250 Years of History: How Far Is the AI Capex Bubble From Bursting?

The AI spending binge will eventually end — but probably not when investors expect it to.

The hard part about capital expenditure bubbles isn't dancing when the music plays; it's knowing when to sit down. As summer fades into fall, investors should keep dancing to AI's rhythm.

With just four months left in 2026, we understand the urge to head for the exits. The S&P 500 is up roughly 12% year-to-date, but barely above its early-June levels. The longer the index stagnates, the easier it is to focus on what could go wrong. And among all the concerns lurking, nothing looms larger than the massive sums companies are pouring into AI.

The scale of AI capital expenditure is staggering. Since early 2024, roughly $500 billion has been spent on chips, $350 billion on power infrastructure, $200 billion on construction, and $100 billion on networking equipment. This total far exceeds the entire S&P 500's capital expenditure for all of 2021 — approximately $575 billion — the year before OpenAI released ChatGPT.

As long as companies were paying for AI out of free cash flow, nobody seemed to mind. But now the former cash cows — Amazon, Alphabet, Meta, Microsoft — are taking on debt to finance AI data centers, and concerns are growing. These concerns aren't unfounded. The amounts they plan to spend are staggering — roughly $2 trillion over the next two years — and the忧虑 has started showing up in credit markets: the cost of insuring against credit default has surged.

Wall Street is anxious too. AI stocks have fallen 20% from their 52-week highs set in June. That's enough to make investors ask: even if the AI buildout isn't over, has the AI trade run its course?

History says no.

Capital expenditure bubbles almost always burst — but rarely when investors suspect they will. That's been the pattern in the U.S. from the railroad boom through the internet bubble and beyond. While the AI capex surge will almost certainly end badly in some form, the odds of it ending badly right now aren't high. According to a Barron's analysis, over 250 years of American history, the economy has been able to absorb roughly 25% of GDP in spending on transformative technologies before things truly get problematic. AI hasn't come close to that threshold yet — and won't for years, suggesting the AI trade — and this stock market rally — still has more room to run.

"I think we're going to continue going higher. This is going to be bigger and longer than anyone thinks," said Ben Reitzes, tech research lead at Melius Research.

The U.S. has a long history of capital expenditure booms and busts. From the great railroad expansion of the 1860s, to the electrification boom of the 1920s, to the 1990s internet bubble and the early-2000s housing bubble, they all followed a similar pattern: a new technology emerges, investors pile in, valuations soar, equity and debt financing increase along with market concentration problems, and then comes the crash. But booms don't end on their own. They need an external shock — a policy trigger like Federal Reserve rate hikes, or an event like the 1906 San Francisco earthquake.

These cycles are regular enough that they can serve as a framework for thinking about where AI sits in its own cycle. One such framework is what might be called the "Rule of 25" — the total spending on transformative technologies the U.S. economy can absorb as a share of overall economic output during a boom period. For example, during the early railroad boom of the 1860s, U.S. GDP was approximately $10 billion per year (according to the National Bureau of Economic Research), and railroad spending eventually reached $2.5 billion before the crash of 1873. Similarly, in the late 1990s, roughly $1.5 trillion was spent building internet infrastructure, while the U.S. economy was only $6 trillion in size. The same pattern held for industrial and electrification buildout in the 1920s.

Using this standard, one can estimate top-down how much more AI can spend before the economy is primed for a crash.