From Railway Mania to Artificial Intelligence Frenzy: How Technological Revolutions Navigate Through Capital Bubbles

Text by Tan Yinliang, Professor of Decision Sciences and Management Information Systems, China Europe International Business School

In 2026, the core debate in the artificial intelligence market has shifted from "whether it is useful" to "whether the massive capital invested can be recouped". Goldman Sachs estimates that global AI-related investment will exceed 1 trillion US dollars in 2026, of which about 581 billion US dollars will be generated in the United States; analysts at BofA Securities predict that global capital expenditure on cloud computing and AI infrastructure may rise to 1.4 to 1.5 trillion US dollars in 2027. The latter figure is equivalent to nearly 5% of the United States' annual GDP. Although these investments are led by American technology companies, they are distributed across the globe, and cannot be directly interpreted as the proportion of domestic investment in the United States reaching 5%. According to Goldman Sachs' stricter geographic caliber, the proportion of US AI investment in GDP is expected to rise from 1.8% in 2026 to 2.5% in 2027.

Why compare the current situation with the railway era? The reason is that in terms of capital expenditure intensity and construction mechanism, the Internet is not the closest historical parallel. Around 2000, the United States' broad investment in information and communication technology once reached about 4.5% of GDP, but that included purchases of computers, software and communication equipment by tens of thousands of enterprises, representing the decentralized diffusion of technology across the whole society. If we only look at the underlying telecommunication network of the Internet, the capital expenditure of listed American telecom operators in 2000 was about 121 billion US dollars, only equivalent to 1.2% of that year's GDP.

The situation today is different. Investments are highly concentrated in a small number of large technology enterprises, and are pouring into chips, servers, data centers, power supply, optical fibers, cooling facilities and land at the same time. If we limit the conditions to private capital-led investment, annual input reaching several percentage points of the total economic volume, long construction cycles, non-withdrawable capital commitments, and the necessity of building networks first before demand takes shape, almost only railways in modern commercial history can be used for analogy. UK railway investment reached 5.7% and 6.7% of GDP in 1846 and 1847 respectively. Therefore, the railway is not just a rhetorical reference, but a mirror reflecting the balance sheet situation.

UK: Stock prices peak first, railway investment peaks later

In the 1840s, UK railway stocks rose by an average of about 106% between 1843 and 1845. Then the market reversed, and the representative railway stock index fell by about 64% from its peak in the summer of 1845 to the end of 1849. However, the peaking of stock prices did not immediately halt construction. The total length of operational railways in the UK increased from 2240 miles in 1844 to 6621 miles in 1850, and the annual new mileage did not reach its peak until 1848; the peak of railway investment as a share of the total economy also appeared after the stock price peaked.

The reason is not complicated. Stock prices are votes on the future, while capital expenditure is the settlement of past commitments. Land has been purchased, contracts have been signed, and projects have started, so even if the market cools down, spending will continue. What is more dangerous is that at that time, investors only needed to pay a small amount of subscription money initially, and were then repeatedly asked to pay in more capital as the projects advanced. With costs exceeding expectations, revenues falling below forecasts, parallel lines being repeatedly constructed, and credit tightening, the technology boom quickly turned into financing pressure.

The most direct reminder for today is that the continuous acceleration of capital expenditure does not mean that investment returns are still improving. The construction peak is often just the most concentrated fulfillment of the previous round of optimistic expectations.

US: After the bubble burst, railways continued to expand

The US railway cycle was longer, and more similar to the possible AI cycle in the future. The annual new railway mileage in the US rose from 1404 miles in 1866 to 7439 miles in 1872. A large number of lines were built in the western regions where demand was not yet mature, and revenues took many years to materialize, while interest payments had to be made immediately. In 1873, the collapse of Jay Cooke & Company, which financed the Northern Pacific Railway, triggered a financial panic; 89 of the 364 railway companies went bankrupt, and the annual new mileage dropped to 1606 miles by 1875.

But the railway era did not end. The new mileage recovered to 5006 miles in 1879, reached 11599 miles in 1882, and rose to 13081 miles in 1887. Overconstruction and high debt then triggered the 1893 crisis again, and more than 125 railways entered bankruptcy receivership in the year ending June 1894.

