October 2026

The bubble bursts, the rails remain

In 1847, Britain's railway shares crashed and thousands of investors were ruined. Yet the lines built during that "overbuilding" went on to carry the British economy for a century.

Nearly two hundred years later, the world's largest tech companies are spending close to $700 billion a year on AI infrastructure, while most businesses have yet to turn AI into profit. "Is this another bubble?" has become the market's defining question.

This piece answers it from two angles: a historian's view of how the steam engine grew from a mine pump into a world-changing force, and an economist's view of AI's current spending, returns and risks. The answer is probably not yes or no, but both.

The historian's view: steam took a century to change the economy

It took roughly 150 years for the steam engine to lift British productivity in a meaningful way. Its path ran through five stages:

1. Origins (1712–1769): a machine that only pumped water. Newcomen's engine converted less than 1% of its fuel into work and paid off only at coal mines, where fuel was nearly free. It was a mining tool, not a general-purpose technology.
2. Breakthrough (1769–1800): better engineering plus a new business model. Watt's separate condenser cut coal use sharply, and rotary motion in the 1780s let engines drive textile machinery. Boulton & Watt charged customers a share of the coal they saved — pay-for-results, two centuries early.
3. Diffusion (1800–1830): strikingly slow. Until around 1830, water wheels still powered most British factories. Coal prices, transport, factory redesign and a shortage of skilled mechanics all held adoption back.
4. Mania (1830–1847): railway fever and crash. Once steam was put on wheels, the 1840s Railway Mania drew in leveraged small investors and collapsed around 1847. Shareholders lost; the track stayed.
5. Maturity (1850–1870): the productivity payoff. Economic historians estimate steam's contribution to British labour productivity peaked only now. The factory system, urbanisation, a national market and modern finance took shape.

There was a cost too. Economic historian Robert Allen calls it "Engels' pause": between about 1790 and 1840, output soared while workers' real wages stagnated, and the gains flowed mainly to capital. Groups such as handloom weavers disappeared almost entirely.

The economist's view: AI is in its "railway mania" phase

Today's AI combines unprecedented investment with profits that have not yet caught up — exactly where steam stood between 1800 and 1830.

Capital spending: unmatched in scale and speed

Microsoft, Amazon, Alphabet, Meta and Oracle have guided to $635–690 billion of combined capital spending in 2026, up roughly 70% on 2025, with spending growing faster than cloud revenue (LongYield). One analysis estimates the cloud era cost about $1–1.5 trillion over 10–12 years, while the AI build-out has already reached about $0.8 trillion in around three years (Future Edge).

Financing: the signal to watch

The giants are shifting from funding with their own cash to funding with debt. Big tech has issued about $100 billion of bonds for AI in 2026, and demand for credit default swap protection has hit records (IEEE ComSoc). Oracle's capital spending rose from 37% of revenue in fiscal 2025 to 66% in the first half of fiscal 2026 (Futuriom). Investor Michael Burry disclosed a short position on a semiconductor ETF this year (Gotrade).

Business returns: people are faster, companies aren't richer yet

• McKinsey's 2026 survey: 80% of respondents say AI improved their own productivity, but only 37% report it contributing to company EBIT, roughly flat on the year before (MarketScale).
• A March 2026 survey of 650 enterprise technology leaders: 78% have AI pilots, but only 14% have reached production scale (ai2.work).
• The yardstick is getting harder: direct financial impact as the top AI ROI metric nearly doubled to 21.7%, while "productivity gains" fell from 23.8% to 18.0% (Futurum).

The parallel is striking. A steam engine bolted into a factory designed for a water wheel delivered little; AI dropped into processes designed for people delivers little too. The real returns come from redesigning the work itself.

Similarities, differences and the cycle

AI is not imitating the Industrial Revolution. It is repeating the pattern every general-purpose technology follows: breakthrough, investment frenzy, infrastructure overbuild, shakeout, organisational rebuild. Electricity and the internet took the same road.

Five similarities

Dimension	Steam era	AI era
From niche to general	From mine pumps to every industry	From chat and coding to every industry
Productivity lag	Nearly a century after Watt before productivity surged	Individual gains real, company profits limited
Bottleneck is organisation	Factories rebuilt, mechanics trained	Processes, data and roles must be rebuilt
Infrastructure-led frenzy	Railway Mania, leveraged retail investors	Hundreds of billions in data centres, rising debt
Uneven gains	Output up, wages flat	Gains concentrated in a few platforms and chipmakers

Five differences

1. Speed. Steam took over a century to spread; generative AI became standard in businesses worldwide in under four years. Bubbles form and clear faster, and society has less time to adapt.
2. What gets replaced. Steam replaced muscle and water power and created many skilled jobs; AI targets cognitive work, hitting white-collar roles first.
3. Asset life. Rails last decades; AI chips are obsolete in a few years. When the railway bubble burst the track remained; if an AI bubble bursts, outdated GPUs will be worth far less.
4. Who invests. Railway Mania was driven by small investors; today's spenders are cash-rich tech giants with far greater resilience.
5. Geopolitics. Steam technology eventually spread across Europe and America; AI has been shaped from the start by chip export controls and great-power rivalry.

The cycle side by side

Stage	Steam	AI
Origins	1712 Newcomen engine	2012–2017 deep learning, Transformer
Breakthrough	1769–1781 Watt's improvements	2020–2023 large models, ChatGPT
Mania and build-out	1830–1847 Railway Mania	2023–now capex race (we are here)
Shakeout	1847 crash	Not yet
Maturity	1850–1870 productivity surge	Depends on when companies redesign their work

Conclusion: a bubble and a revolution

AI is most likely a partial bubble sitting on top of a real technological revolution, much like the railways of the 1840s.

Signs of a bubble: capital spending growing far faster than revenue; falling free cash flow and rising debt; circular flows where one AI company's spending becomes another's revenue; a stalled share of companies seeing profit impact.

Signs it is not just a bubble: demand is real and cloud AI revenue is growing fast; the main investors are financially strong; individual productivity gains are widely confirmed.

History's key lesson is that "is it a bubble?" and "is the technology real?" are separate questions. Most Railway Mania shareholders lost money, yet the railways transformed Britain. The question worth asking is not whether prices will fall, but which investments will last like rails and which will depreciate like old GPUs.

For businesses, history has an answer as well: the winners were not the first factories to buy a steam engine, but the first to redesign the factory around it.

This article is analysis and opinion, not investment advice.