Will the AI Capex Bubble Burst? The "Rule of 25" Drawn From 250 Years of Market History

Data in this article is current as of 2026-08-29. Market information is time-sensitive, so check the date as you read.

August has been a strange month for US stocks. The index is up around 12% for the year, but it has gone basically nowhere since June. When a rally stalls, everyone starts hunting for a culprit — and this time everyone is pointing at the same place: how much money is being poured into AI.

I’ve been fooled by this kind of mood before. When I was younger I traded warrants and chased hot tips, calling myself a swing trader. The swings were real — my capital swung right off a cliff. Those years cost me a painful tuition, but that’s also where I learned something: there is often a gap of several years between “this feels like it’s about to blow up” and “this is actually blowing up.”

That Barron’s cover story from late August is essentially an attempt to measure that gap. Writer Al Root went through 250 years of US capital spending history and pulled out a line they call the “Rule of 25.”

First, how much has AI actually burned?

Since the start of 2024, global AI spending breaks down roughly like this:

For comparison: in 2021 — the year before ChatGPT existed — the combined capital expenditure of all 500 companies in the S&P 500 was about $575 billion.

In other words, what AI has burned through in the past two-plus years is roughly double what the entire top 500 US companies spent in a full year on all their plants, all their equipment, all their infrastructure.

And it isn’t done. The four hyperscalers plan to put in another ~$2 trillion over the next two years. I ran the numbers on this scale in Big Tech’s AI spending has passed the Apollo program. Six months later, the figures have only gone up.

What’s making the market nervous isn’t the amount, though. It’s that the source of the money has changed.

These companies used to build out of free cash flow — clean balance sheets, nobody complained. Now they’re borrowing.

For 250 years, America has broken it the same way

Raymond James strategist Tavis McCourt has mapped this cycle — from the railroad expansion of the 1860s, through the electrification boom of the 1920s, to the dot-com and housing bubbles of the 1990s and 2000s. The script barely changes:

1. A new technology arrives
2. Investor enthusiasm pushes valuations up
3. Heavy equity and debt raises begin
4. The market concentrates into a handful of winners
5. Bust

But here’s the detail most people skip: the boom doesn’t end on its own.

McCourt notes that every one of these cycles needed an outside shock to turn. It can be policy-driven, like a Fed rate hike. It can also come out of nowhere, like the 1906 San Francisco earthquake.

Put differently: if you’re watching valuations and waiting for the thing to “pop naturally,” you may be waiting a long time.

The “Rule of 25”: how the line gets drawn

Barron’s analysis goes like this: every time the US economy has absorbed investment in a transformational technology, things only get genuinely ugly once cumulative spending reaches roughly 25% of GDP.

The historical comparisons:

Apply it to today:

• US GDP is roughly $30 trillion
• Danger zone = $30 trillion × 25% ≈ $7.5 trillion
• Current cumulative spending: ~$1.15 trillion
• Another $5 to $6 trillion to go before the line is crossed

Hyperscalers are projected to spend $3.7 trillion globally through 2029, and they aren’t the only ones building — Oracle, SpaceX, Anthropic, and OpenAI are all in. At this pace, AI spending wouldn’t trip the Rule of 25 until the early 2030s, six or seven years into the boom.

But let me tell you how sloppy this line is

I’m not going to dress this number up as some precision instrument, because it has three obvious soft spots:

One, the units don’t match. That $1.15 trillion is global spending, while the $7.5 trillion threshold is calculated off US GDP. The numerator and denominator aren’t living in the same world.

Two, the sample size is three. Railroads, electrification, the internet — three dots and a line through them. Statistically, you could call that a coincidence.

Three, “25%” is not a law of physics. It’s an observation made in hindsight, not a circuit breaker built into the economy.

So the use of this line isn’t prediction, it’s calibration. What it tells you isn’t “the crash comes in month X of year Y.” It tells you “on a historical scale, calling the bubble over right now is probably too early.” The distance between those two statements is exactly the distance between getting out now and staying in for a few more years.

The real trigger isn’t valuation — it’s interest rates

If the boom needs an outside shock to end, where should you be looking?

Barron’s answer is clear: financing costs.

