI that asks whether it is a bubble or not. I think the question itself is wrong, because the answer is already available. Britain really did get a railway network. And shareholders were ground down. Both sentences are true at the same time.

Whether the technology is real and whether the equity that financed it survives are entirely different questions. The lesson of 1848 is not “railways were fake.” It is: “even if the technology is real, shareholders die when the accounting is false.”

So the question to ask now is this:

Who ultimately receives the bill for these $433.9 billion (4,339 × $100 million)?

Closing

Let me reduce it to 3 lines.

Arthur Smith’s 1848 book was not a bubble warning but an accounting indictment. Account by account, he proved that dividends were coming not from operating profit but from new share payments and debt.

The core of today’s AI capex debate is also not technology but the books. Useful-life assumptions, SPV debt, circular financing: all 3 are questions of “where do we record the cost?”

In 1848, too, the real fight was over audit rights. And in December 2026, mandatory detailed disclosure of depreciation expense begins. It will be worth watching these books open.

If you are budgeting for AI adoption inside your organization right now, I would urge you to check just 1 thing: “How many years are we assuming for the renewal cycle of this investment?” 3 years and 6 years are completely different business plans.

And let me ask 1 question. Have you ever seen a situation at work where the question “Where did this number come from?” changed the whole game? If you can share which account it was and how you spotted it in the comments, I may build a future issue from it.

💬 Share your experience with the question above in the comments · 📨 If you have a colleague in accounting or finance, please send them this piece

📬 Every week, I analyze technology, economics, and the humanities in combination. Subscribe

⚠️ This article is not investment advice regarding any particular security or asset. The depreciation estimates cited here are scenario analyses calculated by changing assumptions, not settled accounting figures.

Arthur Smith’s 1848 book was not a bubble warning but an accounting indictment. Account by account, he proved that dividends were coming not from operating profit but from new share payments and debt.
The core of today’s AI capex debate is also not technology but the books. Useful-life assumptions, SPV debt, circular financing: all 3 are questions of “where do we record the cost?”
In 1848, too, the real fight was over audit rights. And in December 2026, mandatory detailed disclosure of depreciation expense begins. It will be worth watching these books open.

📎 References & Further Reading

Primary sources

Arthur Smith, “The Bubble of the Age; or, The Fallacies of Railway Investment, Railway Accounts, and Railway Dividends”, 2nd ed., London: Sherwood, Gilbert, and Piper, 1848. : The whole book is only 72 pages, so it is not a heavy lift. Read the LNWR capital-account analysis on pages 16-18 and the conclusion on pages 59-63, and you can verify the argument of this piece.
Silicon Analysts, “Hyperscaler AI Capex 2026: $434B Trailing Four Quarters, D&A Lag, Debt Wave”. : This tracks the gap between capex and depreciation quarter by quarter. The key figures in this piece come from here.
Deep Quarry, “Depreciation of GPUs: between useful lives and useful myths”. : This lays out both sides of the GPU useful-life debate. It also points out the weaknesses in the claim that profits are “obviously overstated.”

Background

Quinn Emanuel, “Emerging Litigation Risks in Financing AI Data Centers Boom”. : A legal practitioner’s breakdown of SPV structures. The Meta-Blue Owl Hyperion deal structure is explained clearly.
Noah Smith, “Should we worry about AI’s circular deals?”. : This is the opposing view, arguing that we need not worry too much about circular deals. Since it reaches the opposite conclusion from mine, it is worth reading alongside this piece.

📝 Glossary

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis, including business data management and business analytics, while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master's from Korea University's Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

Capital Account: An account that records money spent to create new assets. Money entered here is not treated as an expense for that year, so it does not reduce profit. By contrast, if the same spending is entered in the Revenue Account, that year’s profit falls by that amount. The same outlay changes profit depending on where it is recorded. ↩
Call: In 19th-century Britain, shares could be bought by paying only part of their par value. When a company needed cash, it could demand that shareholders pay the remainder; this demand was called a call. If shareholders did not pay, their shares could be forfeited. ↩
Useful Life: The period over which an asset is expensed. If the useful life is 6 years, you expense 1/6 each year; if it is 3 years, you expense 1/3 each year. Even for the same equipment, extending useful life makes that year’s profit look larger. ↩
SPV, or Special Purpose Vehicle: A separate company created for a specific project. If the parent company owns only a small stake, the SPV may not be included in consolidated financial statements, so the SPV’s debt does not appear on the parent’s books. ↩
Circular Financing: A structure in which a supplier invests in a customer, and the customer uses that money to buy the supplier’s products. Revenue is recorded, but the money is closer to the supplier’s own capital coming back than to cash from an end customer. ↩
FASB ASU 2024-03: A standard issued by the U.S. Financial Accounting Standards Board in 2024 requiring income-statement expenses to be broken out by line item, including depreciation. It applies to fiscal years beginning after December 15, 2026. ↩