When will the AI bubble burst?
Nobody can time it, and every specific date you'll see is a guess. What's documented: the BIS's June 2026 annual report compares the AI buildout to railway mania and the dotcom boom, hyperscalers are committing over $1 trillion in 2025–26 capex that outpaces their free cash flow, and MIT found ~95% of enterprise AI pilots showed no measurable profit impact. This is not investment advice.
Why — the first-principles explanation
Bubbles are impossible to time for a structural reason, not because analysts are lazy. A bubble is a coordination problem: it pops when enough people simultaneously decide it will pop. That makes the timing reflexive — dependent on beliefs about beliefs — which is a fundamentally different thing from predicting a physical process. If a reliable model for the pop date existed, people would trade on it, and the pop would move to the model's date. The prediction destroys itself. This is why the honest answer to "when" is always "unknowable," and why the useful question is "what would have to be true?"
So, what's measurable. The capex side is documented: per the BIS, the five largest hyperscalers are set to spend over $1 trillion on AI capex across 2025–2026, and critically, these commitments are outpacing earnings and free cash flow, pushing some firms toward debt financing. That last detail is the one that matters. Spending surplus cash on an experiment is survivable. Borrowing against it converts a disappointment into a credit event, and credit events are how technology corrections become recessions.
The returns side is where the gap sits. MIT's study of 300+ enterprise AI initiatives found roughly 95% delivered zero measurable P&L impact, on $30–40 billion of spending. Bulls say this is early — enterprise software always lags. Bears say a $1 trillion bet resting on a 5% hit rate is the definition of the problem. Both readings fit the data, which is exactly why this isn't settled.
The BIS parallel deserves precision, because it's routinely mangled. The BIS compared this to canal mania, 1840s British railway mania, 1920s electrification, and the 1990s dotcom boom, noting all shared one trait: a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify, and that these episodes ended in investment reversals that induced economy-wide recessions. Read that carefully. The BIS is not saying AI is fake. It's saying real revolutions and ruinous overinvestment routinely coexist. Britain's railways were transformative and railway investors were destroyed. Both are true. So "AI is real" is not a counterargument to "AI is a bubble" — historically, it's a prerequisite.
What to actually watch: whether AI revenue growth converges with capex growth, whether debt-financed buildout accelerates, and whether depreciation schedules on GPUs prove optimistic. Those are observable. The date is not.
An example that makes it click
Think about a small town where someone finds gold in the creek.
There really is gold — that's not in dispute. So the town builds. A railroad, three hotels, a bank. The construction is real, visible, and financed. Then a second wave arrives who didn't come for gold; they came because land prices are rising, and land prices are rising because people are arriving. That loop has no natural stopping point, which is why nobody can tell you the day it ends. It ends when enough people look around at once and think "wait."
And when it does end, three things happen that people insist are contradictory but aren't. The hotel investors are wiped out. The gold is still in the creek. And the railroad still stands — carrying freight for the next hundred years, for somebody who bought it at ten cents on the dollar. That's the historical pattern, and it's what the BIS is pointing at.
Key facts
- The BIS Annual Economic Report published June 2026 states the five largest hyperscalers are set to spend over $1 trillion on AI-related capex from 2025 through 2026, outpacing their earnings and free cash flow, with some issuing debt to finance it.
- The BIS compares the AI investment boom to 1830s canal mania, 1840s British railway mania, 1920s electrification, and the late-1990s dotcom boom — episodes that shared a genuine technological breakthrough attracting capital beyond justifiable commercial returns and ended in investment reversals inducing economy-wide recessions.
- An MIT study of 300+ enterprise AI initiatives (July 2025) found roughly 95% delivered zero measurable P&L impact on $30-40 billion of enterprise generative AI spending in 2024-25.
- The IEA projects data centre electricity roughly doubling from ~415 TWh (2024) to ~945 TWh by 2030 — physical infrastructure that persists regardless of financial outcomes.
- The Shiller CAPE ratio exceeded 40 in late 2025, a level previously reached only before the dot-com crash.
- No credible institution — including the BIS — has published a predicted date for a burst; the BIS frames its analysis as highlighting downside risk, not forecasting timing.
▶ The 60-second explainer (script)
When will the AI bubble burst? Nobody knows, and every date you'll see is a guess. But that's not a cop-out — there's a structural reason. A bubble pops when enough people simultaneously decide it will pop. That makes the timing reflexive: it depends on beliefs about beliefs. If a reliable model for the date existed, people would trade on it and the pop would move. The prediction destroys itself. So the useful question isn't when. It's what's measurable. Here's what's documented. The Bank for International Settlements — the central banks' central bank — says the five biggest hyperscalers will spend over a trillion dollars on AI capex across 2025 and 2026. And critically, that spending is outpacing their earnings and free cash flow, pushing some toward debt. That detail matters most. Spending spare cash on an experiment is survivable. Borrowing against it turns a disappointment into a credit event — and credit events are how tech corrections become recessions. On the returns side, MIT studied over three hundred enterprise AI projects. About ninety-five percent showed no measurable profit impact. Bulls say it's early. Bears say a trillion-dollar bet on a five percent hit rate is the problem itself. Both readings fit. Now, the BIS comparison gets mangled constantly, so hear it precisely. They compared this to canal mania, railway mania, electrification, and dotcom — and their point was that all of those were genuine technological breakthroughs that attracted more capital than returns could justify. Britain's railways were transformative and railway investors were destroyed. Both true. So AI being real is not an argument against a bubble. Historically it's a prerequisite. None of this is investment advice.
What authoritative sources say
People also ask
Is AI definitely a bubble?
Not established. The BIS says the boom resembles historical manias and highlights downside risk — it does not declare a bubble or predict a date. The capex-versus-returns gap is real and documented; whether it closes or breaks is genuinely open.
If AI is real technology, how can it be a bubble?
Those aren't opposites — historically they go together. The BIS's whole point is that canals, railways, electrification, and the internet were all genuine breakthroughs that still attracted more capital than returns justified. Real technology is arguably a prerequisite for a big bubble.
What's the single most important indicator to watch?
Debt-financed capex. Companies spending surplus cash can absorb disappointment. The BIS specifically flags that AI commitments now outpace hyperscaler earnings and free cash flow, pushing some to issue debt — that's what turns a correction into a credit event.
What happens to the data centres if it bursts?
They keep standing. The historical pattern is that investors get wiped out while the infrastructure is bought cheap and used for decades — the fiber laid in 1999 carried streaming video in 2006, just not for the people who financed it.
Should I make financial decisions based on this?
This page is not investment advice and no one here can time markets. If you're making decisions with real money, the documented facts above are inputs — the timing question is genuinely unanswerable, and anyone claiming otherwise should be treated with suspicion.