When everyone believes the system is stronger, it is usually more fragile. That is the paradox of booms. The AI wave looks durable. It has revenue, hard assets, and real users. But a simple stress test exposes load-bearing points that do not tolerate error. Call it the four Os of a bubble: overcapacity, overconfidence, overconcentration, and opacity. Run this test on today’s AI buildout and the fault lines appear.
Bubbles that build things tend to break hardest. Railroads in the 19th century, long-haul fiber in 1999, shale wells a decade ago. Capacity arrives before economics. AI is following the same script. Big Tech is laying billions into chips, data centers, and power. Google’s plan to spend roughly 15 billion dollars to build a major data center hub in India shows the scale. Investors may take comfort because data demand will likely keep rising. That was true for fiber and oil too. Yet ownership matters. Overcapacity shifts bargaining power to customers. Prices compress. Balance sheets take the hit. When capex precedes cash flow discipline, the gap must close somewhere: in margins, in write-downs, or in forced consolidation.
Narratives turn uncertainty into a sales pitch. In markets, that often means optionality gets priced as inevitability. AI promises productivity leaps, new revenue pools, and platform control. Many models will deliver. But the base rate says adoption follows S-curves and returns concentrate in a few hands. Even some AI leaders are cautious. Sam Altman has warned that investors may be overexcited and that a bubble risk exists. The CEO of Deutsche Bank’s asset manager called this cycle unprecedented, with no clear playbook. When experts say there is no history to lean on, the rational move is to lean on probability. Fat tails cut both ways. You can be right on the destination and wrong on timing, pathways, and who actually earns the cash.
AI depends on narrow chokepoints: a handful of chip makers, a cloud oligopoly, limited high-voltage interconnects, and constrained power markets. That is concentration risk. It makes returns look smooth until a single point fails. It also amplifies policy and geopolitical shocks. Goldman Sachs has flagged that a pullback in AI spending by Big Tech could compress valuations by 15 to 20 percent. That estimate is not magic. It reflects dependencies. If hyperscalers slow orders, chip suppliers, contract manufacturers, power developers, specialized REITs, and a long tail of startups feel the same shock. In network theory, dense hubs transmit stress fast. In practice, it means a cautious procurement update on one earnings call can become a sector-wide de-rating in a week.
Opacity feeds bubbles because it delays feedback. With AI, opacity is technical and financial. Technical opacity: model quality metrics are easy to game and hard to audit, safety is probabilistic, and reliability varies by context. Financial opacity: cost accounting for training and inference is unsettled. Some firms capitalize software and infrastructure while expensing credits and incentives, muddling unit economics. Are AI features expanding revenue or cannibalizing existing lines like search, support, or software seats? Are gross margins flattered by introductory pricing or subsidized usage? The lack of standardized disclosure keeps investors guessing. Markets hate surprises, but opacity makes them likely. The moment a firm admits that usage is high but monetization lags, the valuation math snaps back to earth.
Herding is not just a retail phenomenon. It is executive game theory. No CEO wants to be the one who underspent and lost the platform shift. So capex races continue until balance sheets or boards push back. This is a prisoner’s dilemma. The dominant strategy is to invest, even if the group outcome is overbuild. AI raises the stakes because speed compresses cycles. The Federal Reserve’s Michael Barr has warned that the same attributes that make AI attractive can create market-wide risks through automaticity and herding. If credit, trading, and procurement decisions are augmented by models tuned on the same data, correlation rises. In stress, correlation goes to one. We learned this lesson in 2008 with structured credit. We may relearn it in 2026 with compute.
Run a basic inversion test. Assume the story is right, but the cadence slips. Hyperscalers trim 2026 capex by 10 to 20 percent to digest buildouts, improve efficiency, or wait for new chips. Goldman’s scenario of a 15 to 20 percent valuation compression is plausible, not punitive. Chip suppliers miss volume targets; backlog converts slower; power purchase agreements get renegotiated; AI infrastructure REITs guide lower; startups built on cloud credits face real bills; second-order effects hit local construction, utilities, and municipal financing around new data center hubs such as India. Retail investors may be less exposed, with surveys suggesting only about 30 percent have bought AI stocks. That lowers forced selling at the edge. It does not remove institutional flows at the core.
The antidote to the four Os is not pessimism. It is design. Firms should favor variable over fixed cost, modular over monolithic builds, and demand-backed expansion over preemptive scale. Lock in power where it is cheap and reliable, not just available. Diversify chip supply and avoid single-vendor dependence where possible. Push for open standards to reduce switching costs. Focus on products with repeatable unit economics and pricing power. Accept that small, specialized models can beat massive ones on cost and latency for many jobs. Disclose AI revenue, cost of compute, and capex-to-revenue glidepaths with clarity. Investors should treat compute cycles like commodities cycles. Demand hurdle rates that clear a real cost of capital. Do not confuse model benchmarks with cash flow.
Every transformative technology carries both overbuild and underestimation. Railroads bankrupted many investors; they also built modern commerce. Fiber glut crushed telecom equity for a decade; it later enabled cloud and streaming. The lesson is not to avoid waves. It is to price fragility. Tight systems snap. Redundant, cash-generating systems bend and gain. In classical engineering, bridges are load-tested beyond expected stress to reveal weak joints before they fail. The four Os provide a similar mental load test for AI. Overcapacity shows up in build plans. Overconfidence shows up in guidance. Overconcentration shows up in vendor maps. Opacity shows up in footnotes. If you cannot pass that test on a name or a theme, you are not investing. You are hoping.
Instead of asking who wins if AI works, ask who survives if AI stalls. Who can cut capex without losing the customer? Who earns returns on legacy businesses while AI matures? Who can pivot from bleeding-edge to efficient-edge when the cost curves flatten? The market will eventually separate narrative from cash. Retail caution today may be a tell rather than a lag. Professionals crowded into the same trades reduce resilience. The only durable edge is a margin of safety. In a boom built on speed, the scarce asset is patience. The four Os do not predict a crash. They map where strain will show. In bubbles, what you do not see is what hurts you. In AI, that is still the load-bearing part.
