AI could transform the economy and still destroy vast amounts of capital.
Britain’s Railway Mania shows how a sound technological forecast can become a poor investment, especially when no competitor can afford to stand still.
On 28 July, Fitch Ratings described an AI market correction as an emerging global credit risk. Its concern was not that artificial intelligence had proved useless. It was that the cost of pursuing it had become large enough to matter well beyond the technology industry.
Alphabet, Amazon, Meta and Microsoft are projected to spend about $700 billion between them in 2026, an increase of more than 75 per cent in a year. Not all of that expenditure is exclusively for AI, but much of the increase is being directed towards the data centres, chips and power infrastructure needed to build and run it. Fitch estimates that booming IT investment directly added 1.4 percentage points to first-quarter US economic growth. If the expected returns fail to arrive, the effects will not stay neatly contained inside Silicon Valley.
The obvious debate asks whether AI is a revolution or a bubble. It is probably the wrong question.
A technology can be genuinely transformative while the businesses financing its expansion earn disappointing returns. It can change daily life, raise productivity and leave behind useful infrastructure, even as share prices fall and investors lose money.
Britain discovered this in the 1840s. The railway was real. The transformation was real. So were the losses.
A boom built on evidence
The Railway Mania is easy to remember as an episode of Victorian credulity: promoters sketching improbable lines across maps, newspapers feeding public excitement and inexperienced investors chasing easy riches. All of that happened. It is also an incomplete account.
By the early 1840s, railways had already demonstrated their value. The Liverpool and Manchester Railway had been operating since 1830. By 1842, 1,951 miles of track were open across Great Britain. Trains were moving people and goods faster than the roads and canals they replaced, while the leading companies were becoming more profitable.
Across ten of the largest railways, passenger and freight receipts rose by 42 per cent between 1842 and 1846, even though their open mileage increased by only 25 per cent. Dividends at established railway companies climbed too. There was speculation in the market, but there was also visible demand, improving productivity and real cash flowing through the system.
That distinction matters. Investors were not being asked to believe in a machine that existed only in a prospectus. They could see locomotives arriving, journey times shrinking and successful lines paying their owners. The mistake was subtler: evidence that some railways were valuable became confidence that almost any additional railway would be equally valuable.
The mechanism for turning that confidence into a frenzy was unusually effective. A railway needed Parliament’s approval, largely because construction required landowners to sell property along the proposed route. Hundreds of schemes therefore arrived at Westminster seeking authorisation. By the autumn of 1845, 562 new proposals had been submitted for the next parliamentary session. The Times estimated that 1,263 railway projects had been promoted during that year alone.
Investors did not usually pay the full value of a share at the start. They put down a relatively small deposit and promised to provide the rest when the company issued a call for capital. This made railway shares feel affordable and magnified the early gains. It also left shareholders liable for much larger payments later.
More than 33,000 named individuals subscribed to new schemes considered by Parliament in 1845 and 1846, and the surviving records understate the true number. Many were members of a growing middle class entering the equity market for the first time. Railway shares more than doubled between 1843 and the autumn of 1845.
Britain was not merely imagining a national railway system. It was trying to finance one all at once.
When the promise became a payment
The mood changed before the technology did.
A poor harvest and the failure of the potato crop in Ireland placed pressure on food supplies and the financial system. Gold reserves fell as Britain paid for grain imports. The Bank of England raised its rate twice in the autumn of 1845. At the same time, newly authorised railway companies moved from selling a vision to building lines, which meant calling on shareholders for the money they had promised.
From January 1843 to September 1845, railway companies called up an average of about £300,000 a month. In the final three months of 1845, they demanded £4.4 million. Investors who could not meet a call sold the affected share, or sold other investments to find the cash. Those sales pushed prices down, which made confidence weaker and further payments harder to meet. A structure that had amplified the boom now amplified the decline.
One reconstruction of the market shows railway shares falling by roughly 64 per cent from their peak in October 1845 to their trough in April 1850. Many planned lines were abandoned or suspended. Companies failed, dividends were cut and households suffered losses.
Yet the building did not simply stop. Capital committed during the enthusiasm of 1845 continued to pass into iron rails, bridges, cuttings and stations long after railway shares had begun to fall. Britain’s open network tripled from its 1842 level to 6,123 miles by 1850.
The result was a strange separation between public usefulness and private reward. A denser network connected more places, carried more traffic and intensified competition. It also pushed the industry’s average dividend as a share of paid-in capital from a peak of 7 per cent in 1847 to 2.4 per cent by 1852. Expansion made railways more useful as a system and less lucrative for many of their owners.
The longer-term social gain was substantial. Economic historian Timothy Leunig estimates that, by 1912, the time and fare savings created by passenger rail were equal to 14 per cent of national income. Across the period he studied, railways accounted for around a sixth of economy-wide productivity growth. None of this restored the money lost by a shareholder who had bought the wrong line at the wrong price.
A technology can be transformative in public and ruinous in private.
The trap of standing still
Railway Mania was not driven only by naïve investors or dishonest promoters. Established railway managers faced a genuine strategic problem.
An existing company might have preferred to operate its profitable route without funding an expensive expansion. But if a rival built a competing line, it could take passengers, freight and control of an important junction. Once competitors began proposing new routes, standing still became dangerous. Companies promoted defensive lines, entered less profitable areas and absorbed other railways to protect their territories.
Research by Gareth Campbell and John Turner describes this as a prisoner’s dilemma. The industry as a whole might have earned better returns with less construction, but each individual company had reason to expand because others were expanding. What was rational for one firm helped produce a worse outcome for the group.
