From 19th-century railways to 21st-century artificial intelligence, every major technological innovation in history has sparked a surge in capital expenditure (Capex)—yet such frenzies often end in a burst bubble.
In a special report titled When Capex Booms Turn Into Busts: Lessons From History, released this November, BCA Research reviews four classic Capex booms. The report reveals the underlying dynamics of the transition from boom to bust and issues a warning regarding the current AI frenzy.
The report identifies five common patterns: investors overlooking the S-curve of technology adoption; revenue forecasts underestimating the magnitude of price declines; a heavy reliance on debt for financing; asset price peaks preceding the decline in investment; and a mutually reinforcing relationship between the Capex collapse and economic recession. Signs of these patterns are already emerging in the AI sector—stagnant technology adoption rates, token price crashes exceeding 99%, surging corporate debt, and falling GPU rental costs.
Based on this historical comparative analysis, BCA Research concludes that the AI boom is following the trajectory of historical bubbles and is likely to end within the next 6 to 12 months. The report advises investors to maintain a neutral stance on equities in the short term and a moderately underweight position in the medium term, while closely monitoring forward-looking indicators such as revisions to analyst expectations, GPU rental costs, and corporate free cash flow.
The report highlights additional concerns regarding the current economic environment, noting that US job openings have fallen to a five-year low. If the AI boom fades without a new bubble to offset the shock, a future recession could prove more severe than the one following the 2001 dot-com crash.
Lessons from History: The Collapse Trajectories of Four Capital Frenzies
BCA notes that Capex booms are fundamentally driven by collective investor optimism regarding the commercialization prospects of new technologies. However, history has repeatedly shown that such optimism often ignores the objective realities of technology implementation, ultimately leading to a collapse fueled by supply-demand imbalances, debt accumulation, and inflated valuations.
The 19th-century railway boom in the UK and the US illustrates the destructive power of overcapacity. The report notes that the success of the Liverpool and Manchester Railway in 1830 ignited an investment frenzy in Britain, with railway stock prices nearly doubling between 1843 and 1845.
By 1847, spending on railway construction had surged to a record 7% of Britain’s GDP. A tightening of liquidity eventually triggered a financial crisis in October 1847, causing the railway index to plummet 65% from its peak.
The report states that the U.S. railway boom culminated in the Panic of 1873; the New York Stock Exchange was forced to close for ten days, and losses from corporate bond defaults between 1873 and 1875 reached 36% of face value.
After U.S. railway track mileage peaked at over 13,000 miles in 1887, overcapacity led to a collapse in transport prices; by 1894, approximately 20% of U.S. railway mileage was in receivership.
The electrification boom of the 1920s exposed the fragility of pyramid-style capital structures.
The report points out that the proportion of households with electricity climbed from 8% in 1907 to 68% in 1930, though this process was concentrated primarily in urban areas.
Wall Street was deeply involved in this boom; utility company stocks and bonds were marketed as safe assets suitable even for “widows and orphans,” and by 1929, holding companies controlled over 80% of U.S. power generation.
The report notes that following the 1929 stock market crash, the major utility group Insull went bankrupt in 1932, reportedly wiping out the life savings of 600,000 small investors. Construction spending by U.S. electric utilities plummeted from a peak of approximately $919 million in 1930 to $129 million in 1933. The internet boom of the late 1990s demonstrated that innovation does not equate to profitability.
BCA notes that between 1995 and 2004, the annualized productivity growth rate of U.S. non-farm businesses reached 3.1%, far exceeding the rates seen in subsequent periods.
However, the ratio of tech-related capital expenditure to GDP surged from 2.9% in 1992 to 4.5% in 2000; this overinvestment placed immense strain on corporate balance sheets.
The report points out that free cash flow in the telecommunications sector peaked in late 1997, declined steadily thereafter, and then plummeted in 2000. After rising sixfold between 1995 and 2000, the Nasdaq Composite Index crashed by 78% over the following two and a half years.
Repeated oil booms perfectly illustrate the cyclical nature of supply-demand imbalances.
BCA notes that following the discovery of massive oil reserves in East Texas in 1930, daily production surpassed 300,000 barrels within 12 months; however, the deepening Great Depression caused oil prices to crash to 10 cents per barrel.
In 1985, Saudi Arabia abandoned production limits, causing oil prices to briefly drop to $10 per barrel.
Between 2008 and 2015, the U.S. shale oil boom drove crude production from 5 million to 9.4 million barrels per day; meanwhile, OPEC’s refusal to cut production in 2014 caused prices to fall from $115 per barrel in mid-year to $57 by year-end.