US railways did not experience a simple "boom-bust" cycle, but repeatedly went through construction, crisis, refinancing, merger, standardization and re-construction. The burst of the bubble did not negate railways, but changed their owners, capital structure and industry order. The first boom answered the question of "whether railways exist", and the post-crisis integration answered the questions of "who will operate them, how to interconnect them, and how to improve their utilization rate". Artificial intelligence will most likely go through the same process.

The biggest mismatch: social value does not equal shareholder return

The value created by railways far exceeds the profits of railway companies. A study published in the Journal of Political Economy estimates that without railways, the overall productivity of the United States in 1890 would have been about 27% lower; the annual social return on railway capital was about 48%, but railway enterprises only obtained about 7% of it.

The same is likely to be true for artificial intelligence. It reduces the cost of information processing, knowledge production, organizational coordination and decision-making. The biggest beneficiaries may be enterprises that use AI to restructure R&D, marketing, supply chains and customer services, as well as consumers and new entrepreneurs, not necessarily the companies that bear all the infrastructure investment.

Therefore, the conclusion that "AI will improve productivity" cannot automatically lead to "all AI infrastructure investments will generate high returns". A technology can achieve great social success while resulting in financial failure for some shareholders.

The real turning point: shifting from free cash flow investment to debt financing

What is most worth observing today is not just the scale of capital expenditure, but where the funds come from. In the past, large technology companies mainly used their strong free cash flow to build data centers; now, capital expenditure is approaching operating cash flow, and bonds, long-term leases, project financing and private credit are starting to play a bigger role. Goldman Sachs estimates that about one-third of the capital expenditure of hyperscale technology enterprises in 2026 may be financed by debt, and the proportion may rise to 35% in 2027; the time lag from capital input to revenue generation may be extended from three months to two years.

The investment logic has thus undergone fundamental changes: in the past, companies used today's cash to buy future growth, but now it is increasingly using today's financing cost to buy uncertain future cash flow. The perspective for observing enterprises must also shift from the income statement to the balance sheet, from revenue growth rate to the comparison between debt cost and return on investment, and from annual performance to refinancing window, contract term, residual value of assets and computing power utilization rate. This is a meso-level change in the industry.

At the micro level, the US capital market is partially shifting from an investment market that "prices" for growth to a financing market that "supplies blood" for construction. Investors should not only ask how much the company plans to spend, but also who provides the funds, at which layer the risk remains, and once revenue falls short of expectations, whether the loss will be borne by shareholders, creditors, data center owners or private credit funds.

At the macro level, liquidity will not always be abundant. The divergence in US fiscal governance makes it difficult to rapidly reduce the deficit. The Congressional Budget Office estimates that the federal deficit in fiscal year 2026 will account for about 5.8% of GDP, and rising interest payments make the debt issue more persistent; the normalization of Japan's monetary policy may lead to the gradual exit of hidden leverage and yen financing transactions accumulated in the long-term low interest rate environment. Both factors may raise the global cost of capital, and reduce the tolerance space for high-valuation assets that rely on cash flow from the distant future.

These changes do not mean that problems will occur immediately, nor can we conclude that a crisis is inevitable. Technology giants still have strong balance sheets, and real demand is still growing. But the observation window for risks has changed: a one-dimensional stall may only be an industry revaluation; if the three dimensions of declining return on investment, rising refinancing cost and tightening global liquidity resonate at the same time, risks may spread beyond the AI industry and evolve into a broader credit event, or even a financial crisis in extreme cases.

From construction race to utilization rate race

In the next stage, the market will shift from "who builds more" to "who can fill the assets", from comparing capital expenditure to comparing return on capital, and from rewarding companies that own the most chips to rewarding enterprises that can best embed AI into customer processes and generate continuous cash flow.

Railways did not fail. What failed was the capital that overestimated demand, underestimated costs, relied on fragile financing and repeatedly constructed lines. Artificial intelligence will not fail just because of falling stock prices, suspended projects or corporate bankruptcies. What really needs to be guarded against is misinterpreting "AI is bound to change the world" as "every AI investment is bound to make money". History will not excuse capital allocation errors just because the technology direction is correct.