Railway Dividends Funded by New Share Capital, Not Profit

1848: Arthur Smith Digs Into Railway Accounting

In London in 1848, a man named Arthur Smith published a slim book. Its title was The Bubble of the Age, and its subtitle: “Railway Investment, Railway Accounts, and the Fiction of Railway Dividends.” He wasn’t a famous economist or a politician. He was just someone who had gone all the way to the bottom of other companies’ ledgers.

Most people who bring up this book today call it “an 1848 warning about the railway bubble.” I think that’s the biggest misreading of it you can make.

Smith never once disputes railway technology itself. The trains really were running, and the fare revenue was real. The one thing he wouldn’t let go of was this: which account a given cost had been booked to.

This book isn’t a bubble warning—it’s an accounting indictment. And the debate now swirling around AI infrastructure investment turns on the exact same question it did 178 years ago: which account do you book the cost to.

🔍 1848: An Accountant’s Indictment

Smith tore into the books of the London and North Western Railway (LNWR). The company’s main lines had already been open for years — Liverpool to Manchester since 1830, Manchester to Birmingham since 1842.

But something didn’t add up. On lines that were already fully built, close to £1,000,000 a year kept flowing into the capital account¹. In just the last two years alone: £1,983,472. And £489,589 of that was spent in the very half-year the company itself described in its own report as “a period of severe depression and unprecedented crisis.”

What did this mean? Smith put it this way:

“Railways can only be worked by a constant addition to the capital account, greater than the dividends declared.” (Railways can only be kept running by continually adding to the capital account more than the dividends declared.)

In other words, they were pushing the costs of wear and replacement into the capital account rather than the profit-and-loss account. That made the income statement look profitable. Profits let them declare dividends; dividends pushed up the share price; a higher share price let them issue new shares; and the money raised from those new shares funded the next round of dividends.

Smith summed up the whole structure in a single line.

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

And here’s the sentence from this book that I sat with the longest.

“What information, the single fact, that no Company has closed its capital account, should convey to the cautious.” (How much a single fact should tell the cautious observer — that no company has ever closed its capital account.)

This wasn’t an asset you build once and be done with — it was an asset that had to be kept alive by continually injecting new capital. Yet these companies were booking it as if it were a one-time construction cost.

Smith didn’t stop there. He also dug into the deals directors were making with each other. The Hull & Selby Railway had a market capitalization of under £250,000, yet the North Midland board guaranteed to buy it for £2,000,000. The Great North of England committed to buying a company worth £438,900 on the market for £4,000,000. Smith’s phrase for it was exact:

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

There’s one name this book keeps coming back to: George Hudson, the man known at the time as the “Railway King.” Smith called him out by name in 1848, and the following year, in 1849, Hudson was exposed for paying dividends out of capital — and it destroyed him.

📊 2026: Hyperscaler Capex and Depreciation

Let’s come back to the present.

Start with scale. Between Microsoft, Alphabet, Amazon, and Meta, spending on property and equipment over the four quarters through Q1 2026 came to about $433.9 billion. Over the same period, these four companies recognized roughly $149.0 billion in depreciation. Combined capex for Q1 2026 alone was $129.8 billion, up 80% year over year.

This is exactly where Smith, if he were watching, would ask the obvious question: why does spending $433.9 billion only produce $149.0 billion in depreciation?

Part of the answer lies in useful life³. Between 2022 and 2023, Amazon, Alphabet, and Microsoft all extended the accounting useful life of their server and networking equipment from the previous 3–4 years to 6 years. Stretch out an asset’s assumed lifespan, and the annual depreciation charge shrinks — which inflates reported earnings by exactly that much.

How big a difference does this make? According to a sensitivity analysis by Goldman Sachs, if GPU useful life were shortened from 5 years to 3 years, cumulative depreciation from 2026 through 2031 would rise from roughly $3 trillion to $4 trillion. That’s a swing of $1 trillion. Michael Burry estimated the resulting earnings overstatement between 2026 and 2028 at more than $176 billion.

To be fair, though, there’s something worth flagging honestly here. These figures are scenarios built on altered assumptions, not settled facts. And whether GPUs actually last 3 years or 6 is genuinely contested within the industry. What’s interesting is that companies have started diverging on this. Starting in January 2025, Amazon walked back the useful life of some servers from 6 years to 5, while Meta, over the same period, extended its own further. Different companies are reaching different answers about the very same assets.

Second is the question of where the debt actually sits. AI infrastructure debt that Oracle, Meta, xAI, and CoreWeave have moved off their balance sheets through SPVs⁴ now totals roughly $120 billion. The signature example is Meta’s Hyperion data center. Meta created an SPV in which it owns 20% and Blue Owl Capital owns 80%, and into that vehicle PIMCO, BlackRock, Apollo, and others lent roughly $27 billion. None of this debt appears on Meta’s consolidated financial statements.