That is the most revealing connection to the AI boom.
In 2024, Alphabet chief executive Sundar Pichai told investors that the company regarded the risk of underinvesting in AI as dramatically greater than the risk of overinvesting. Surplus infrastructure could be used across Google’s businesses, he argued, while failing to remain at the frontier carried a far more serious competitive cost.
Two years later, that logic is visible in the numbers. Alphabet has raised its 2026 capital expenditure guidance to between $195 billion and $205 billion. It says demand continues to exceed available capacity. Its cloud revenue grew by 82 per cent in the second quarter, while its cloud backlog reached $514 billion. These are not the figures of a market with no customers.
They do not prove that every investment will earn an attractive return either.
Amazon, Meta and Microsoft face versions of the same decision. If a company believes AI could reshape search, advertising, software, retail or cloud computing, holding back may place its existing business at risk. Even a chief executive who suspects the industry is building too much capacity cannot know which rival will discover the next improvement or secure the most valuable customers.
Competition can make overinvestment rational.
If every major company reaches the same conclusion, the result may be far more computing capacity than any of them would have chosen in a coordinated market. That abundance can lower the price of AI, improve the products built on top of it and accelerate adoption. It can also reduce the return earned on each new data centre.
This is where public excitement often merges two separate forecasts. The first is that a technology will become important. The second is that today’s providers, assets and prices will capture the value it creates. The first can be correct while the second is badly wrong.
Rail passengers gained from a larger network and more competitive routes. Railway shareholders discovered that carrying more passengers did not guarantee generous dividends. With AI, the largest gains could similarly flow towards the businesses that use cheap intelligence, or towards consumers who receive better services, rather than remaining with every company that paid to create the underlying capacity.
The case against calling it a bubble
There is a serious objection to this argument: the spending may be broadly justified.
AI adoption is no longer confined to demonstrations and speculative start-ups. Stanford’s 2026 AI Index estimates that generative AI reached 53 per cent population-level adoption within three years. Reported organisational AI adoption reached 88 per cent in 2025, while estimated annual consumer surplus in the United States reached $172 billion by early 2026. Revenue at leading AI companies is rising quickly, and major cloud providers still report capacity constraints.
The firms doing most of the building are also not lightly financed Victorian promotions. They are among the world’s largest and most profitable companies, with established customers, valuable distribution networks and several ways to use the infrastructure they create. Alphabet can deploy the same capacity across cloud services, search, advertising and internal operations. Even if one application disappoints, the asset may support another.
There is also a genuine historical dispute about how irrational Railway Mania really was. Campbell’s analysis suggests railway shares were not obviously mispriced at their peak when judged against the dividends investors could observe at the time. The subsequent fall reflected changing conditions, heavier competition and lower future earnings, developments that were much clearer afterwards than beforehand.
This does not make the railway losses imaginary. It makes the lesson more uncomfortable. Wasteful outcomes do not require everyone involved to be foolish. Sensible decisions, made with incomplete information and under competitive pressure, can accumulate into a poor collective result.
Calling AI a bubble can therefore obscure as much as it reveals. The useful question is not whether excitement exists. It is whether the eventual cash generated by the infrastructure will justify its cost, after competition has lowered prices and technology has moved on.
Where the rails end
A railway line is a long-lived physical asset. Embankments, tunnels and routes can remain useful for generations. AI accelerators can lose economic value much faster as newer chips and more efficient models arrive. A data-centre building and its power connection may endure, but the expensive equipment inside it is not equivalent to a Victorian viaduct.
The financing is different too. Railway Mania drew thousands of individuals into part-paid shares that could expose them to ruinous capital calls. Today’s build-out is led by profitable incumbents using operating cash flow, conventional debt and partnerships. Losses could still spread through shareholders, lenders, suppliers and the wider economy, but the route would be different.
Most importantly, the AI builders may be better placed to capture downstream value. Nineteenth-century railway companies transported the passenger but did not own the factory made more productive by the journey. The largest technology groups own infrastructure, models, software, advertising systems and customer relationships. Their control of several layers may allow them to retain returns that railway owners lost to competition and their users.
These differences are large enough to rule out any claim that AI must follow the railway market’s path. The analogy identifies a pressure, not a timetable.
What to watch next
How the expansion is financed. Capital expenditure funded comfortably from operating cash flow creates a different risk from spending increasingly supported by borrowing, private credit or guarantees. Rising debt does not prove a bubble, but it widens the number of people exposed if returns disappoint.
Price and utilisation of computing capacity. Persistent shortages and stable pricing would support the case that construction is meeting real demand. Falling rental prices, idle capacity or cancelled data-centre projects would suggest that supply is arriving faster than profitable uses.
Quality of revenue disclosure. Broad claims about AI-assisted growth are less informative than recurring revenue that can be tied to AI products, set against the depreciation, energy and operating costs required to provide them.
Where the economic benefit appears. If companies using AI achieve lasting productivity gains while model and infrastructure providers struggle to protect margins, the technology may be succeeding while value moves away from its builders.
What companies do with yesterday’s hardware. Shorter asset lives, large write-downs or repeated upgrades before equipment has paid for itself would weaken the railway comparison’s most optimistic element: the idea that overbuilding at least leaves durable infrastructure behind.
The lesson of Railway Mania is not that AI is destined to crash. It is that technological importance and investment return answer different questions. The rails survived. Many of the expected returns did not. If a similar pattern is forming around AI, the revealing moment will not necessarily be when the technology stops working. It may be when it works, spreads and becomes cheaper, while leaving less profit where the money was first poured in.
