I Tried to Prove AI Was a Bubble. The Evidence Made Me More Bullish.

​I did not begin with the assumption that AI must work as an investment theme. I began with the opposite question: what if AI is already a bubble? I examined the dot-com boom [1], China’s 2015 equity mania [2], the 2021 meme-stock episode [3], the 2026 Korean semiconductor surge [4], and the debated interpretation of Tulipmania [5]. These are comparisons, not a statistically validated bubble test.

My conclusion is not that AI is risk-free or that a bubble cannot form. The current cycle still looks to me more like a technological and earnings expansion than a broad, terminal bubble. But growing financing links, weaker cash conversion and speculative trading deserve more weight than a simple all-clear scorecard would suggest.

My working conclusion: remain heavily exposed while testing whether commercialization and operating returns justify the capital being committed.

​AI is closer to a Kondratiev-style long wave than a normal sector cycle. The term describes the hypothesis that major technologies drive multi-decade waves of investment and economic change. I use it as an analogy for diffusion and capital formation, not as an established timetable for stock-market returns.

Many investors compare AI with Chinese baijiu (premium liquor), biotech or clean energy: hot narrative, capital inflows, valuation expansion, and then mean reversion. The problem is scale. Those were sector stories. AI can reshape the common production tool used across finance, software, research, medicine, education, law and manufacturing. A sector cycle can revert to the mean; a general-purpose technology can rewrite the mean itself.

Human progress has not been linear. Before the Industrial Revolution, civilization was crawling. Then it learned to run. Electricity and automobiles put us in a car. Computers and the internet put us on an airplane. AI may be the first technology that makes the curve feel like a rocket.

Each wave inherits the infrastructure and knowledge of the previous one. AI goes further: it can increasingly assist coding, research, design and engineering. Innovation is beginning to accelerate innovation itself.

If AI scales broadly, the economic value and market wealth it creates could exceed the internet and mobile-internet eras.

AI addresses cognitive labor, research, operations and decision-making, and it arrives in a much larger capital market. SIFMA’s latest annual fact book puts global equity market capitalization at $157.8 trillion at year-end 2025 [6]. That scale creates opportunity; it does not establish AI’s addressable revenue, investor returns or a guaranteed bull-market duration.

I believe AI could widen the gap between owners of productive capital and people who depend mainly on labor income. As machines perform more cognitive tasks, ownership of businesses, models, data, brands and intellectual property, along with energy and compute infrastructure, may become more valuable. An IMF scenario model identifies a similar channel through capital returns, although its results depend on productivity gains and how AI complements labor. [7]

The historical evidence gives this concern context. As Table 1 shows, the top 0.1% share of U.S. household net worth rose from 8.7% in Q4 1989 to 15.0% in Q2 2026. Over the same period, the next 9% below the top 1% saw their share fall from 37.9% to 36.4%. Wealth became more concentrated at the very top; not every wealthy group increased its share. [8]

​Table 1. U.S. Household Net Worth Shares by Wealth Group

Table 2 summarizes my assessment. “Supportive” means consistent with continued expansion; “mixed” or “yellow” means the evidence is incomplete or the risks are rising. These judgments are neither probabilities nor a mechanical trading rule.

Table 2. Ten Indicators of a Late-Stage AI Bubble

1. Valuation

I do not call something a bubble simply because its price rose. I ask whether earnings, free cash flow and returns on capital can support it. At the October 23 close, Stock Analysis reported NVIDIA at about 40.1 times trailing free cash flow [10]. That is a demanding multiple. Forward P/E ratios depend on forecasts and should not be compared casually with historical trailing multiples. My case rests on future earnings delivery, not on declaring the sector cheap.

2. Capex Is Already Showing Up in Revenue

The latest reported June-quarter results show Microsoft Cloud revenue of $59.3 billion, up 27% year over year; Google Cloud revenue of $24.8 billion, up 82%; and AWS revenue of $42.2 billion, up 37% [11][12][13]. Microsoft also reported $678 billion in commercial remaining performance obligations. These support the demand thesis, but cloud totals are not pure AI revenue and backlog is not cash already earned.