Third: circular transactions⁵. Nvidia invests in OpenAI. OpenAI commits to buying cloud capacity from Oracle. Oracle, in order to fulfill that commitment, buys Nvidia GPUs. Analyses in 2026 put the total size of this kind of circular financing at over $800 billion — the Nvidia–OpenAI commitment worth $100 billion, AMD’s at $200 billion, Oracle’s at $300 billion, all locked into each other.

Worth rereading Smith’s line here: “the directors of one company bought, guaranteed, and leased assets belonging to themselves in another.”

Looking at the draft, it accurately preserves the structure, numbers, and meaning. No Hangul remains, all numbers match, headings/footnotes align. Only minor polish needed.

🧩 Three Ways 1848 and 2026 Overlap

There are three points, dear reader, where these two eras converge.

First, which account absorbs wear-and-tear and renewal costs. In 1848, renewal expenses were hidden in the capital account to manufacture profit. Today, useful-life assumptions are stretched out to defer depreciation. The account names have changed, but the function is identical. The gap between an asset’s actual economic lifespan and its lifespan on the books becomes reported profit, directly and mechanically.

Where Smith once wrote that “no company has ever closed its capital account,” we now have the GPU refresh cycle, which comes around every 2-3 years. A data center isn’t an asset you build once and finish with. It’s an asset you have to keep buying, over and over.

Second, demand is manufactured with internal money. Railway directors bought lines from other companies they themselves controlled at 8 times their value, and in doing so, built up that company’s share price. Today, chip companies invest in model companies, and that money flows straight back into chip purchases. In both cases, what looks like external demand is actually money circulating within a closed loop. The issue isn’t that this is illegal—it’s that within this structure, nobody can answer the question of what real, final demand actually is.

Third, the 1848 dispute wasn’t about technology at all—it was about who held the authority to audit the books.

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

Think about it—it’s strange, isn’t it? If the accounts were clean, a government audit should have been good for the share price. Smith made this very point. And yet the directors fought it tooth and nail.

The question at stake was never “is the railway real.” It was who gets to hold the authority to verify the ledger.

The same is true now. Useful-life assumptions are management’s discretion. They’re disclosed, sure, but there’s no strong external mechanism forcing verification of whether those assumptions match economic reality. For reference, in 2024 the FASB issued a standard (ASU 2024-03)⁶ requiring depreciation to be disclosed broken down by category, and it takes effect for fiscal years beginning after December 15, 2026. So itemized depreciation disclosures will only become visible once this standard actually kicks in.

⚡ But the difference matters more than the parallel

If I stopped the analogy here, this would just be a trite “history repeats itself” piece. I think the point where the two diverge is far more practically important.

Railway capital was raised through a partly-paid share structure. When you bought a share, you paid only a fraction of its face value upfront, and the company issued “calls” demanding further installments whenever it needed cash. Smith quotes one shareholder’s line, and it’s brutal:

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

This structure was designed, structurally, to bankrupt individual investors. That’s why, when more than £200 million evaporated from British railway shares within two years of the 1845 peak, it immediately cascaded into household bankruptcies and shuttered shops.

Today’s AI capex is different in one big respect: a large chunk of it comes straight out of hyperscalers’ operating cash flow. They’re not building data centers with debt or share issuance the way dot-com companies did. The funding structure itself is different from back then. This counterargument deserves to be taken seriously.

But stopping there only gets you half the picture. Leverage hasn’t disappeared — it has migrated to the periphery. The $120 billion in SPVs, the private credit, the neocloud vendor financing we looked at earlier — the channel through which losses get transmitted has changed, but the losses themselves haven’t gone away. And heading into 2026, reports have started emerging that Big Tech’s free cash flow is effectively converging toward zero. The very defense line of “we can cover it with operating cash flow” is now being put to the test.

Oswarld’s Lens

When I’m building a GTM strategy, I’ve made a habit of asking, “Where exactly did this number come from?” I’ve seen plenty of cases where the pipeline looked great, but the revenue behind it was actually just a partner buying in to make the numbers work. On paper, everything’s clean. Nobody’s lying. But if you build next quarter’s hiring plan on that revenue, you’re heading for trouble.

That’s exactly what Smith did in 1848. He didn’t say “the railways are a fraud.” He didn’t say the trains weren’t running. He simply went down the accounts line by line and asked where the money had come from. And the answer he found was: from new shareholders.

I’m honestly a bit tired of the debate over whether AI is “a bubble or not.” I think it’s the wrong question. Britain genuinely got its railway network, and at the same time, railway shareholders took real losses. Both things can be true at once.