Five Common Patterns: The Inevitable Path from Boom to Bust
By reviewing the rise and fall of four typical booms, BCA Research identified five common patterns that serve as a key benchmark for assessing the trajectory of the current AI boom. Specifically:
The first pattern is that investors overlook the S-curve of technology adoption. Technology adoption never proceeds linearly; instead, it follows an S-curve pattern: adoption by early enthusiasts, followed by mass adoption, and finally, uptake by laggards. Stock prices typically rise during the first phase and peak midway through the second phase—specifically when the rate of adoption growth shifts from positive to negative.
The AI sector is currently exhibiting this characteristic: while most enterprises express an intention to increase AI usage, actual adoption rates are showing signs of stagnation, with some metrics even declining in recent months. This divergence between intent and action is a classic signal that technology adoption has entered the latter part of the second phase.
The second pattern is that revenue forecasts underestimate the magnitude of price declines.
New technologies often command pricing power due to scarcity in their early stages, but prices inevitably plummet as adoption spreads and competition intensifies. Between 1998 and 2015, internet traffic grew at an annualized rate of 67%, yet the price per unit of data transmission fell sharply in tandem. Solar panel prices have steadily declined since their inception, dropping 95% between 2007 and the present.
The AI industry is repeating this pattern: since 2023, the introduction of faster chips and superior algorithms has driven the price of tokens down by more than 99%. Despite the emergence of new applications like video generation, user willingness to pay for such services remains unclear.
The third pattern is that debt becomes the primary source of financing.
In the early stages of a boom, companies can usually fund capital expenditures through retained earnings; however, as the scale of investment grows, debt increasingly becomes the main source of capital.
In October 2025, Meta announced a $27 billion data center financing agreement structured through an off-balance-sheet special purpose entity. Meanwhile, after securing $38 billion in loans, Oracle raised an additional $18 billion in the bond market, bringing its total debt to nearly $96 billion.
Of even greater concern are “new cloud providers” like CoreWeave; as of October 2025, CoreWeave’s credit default swap (CDS) spread had risen from 359 basis points at the start of the month to 532 basis points. The fourth pattern is that asset prices peak before a decline in investment occurs.
Historically, during capital expenditure booms, the prices of assets such as stocks often peak before actual investment spending begins to fall. Even as investment spending retreats from its highs, the absolute level may remain elevated, further exacerbating overcapacity. This implies that investors who wait for clear signals of a “decline in investment” before acting often miss the optimal window.
The fifth pattern is the mutually reinforcing nature of a capital expenditure collapse and economic recession.
The bursting of a technology bubble typically unfolds in two stages:
The first stage involves the subsiding of tech hype and the emergence of overcapacity; the second stage sees the collapse of capital expenditure dragging down the broader economy, leading to deteriorating corporate earnings and creating a vicious cycle.
The report notes that the 2001 US recession was not triggered by a deterioration in economic fundamentals but rather by the collapse of capital expenditure following the dot-com bubble burst. While the rise of the real estate bubble in 2002 temporarily mitigated the shock of the dot-com crash, it remains uncertain whether a new bubble will emerge to offset the impact of a potential collapse in the AI boom.
Risk signals for the AI boom: A turning point within 6 to 12 months
Based on a comparative analysis of historical patterns, BCA Research believes the AI boom is following the trajectory of past bubbles and expects it to end within the next 6 to 12 months. This assessment is based on multiple risk signals currently emerging in the AI sector.
Regarding technology adoption, the pace of actual AI implementation has failed to keep up with the frenzied expectations of capital markets; corporate adoption rates have stagnated, and consumer willingness to pay for AI applications remains unproven.
Regarding price trends, sharp declines in token prices have signaled deflationary pressure, while the commercial value of new applications—such as video generation—remains questionable.
Regarding debt risk, the financing structures of AI-related companies are becoming increasingly reliant on debt, and credit risks for some firms have already begun to surface. The report recommends focusing on four key leading indicators:
First, revisions to analyst expectations regarding future capital expenditure; if expectations that have been steadily rising begin to plateau, it could signal danger.
Second, GPU rental costs; these costs began to decline after May 2025.
Third, the free cash flow status of hyperscalers; while currently at historically high levels, it has recently shown signs of deterioration.
Fourth, the emergence of a “Metaverse moment”—a scenario where an AI company’s stock price falls following the announcement of a major project—which would serve as a clear indicator of a shift in market sentiment.
For investors, BCA Research recommends adopting a “moderately defensive” strategy at present. This entails maintaining a neutral allocation to equities in the short term (three months), moving to a moderate underweight position in the medium term (12 months), and further increasing defensiveness over the coming months.
Specifically, investors should closely monitor the four aforementioned leading indicators to avoid reactive adjustments only after investment spending has clearly declined. At the same time, they should consider defensive sectors and high-quality bonds to hedge against potential significant volatility in AI-related assets.