J.P. Morgan analyst Tarek Hamid estimates more than $5 trillion in AI data-center spending from 2026 to 2030, funded like this:

See the point? Nearly 80% has to come from outside. Which means the lifeline of the entire AI buildout runs through the bond market.

And long-dated Treasury yields are currently near their highest levels since 2007. 22V strategist Dennis DeBusschere puts it bluntly: the sensitivity of AI buildout names to 10-year yields is going to increase meaningfully from here.

Some people are already buying insurance

The tension in the debt market is already visible. Credit default swaps (CDS — basically insurance you buy on corporate debt) are getting more expensive: it now costs about twice as much to insure against an Alphabet default as it did at the start of the year, and the story is similar for Meta and Amazon.

To be clear, these numbers are still nowhere near the warning thresholds for investment-grade companies. The four hyperscalers have essentially no net debt and are expected to generate around $1 trillion in EBITDA in 2027. This is not a sign of financial distress.

It’s saying one thing only: Big Tech is no longer that debt-free juggernaut. The market has started treating these companies as businesses that borrow, carry leverage, and feel interest rates — like everyone else.

And then there’s debt you won’t find on the balance sheet

Meta’s approach deserves its own look.

They set up a special purpose vehicle called Beignet Investor to build 5 gigawatts of compute in Richland Parish, Louisiana, funded with about $27 billion in debt and $3 billion in equity. Meta holds only 20%; Blue Owl Capital and other private-equity players hold 80%.

But Meta is the tenant, pays the operating expenses, and — if it decides not to renew the lease, it’s on the hook for the debt. Which is also why Beignet’s debt is rated A+.

How to evaluate this kind of off-balance-sheet structure is honestly outside my wheelhouse, and I’m not going to pretend otherwise. I’ll only speak to the half I do understand: the question “whose risk is this, exactly?” cannot be answered by reading the financial statements alone.

So if you want to measure the leverage in the AI buildout using the balance sheets of the four hyperscalers, you’ll come up short. Part of the real leverage isn’t on that sheet.

The other side: this time, the money might actually come back

Talking only about risk isn’t analysis, it’s fear-mongering.

The spending is converting into revenue, and not slowly:

• Alphabet’s cloud business grew 82% year over year in Q2
• Anthropic’s annual revenue run-rate is around $65 billion, up from $9 billion at the end of 2025 — the business it added in eight months is roughly the size of 50-year-old Oracle
• SpaceX (which absorbed Elon Musk’s xAI in February) generated $3.2 billion in AI revenue in 2025 and could surpass $60 billion in 2027
• Nvidia beat expectations in its latest quarter and guided to 70% revenue growth for fiscal 2028

The most important part is asset efficiency. The S&P 500 on average needs $2 to $3 of assets to generate $1 of sales. A gigawatt of computing power costing $40 billion to build can produce about $40 billion in sales.

One to one. You almost never see that in traditional industries.

Of course, that’s the math under ideal conditions — it assumes that compute gets rented, stays rented, and holds its price. Three “ifs.” If any one of them breaks, that one-to-one turns into a very ugly number.

Where I stand on this

With the numbers laid out, my position is pretty simple: the bubble will burst eventually, but not now.

Nvidia CEO Jensen Huang has struck this same tone whenever he’s asked about a bubble, and I agree with his reasoning — we’re still in the infrastructure phase. The track is still being laid, the power is still being run, the facilities are still going up. The defining feature of an infrastructure phase is that money keeps going in and things keep getting built, but nobody has hit the moment where everyone realizes there’s overcapacity.

That moment will come. It came in 1873, and it came in 2000. The only difference is that when you measure it with the Rule of 25, it’s still a few years away.

So what should you do with your ETFs?

If you’ve read this far hoping I’ll say “sell” or “buy more,” I’m going to disappoint you.

First, get one thing straight: you’re probably already exposed and just don’t feel it.

If you hold VOO or any S&P 500 tracker, the four hyperscalers plus Nvidia make up a meaningful chunk. If you hold 0050, TSMC’s weighting is overwhelming — and TSMC is the one taking these orders. Both positions carry more AI capex exposure than most people assume. This is the point I keep hammering in the M7 vs VOO piece: what you think is diversification often isn’t.