Whether a technology is real and whether the equity invested in it survives are two completely different questions. The lesson from 1848 isn’t “the railways were fake” — it’s that even when the technology is real, if the accounting doesn’t match reality, shareholders lose money.

So here’s the question worth asking right now. Of the $433.9 billion your company has spent over the last four quarters, the portion that hasn’t yet been written off as an expense — who ultimately ends up bearing that cost?

Looking at the fragment, I compared it against the Korean source and found no meaning distortions, omissions, number mismatches, or glossary violations. All numbers (1848, 3, 6, 2026) are preserved as digits, structure matches exactly, and there is zero Hangul remaining. The translation is faithful and natural.

Closing

Let me sum this up in three lines.

Arthur Smith’s 1848 book wasn’t a bubble warning—it was an accounting indictment. He proved, account by account, that dividends were coming not from operating profit but from new share subscriptions and borrowed money.

The core of today’s AI capex debate isn’t technology either—it’s the books. Useful-life assumptions, SPV debt, circular deals. All three boil down to one question: “where do you record the cost?”

Even in 1848, the real fight was over audit rights. And in December 2026, mandatory itemized depreciation disclosure begins. Once that disclosure starts, we’ll be able to check line-item depreciation figures directly.

If you’re currently building an AI-adoption budget in your organization, I’d suggest checking just one thing: “What renewal cycle did we assume for this investment?” A 3-year cycle and a 6-year cycle are completely different business plans.

And let me ask you one thing. Have you ever had a moment in your own work where a single question—“where does this number actually come from?”—completely changed the judgment call? Tell me in the comments which account it was and how you caught it; I’ll use it as material for the next issue.

---

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

---

📬 Every week, I analyze technology, economics, and the humanities at their intersections. Subscribe

⚠️ This piece is not investment advice regarding any specific stock or asset. The depreciation estimates cited are scenario analyses based on varied assumptions, not confirmed accounting figures—please keep that in mind.

1. Arthur Smith’s 1848 book wasn’t a bubble warning—it was an accounting indictment. He proved, account by account, that dividends were coming not from operating profit but from new share subscriptions and borrowed money.
2. The core of today’s AI capex debate isn’t technology either—it’s the books. Useful-life assumptions, SPV debt, circular deals. All three boil down to one question: “where do you record the cost?”
3. Even in 1848, the real fight was over audit rights. And in December 2026, mandatory itemized depreciation disclosure begins. Once that disclosure starts, we’ll be able to check line-item depreciation figures directly.

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. : It’s only 72 pages, so don’t let the length put you off. If you just read the LNWR capital account analysis on pages 16–18 and the conclusion on pages 59–63, you’ll have everything you need to check this piece’s argument.
• 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 came 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 pinpoints the weaknesses in the “it’s all overstated” argument.

Background

• Quinn Emanuel, “Emerging Litigation Risks in Financing AI Data Centers Boom”. : This dissects SPV structures from a legal practitioner’s perspective. It does a good job explaining the structure of the Meta–Blue Owl Hyperion deal.
• Noah Smith, “Should we worry about AI’s circular deals?”. : This is the counterargument — that we shouldn’t overworry about circular deals. It reaches the opposite conclusion from mine, so it’s worth reading alongside this piece.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Capital Account: An account that records money spent creating new assets. Money entered here isn’t booked as an expense for that year, so it doesn’t reduce profit. If the same spending were entered in the Revenue Account instead, that year’s profit would shrink by the same amount. Where you record the same expenditure changes how much profit you show. ↩

2. Call: In 19th-century Britain, shares could be bought by paying only part of their face value up front. When a company needed more money, it would demand shareholders pay the rest — this demand was called a “call.” Fail to pay, and your shares could be forfeited. ↩

3. Useful Life: The period over which an asset’s cost is spread as an expense. Set it at 6 years, and you write off one-sixth each year; set it at 3 years, and you write off one-third. Stretch out the useful life on the same equipment, and that year’s profit looks bigger. ↩

4. SPV (Special Purpose Vehicle): A company set up solely to run a specific business. If the parent company holds only a small stake, the SPV isn’t included in the consolidated financial statements — so any debt the SPV carries stays invisible on the parent’s books. ↩

5. Circular Financing: A structure where a supplier invests in its customer, and the customer uses that money to buy the supplier’s products again. Revenue does get recorded, but it’s less money coming from an end consumer than money the supplier itself put in, circling back around. ↩

6. FASB ASU 2024-03: A standard issued by the U.S. Financial Accounting Standards Board in 2024 requiring companies to disclose income statement expenses broken out by category, including depreciation. It applies to fiscal years beginning after December 15, 2026. ↩