The AI boom may follow the path of the 19th-century railway bubble, and builders may not be the ultimate winners

Nvidia's market capitalization surged from $1 trillion to a peak of $5.8 trillion, before falling back to an 18-month low. This wave of ups and downs reminds Goldman Sachs strategist Bobby Molavi of the British railway mania 180 years ago.

In his latest "Ruminations" strategy note, Molavi raises a core question: If this is a bubble, who will ultimately end up with the value? His answer, drawing from the experience of several historical capital expenditure booms, is—"most likely not those who built it."

Bubbles and eventual utility are two different things. Technological innovation can be epoch-making but may not create value immediately. Take previous bubbles for example—the internet, railways, and tulips—none of these technologies disappeared after their respective bubbles burst. I personally am quite fond of tulips.

Tulips survived, the internet survived, and so did the railways. What really changed was who ultimately ended up with the value—the overall tulip industry, the application layer of the internet, and the "overall economy that took off thanks to the infrastructure." But he quickly points out:

Early capital chasing these bubbles got hurt... especially when speculation and hope were 'valued', becoming detached from fundamentals such as income and cash flow.

In other words: You might be totally right about the technology, but could still lose money on the stock.

500 Companies, 8,000 Miles of Railway: The Bill From Back Then Is Still Unclear Today

Molavi's most referenced historical case is the British railway boom in the 1840s:

During the British railway mania of the 19th century, 500 new railway companies competed to build 8,000 miles of new track (the equivalent of 20 times England's land length), costing more than the country's total annual national income.

By comparison, annual U.S. data center capital expenditures currently reach several hundred billion dollars, and cloud computing giants, with seven weeks left in 2026, have already exceeded their annual debt ceiling projections. While the scale hasn't yet reached the point of "using up a year's GDP," the direction is clear.

The dot-com bubble followed a similar playbook. Cisco and Qualcomm "rose and fell," while 360networks, AltaVista, and Webvan "burned brightly and then went out completely." The fiber was laid, the networks built, but a decade later, it was a new batch of players profiting from this infrastructure.

Molavi's conclusion is relatively restrained: "The value of innovation is beyond doubt... but the price of innovation and the timing of returns may be open to question."

“Every Industry Is Tech, and Tech Is Every Industry”

So what’s different this time? Molavi believes that, in some ways, the risk is higher.

During the dot-com bubble in 2000, the technology sector mainly served itself, so a problem in one link had a relatively limited spillover effect. Today, "every industry is tech, and tech is every industry." The degree of cyclical dependence brought about by AI is "rare":

"This is a mutually reinforcing, interdependent theme—Nvidia’s deals with OpenAI, Nvidia holding Intel shares, Oracle purchasing Nvidia GPUs, major hyperscalers owning stakes in AI startups... For quite some time this year, we have been in a large-scale FOMO trade—a gold rush in the form of AI capital expenditure and technological dominance. The fear of 'doing nothing' far exceeds the risk of 'acting too early or excessively.'"

Even after last week's turbulence, OpenAI still completed about $11 billion in financing at a $1.25 trillion valuation, and Anthropic is reportedly preparing for an IPO sized between $80 billion and $100 billion.

Momentum Is a Double-Edged Sword

Why have valuations run ahead of cash flow to such an extent? Molavi points to market structure as the explanation.

In a market dominated by passive funds, ETFs, market-neutral hedge funds, and systematic quantitative strategies, price momentum itself has become a fundamental input:

In a market where the S&P 500 is increasingly driven by 5 to 10 stocks (sometimes only 1 or 2), more than 60% of this year's gains in the first half came from 6 stocks, and 35% from just 1 stock... More and more assets under management are wrapped up in index-tracking or index-hugging funds—the higher the prices go, the more people 'have to' buy—so prices rise even higher. This group will never be sellers when stocks are going up, because they make no judgment about fair value or whether prices have overshot.

Retail investors follow the winners (the Mag 7 and Nvidia, instead of "cheap but underperforming value stocks"), stoking the fire further. Molavi acknowledges that "a large part" of Nvidia's rise is due to fundamentals and is well deserved. But also "a large part" is technical; "it goes up for this reason, but it also falls for this reason." In a few weeks, the momentum factor erased all gains projected for 2026.

Cheaper, But Far From Cheap

So has the current market priced in these risks? Molavi lists the S&P 500's forward price-to-earnings ratio at past market troughs for comparison:

• March 2009: 10.4x

• March 2020: 13.3x

• October 2022: 15.7x

• Current (if last week was a short-term bottom): About 18x

His judgment: "Cheaper, but not truly cheap."

He also admits that this comparison ignores growth rate differences, while U.S. corporate earnings are still growing at more than 30%. But this is a double-edged sword—if the AI capital expenditure cycle reverses, most of that earnings growth will also reverse, and multiplying an 18x P/E by cycle-top earnings is a very different proposition from 18x at cycle-bottom earnings.

Who Is Left Holding the Bag When the Music Stops

Molavi ends with a cautiously agnostic view: The value of AI itself is not in doubt, but the timing of the price and returns is.

Notably, there is also a structural difference between this cycle and the dot-com bubble. The dot-com bubble mainly hurt shareholders; the telecom bubble took bondholders down as well. This time, the AI infrastructure boom is increasingly relying on debt, SPVs, and vendor financing among the same group of companies. This means that if value migrates from the 'builders' the way it did in 1847 and 2001, the loss may not be limited to just stock prices.

Tulips survived, and GPUs will survive too. The question is, when the music stops, who is left holding the bag—and unlike the railway mania, whether this bill ends up on the entire market’s balance sheet.