So what should you do? Three things, all of them boring:

1. Don’t try to call the top. The value of the Rule of 25 isn’t telling you the crash date — it’s telling you that “running for the exit right now” has historically been too early. And since the boom needs an outside shock to end, that shock is by definition something nobody sees coming. You can’t guess it. The math in Does buying at the highs make you the sucker? is worth a look.

2. Shift your attention from “guessing direction” to “minding structure.” Your position mix, whether your automatic contributions are still running, whether your emergency fund can survive a 30% drawdown. You control these. You don’t control the market — focusing on what you can control is advice I’ve been repeating for years, and it still works.

3. If you genuinely want to track the risk, watch rates, not prices. The 10-year Treasury yield is the cost valve on this entire buildout. It grinding higher matters more than an AI stock dropping 5% in a day.

As for what I do myself, it’s so boring it’s hard to write into an article: just keep buying ETFs, keep buying, don’t stop the automatic contributions, and pull money out when I actually need it. Since this AI argument started, the only action I’ve taken is continuing to make my contributions.

No adding, no trimming, no “let me lock in some profit first.” Not because I have unusual conviction in AI, but because I have unusually little conviction in my own ability to call a top.

One more thing that should make you feel better: after a bubble pops, what was built stays.

The railroad bubble burst and the tracks remained, running for another hundred years. The dot-com bubble burst and the fiber remained — streaming and the cloud were built right on top of it. Financial markets need time to heal, but the foundation for the next expansion has already been poured by this one.

As for when the music stops — railroads took over a decade, the internet took five years. On a historical scale, we’re probably still mid-set.

FAQ

Further Reading

Disclaimer

This article is for educational purposes and does not constitute personal investment advice. Investment products carry risk, and past performance does not indicate future results. Before making any investment decision, please evaluate your own financial situation and risk tolerance. For personalized advice, please consult a licensed financial advisor or investment planner.

Data Updates and Sources

Data in this article is current as of 2026-08-29:

• Primary source: Barron’s cover story, August 28, 2026, “Is the AI Capex Bubble About to Burst? What 250 Years of Market History Tell Us.” by Al Root (subscription required). The views and figures cited from Raymond James (Tavis McCourt), J.P. Morgan (Tarek Hamid), 22V Research (Dennis DeBusschere), Melius Research (Ben Reitzes), and Evercore (Amit Daryanani) all come from that report.
• US 10-Year Treasury Yield (FRED, Federal Reserve Bank of St. Louis)
• US GDP (Bureau of Economic Analysis)

Any data updates will be noted at the bottom of this article. If you spot an error, please contact Lazy Da.

The Lazy Conclusion

懶大

金融科技新創 CEO | 獨立財務顧問

1. The “Rule of 25” says it’s still early: historically in the US, cumulative investment in a transformational technology has to reach 25% of GDP before things really break. AI is at roughly $1.15 trillion against a threshold near $7.5 trillion — at this pace, that’s the early 2030s.
2. Booms don’t end on their own; they need an outside shock: true from 1860s railroads to the 1990s internet. So “this feels expensive” is not the same as “this crashes tomorrow.”
3. The real trigger is rates, not valuation: nearly 80% of this buildout depends on outside financing, which makes the 10-year Treasury yield the cost valve. That’s the thing to watch.
4. You’re probably already exposed: VOO holds the four hyperscalers and Nvidia; 0050 holds TSMC. What you think is diversification often isn’t.
5. When the bubble pops, the buildout stays: the tracks remained, the fiber remained. Markets need time to heal, but the foundation for the next expansion is already poured.

My answer is the same as Jensen Huang’s: it will burst, but not now — we’re still at the laying-track stage. And over the past 250 years, every time someone shouted “it’s about to stop,” it was usually still early.

That said, I have far less confidence in that answer than I do in this one: rather than guessing which beat the music stops on, make sure the dance you’re doing is one you can get up from after a fall. What I do is keep my contributions running, then go to sleep.

That’s it.

📩

訂閱電子報，獲取更多理財觀點

立即訂閱

🚀 已有 1,000+ 讀者加入理財成長之路

About Lazy Da I’m Mars, CEO of Hippo Insurance, with over 15 years in insurance and finance. Every week I use “Lazy to Be Rich” to break down money concepts, with the goal of helping beginners in Taiwan skip the detours. All data in this article comes from official sources, with the cutoff date noted.