The argument:
Three of the four mechanisms that have preceded every major asset bubble of the past fifty years are now present in AI markets.
Hyperscaler capex of $725 billion in 2026 against enterprise AI revenue of roughly $100 billion. The three largest AI-native companies : OpenAI, Anthropic, and SpaceX/xAI are asking public markets for over $200 billion at trillion-dollar valuations, against a US IPO market that raised $45 billion in all of 2025.
The technology is real. The bubble is real. Both have always been true simultaneously in productivity-revolution cycles. The first survives; the equity that financed it usually does not.
The framework: four mechanisms across fifty years
Bubbles look distinctive at the time. In hindsight, they rhyme. Japan in 1989, the Asian crisis of 1997, LTCM in 1998, dot-com in 2000, US housing in 2007, two Chinese A-share bubbles, crypto in 2017 and 2021 each looked novel to its participants, each shared the same anatomy in retrospect.
- Liquidity arrives Cheap money seeds every bubble, engineered by central banks or structural liquidity.
- Retail arrives By the time the public is excited, institutional positioning is complete.
- Leverage hides in the Not on speculator books but in supply chains, financing structures, vendor arrangements. The hiding is part of the bubble.
- A trigger ends Almost always a rate hike or a rule change & not a sudden recognition that the asset is overvalued. The marginal buyer simply runs out.
Three historical bubbles worth keeping in mind
Dot-com (1995–2000) — the technology being real did not save the equity
The internet was not an illusion. Every claim made for it in 1995 was on its way to being true by 2000. The Nasdaq still fell 78%. Cisco lost 89%, Amazon 93%. WorldCom became the largest US bankruptcy of its time. The fiber-optic capacity laid in the boom survived to power today’s internet. The equity that financed it largely did not.
Japan (1986–1991) — concentration produces decades of consequence
At peak, Japanese equities were 45% of global market cap. Land in central Tokyo traded above $1 million per square meter. The Nikkei did not reclaim its 1989 high until February 2024 — a 34-year drawdown. When market cap concentrates in a single thesis, the depth of the fall depends on how tightly the system has tied itself to it.
US housing (2002–2007) — hidden plumbing turns downturns into systemic events
The lesson is not predatory subprime lending. It is the structural inability of a leveraged financial system to locate its own exposure when conditions reverse. Trillions in mortgage-linked exposure, 30:1 leverage at investment banks, AAA ratings based on pre-bubble default histories.
Running the diagnostic against AI in 2026
| Mechanism | Status | Evidence |
| Liquidity | Present | $200B+ in annual hyperscaler operating cash flow sustains the spend; ZIRP-era policy seeded the position. |
| Retail involvement | Present | Magnificent Seven ≈ ⅓ of S&P 500. The 2026 IPO wave brings public capital in directly. |
| Hidden leverage | Present | Nvidia takes 85% of revenue from six customers; the top four account for 60%. Nvidia → OpenAI → Azure → Nvidia chips. Same structure as 1999 telecom vendor financing. |
| Trigger | Not yet visible | True by definition of every bubble that has not yet popped. |
Three additional signatures specific to AI:
- Capex vs $725B planned 2026 hyperscaler AI capex against ~$100B in enterprise AI revenue. Seven times wide and widening.
- The IPO OpenAI (targeting >$1T in September), Anthropic ($965B post-money, IPO this fall), SpaceX/xAI ($1.8–2T target) collectively asking public markets for $200B+ in
2026 vs $45B raised by the entire US IPO market in 2025. OpenAI alone has committed to $1.4T in data-center spending through the decade with no profitability path before 2030.
- The adoption An MIT study found 95% of enterprise generative-AI pilots produced no measurable EBIT impact.
Implications
Institutional investors: S&P 500 “diversification” is concentrated AI capex exposure marketed as something else. Position sizing should reflect that.
AI IPO buyers: fundamentals matter again on a five-year view that includes 2030 not on a 24-month view that assumes the next round arrives at a higher price.
Corporate leaders: the dot-com lesson holds. The technology will outlast the financial cycle. The question is whether to fund AI capabilities at peak prices or wait for fair ones.
Bottom line
Three of four mechanisms are present. The fourth trigger is invisible until it arrives. Every productivity-revolution bubble in history canals, railroads, electricity, radio, telecom, the internet corrected sharply before the technology’s value emerged. The technology stayed. The equity did not.
AI is a bubble. The technology will outlast it. The equity priced at the peak will not.
Disclaimer: This article is for educational purposes only and should not be considered investment advice. Mutual fund investments are subject to market risks. Consult a qualified financial advisor before making investment decisions.