What if the bubble is real, but the hazard is elsewhere? The latest claim that artificial intelligence is a bubble 17 times the size of dot-com and four times subprime sounds explosive. It also misdirects. Focus on the numerator and you miss the mechanism. Markets do not break because something is expensive. They break when concentrations, feedback loops, and chokepoints turn a stumble into a cascade. We have seen this with railways in the 1840s, canals before them, telecom fiber after. The pattern is not valuation alone; it is structure. The AI trade has the three elements that convert excitement into fragility: extreme concentration, reflexive capital cycles, and single points of failure. If you want to measure risk, put away the megaphone and pick up a map.
Seventeen times dot-com makes for a crisp headline, but what is being compared to what? If the measure is nominal market cap, it ignores a larger economy, altered index composition, and a different rate regime. If it is enterprise value, it ignores lower financial leverage than subprime. Scale without context invites false confidence. Bubbles are about embedded assumptions. Dot-com’s embedded assumption was that demand curves did not matter on the internet. Subprime’s was that correlation could be diversified. The AI trade’s embedded assumption is that compute demand is infinite and evenly monetizable. That assumption is linear. Real systems are not. The relevant unit of analysis is not the size of the story. It is the fragility of the structure carrying it.
Technology now makes up roughly a third of the S&P 500, exceeding the dot-com peak. That is not merely a headline ratio; it is a wiring diagram. Indexation channels savings into the largest constituents mechanically, which in turn boosts their weight further, tightening the loop. The market’s breadth narrows, correlations rise, and the system becomes sensitive to a change in a few nodes. This is a classic common-mode failure: many independent decisions produce the same exposure because the decision rule is the same. Value-at-Risk and risk-parity frameworks, designed on recent vol, reinforce the concentration by reducing nominal risk assessments as prices rise. When a shock comes, de-risking propagates across funds that look diversified on paper but share the same top holdings. The second-order effect is the risk.
Earnings that justify AI valuations are increasingly a function of capital expenditures from a handful of buyers. Hyperscalers spend on accelerators and data centers, which lift semiconductor revenues and margins, which validate the narrative, which lowers the cost of capital for the same buyers, which funds the next wave of spending. George Soros called this reflexivity. Game theory calls it an arms race. No CEO wants to be the one who underinvests and loses the option value of AI adoption, so spending persists beyond standalone return thresholds. That is not inherently bad; it is how general-purpose technologies diffuse. But when two or three buyers set the marginal price of the supply chain, earnings momentum depends on their boardrooms and their power bills. If a single cohort sneezes, revenue lines across a dozen vendors catch a cold.
Hardware scarcity is not a myth; it is the business model. Leading-edge fabs are concentrated in one geography. High-bandwidth memory is concentrated in two suppliers. Data center buildouts depend on transformers, switchgear, and interconnects that have multi-year lead times. Power availability is the most underestimated input. The electrical grid is the levee in this story. Demand for compute is elastic; transmission capacity and permitting are not. That is a single-point-of-failure problem masquerading as a growth story. When the bridge begins to sway under harmonic load, adding cars does not help. An industry can solve these constraints over time. Markets, however, price perfection first and delivery later. Slippage at any chokepoint is a volatility amplifier.
Investors obsess over the supply of compute and neglect the demand for paid outcomes. AI, unlike past automation waves, targets cognitive work directly. That means white-collar wage pressure, re-pricing of professional services, and a reallocation of purchasing power. In the short run, displaced knowledge workers reduce demand for the very software and subscriptions investors assume will be cross-sold with AI features. The longer-run productivity dividend is real, but the path function matters for cash flows. Markets price point estimates. Economies live in distributions. If the distribution includes a period where unit demand lags capability, revenue projections built on linear adoption curves will be wrong in the direction that matters to leveraged expectations.
The regulatory environment is not a background variable. It is a switch. Export controls can remove entire markets from the revenue model overnight. Copyright and privacy rules can reroute monetization from data-rich incumbents to rights holders or push training offshore. Safety regimes can introduce licensing and audits that slow deployment and favor incumbents with compliance scale while killing margins for suppliers. Utility interconnection queues, environmental reviews, and local zoning can delay thousands of megawatts of data center power. These are not tail risks; they are scheduled realities. The path of rules is never monotonic, and it will not be synchronized across regions. Global firms face a combinatorial problem of compliance. Valuations that assume frictionless global scaling are backwards-looking.
Jumps break models, not drifts. The rise of ultra-short-dated options in mega-cap tech has turned daily flows into a volatility machine. Dealers who sell calls must buy stock into strength and sell into weakness, introducing intraday reflexivity and creating the illusion of stability when it is most dangerous. Index hedges that key off correlations can fail when the top few names gap on idiosyncratic news, forcing hedgers to scramble in the same direction. ETF liquidity, which feels deep, can thin out when the underlying basket gaps. This is not hypothetical; we have seen it across asset classes when concentration and derivatives mix. Liquidity is a state variable, not a constant. When everyone wants the same door, the rule is not capacity, it is fire code.
Great technologies often overshoot in markets and undershoot in life. Railways transformed commerce long after their first crash. Fiber lit the world years after telecom equity was wiped. The important distinction for investors is between fragility and antifragility. Fragile structures depend on continuous capital, smooth regulation, and single suppliers. Antifragile structures benefit from volatility, redundancy, and hard constraints. In this cycle, the fragilities are clear: concentrated buyers, bottlenecked supply chains, power scarcity, policy switches, and options-fueled flows. The antidotes are equally clear but less exciting: diversified cash flows, balance sheet self-reliance, pricing power not tied to a single hype cycle, and exposure to real-world constraints that cannot be copy-pasted. The AI story will persist. The market structure supporting it may not.
