In 1850, British railway shares had fallen to less than half the capital invested in them, according to investment records from the period. The railways would eventually transform Britain's economy, but most original investors never recovered their money.

Today's artificial intelligence infrastructure boom is following a remarkably similar pattern. Technology companies plan to invest $432 billion in AI infrastructure by 2026, according to Goldman Sachs research published in October 2025. To put this in perspective, that's nearly triple the $154 billion spent in 2023.

The scale of current building is staggering. In Northern Virginia, data centers now consume 2.5 gigawatts of electricity—enough to power 1.9 million homes, according to local utility Dominion Energy. Ireland halted new data center approvals after the facilities began consuming over 10% of national grid capacity.

These physical constraints represent something new in technology bubbles. During the 1990s telecom boom, companies could always lay more fiber optic cable. The only limit was money. But electricity generation and transmission take years to expand. Grid connections for new data centers now face multi-year waiting lists.

The financial projections for AI companies stretch credibility. According to FactSet consensus data as of October 2025, analysts expect one major AI chip company to grow revenue from $61 billion to $361 billion by 2031 while maintaining operating margins above 65%. For historical context, Intel's operating margins during its years of near monopoly in microprocessors peaked at around 35%.

The telecom bubble of the late 1990s offers the clearest precedent. WorldCom, which declared bankruptcy in 2002 with $107 billion in assets, had borrowed heavily to build fiber optic networks based on projections of infinite demand growth. By 2005, according to Federal Communications Commission data, approximately 85% of fiber optic capacity in the United States sat unused.

Yet that "overbuilt" fiber network became the foundation for today's internet economy. Companies like Google, Netflix, and Amazon built their businesses on infrastructure acquired at fraction of its original cost.

History suggests a predictable pattern. The Bridgewater Canal in Britain, completed in 1761, halved the price of coal in Manchester and made enormous profits. This triggered a canal-building boom in the 1790s. By 1810, most newer canals struggled to pay dividends despite heavy usage. The infrastructure proved valuable, but the builder’s lost money.

Similar dynamics played out with American railroads (major panic in 1873, followed by widespread bankruptcies), electricity infrastructure (the "War of Currents" between AC and DC systems led to massive stranded investments), and radio manufacturing (RCA's stock fell over 95% from its 1929 peak despite radio's continued growth).

Today's AI infrastructure race shows similar characteristics: competing technical standards (each major tech company is developing proprietary AI chips and frameworks), vendor financing emerging (companies increasingly offering payment terms to maintain sales growth), and physical constraints creating bottlenecks (power grid limitations in key markets).

The question isn't whether AI will transform the economy, it almost certainly will. The question is whether today's infrastructure builders will profit from that transformation. History suggests they won't.

The pattern across centuries is consistent: transformative technologies create infrastructure booms, overbuilding leads to price collapse, original investors lose heavily, and new companies build successful businesses on the cheap infrastructure left behind.

For investors, the lesson is clear: the real opportunities in transformative technologies often come not from financing the initial build-out, but from buying the infrastructure after the bubble bursts. It happened with canals, railways, electricity, and telecom. There's little reason to think artificial intelligence will be different.

Source: Goldman Sachs