Cash conversion is the counterweight: Amazon’s trailing twelve-month free cash flow was negative $7.6 billion, versus positive $18.2 billion a year earlier, primarily reflecting higher AI investment [13]. Strong supplier sales do not prove that customers will earn adequate returns. NVIDIA’s latest quarterly revenue of $96.2 billion and its $108 billion Q3 guidance support continued expansion, but the guidance remains a forecast [17].

3. GPU Rental Pricing

GPU rental pricing remains one reason I stay constructive. Figure 1 shows H100 one-year contract prices recovering from their late-2025 low, B200 composite pricing rising sharply, and A100 pricing remaining relatively stable. These patterns are consistent with tight conditions in parts of the compute market, with no broad rental-price collapse across the series shown. [14][15] Alongside the earnings and commercialization evidence discussed above, this supports my view that the AI expansion still has fundamental support.

Figure 1. GPU Rental Pricing: H100 One-Year Contracts and Spot-Contract Composites

Falling rental prices alone would not overturn my thesis. Efficiency improvements, technological substitution and additional supply can lower prices even as adoption expands. I would become more cautious if sustained declines across comparable rental markets coincided with weaker utilization, contract cancellations and deteriorating customer economics. Prices alone cannot establish utilization or customer profitability.

4. Leverage

Leverage can turn a valuation correction into a financing crisis. The Federal Reserve’s May report described resilient banks alongside record-high hedge-fund leverage [21]. The IMF’s October analysis adds a more direct AI warning: 14 transactions since 2023 involved about $300 billion of equity and $650 billion of debt against $1.4 trillion of future revenue commitments [18]. Those commitments are not realized sales. Stress at a major customer could simultaneously impair revenue, investment values and debt. I therefore treat financing risk as yellow, with greater urgency than before.

Speculation also deserves a less reassuring label. Cboe reported that estimated retail options flow reached a record in September, even as overall Q3 options activity moderated [19]. Renaissance Capital’s public Q3 preview counted 30 US IPOs raising $32.8 billion, including SK Hynix’s $26.5 billion offering [20]. That concentration does not establish a broad flood of weak issuers, but neither does it rule out excess in private AI markets.

5. Vertical Price Action

Acceleration matters more to me than a large cumulative gain. What makes me cautious is not just that an asset rises a lot, but that the final leg becomes unusually steep or almost vertical. That pattern often reflects a shift from fundamentals-led repricing to momentum, leverage and fear of missing out.

Table 3 provides historical context for that warning sign. The episodes have different assets, contract structures and measurement windows; they are not comparable return benchmarks or a threshold that identifies a bubble. In particular, a semiconductor stock rally can reflect both earnings growth and speculation. I therefore view vertical price action as a warning signal that deserves attention, but not as conclusive evidence of a market-wide bubble.

Table 3. Historical Context for Rapid Price Acceleration

​​​​​Figure 2 translates the scorecard into my five-stage framework. The larger marker highlights commercialization and capex expansion as the dominant operating phase; Table 4 adds the qualifications. These stages can overlap: financialization is already visible, even if a broad terminal bubble is not my base case.

Figure 2. AI Cycle Framework: My Assessment of the Dominant Phase

• Cloud and enterprise-AI demand slow materially for several quarters, with backlog, paid usage and customer economics weakening together.

• Multiple GPU generations — not just one contract tier — begin to show sustained rental-price declines and weaker utilization.

• AI capex keeps rising but becomes increasingly dependent on financial engineering rather than customer cash flow and operating demand.

• Core AI companies stop receiving earnings upgrades while prices continue to accelerate vertically.

• Weak AI concepts flood IPO and private markets while leverage and retail speculation become socially broad.

AI may develop a bubble, as earlier technological revolutions have. What matters to me is whether capital formation has clearly detached from real demand. So far, I do not think the evidence establishes that across the sector, although financing risks are rising.

That is why I remain willing to be heavily exposed. I do not want to stay structurally underweight what may be the defining productivity revolution of the next decade. As long as the core investment thesis holds, I am prepared to stay invested through short-term volatility. A decade-long technological trend, however, does not guarantee a decade-long bull market.

I am not constructive because I do not see the risks. I am constructive because I have studied the risks — and, for now, the evidence still supports the thesis.

Disclosure: This is independent investment research and not investment advice. Assessments and portfolio views are the author’s. Sources distinguish reported results, management guidance, market observations and the author’s calculations