The paradox of progress is that we need waste to find what works. Overbuild and you discover the efficient frontier. But waste also weakens the beams of any system if it piles up in the wrong place. The AI boom sits on that fault line. Jeff Bezos recently called it an “industrial bubble,” the kind where every experiment gets cash regardless of quality. The claim is seductive: spray capital and the winners will subsidize the learning. The question is not whether the frenzy builds something useful. It will. The question is whether the way we are building it leaves markets, firms, and investors more fragile than they think.
Bubbles can solve a coordination problem. In a new platform era, nobody wants to be the last to invest. A broad mania lowers the cost of capital and gets everyone moving at once. That can be good public-policy-by-accident. The risk is price signals fail when every pitch gets funded. Bridgewater’s co-CIOs have warned that investors are underpricing the actual risk in this run-up. The S&P 500’s strong gains rest on a narrow narrative: AI will raise profits fast enough to justify pulling forward a decade of returns. That narrative can be true in direction and still be wrong in timing, magnitude, or distribution. When capital is indiscriminate, the market stops ranking ideas. In game-theory terms, we swap discovery for herding. Herds move fast but off cliffs as easily as across valleys.
The comforting view is that this boom is different because it is not debt-fueled. The IMF notes much of the spend comes from cash-rich tech giants, not leveraged balance sheets. That reduces the odds of a 2008-style chain of forced selling. But risk does not vanish; it migrates. Equity duration risk rises when future cash flows get pulled forward. Option and structured exposure piles up in places that do not file 10-Ks. Vendors extend generous terms. Cloud credits and take-or-pay compute commitments look like soft leverage. Utilities and data center operators are scaling on the back of AI demand forecasts. If those forecasts slip, you have real assets with long lives matched to cash flows with short lives. Debt-light at the buyer can still be debt-heavy down the supply chain.
Concentration is both a symptom and a cause. The Magnificent Seven now account for roughly a third of the S&P 500 by market value, a share last seen at the dot-com peak. The ECB has flagged the obvious: when a few names carry the index, the system’s fate hinges on them meeting escalating expectations. That is reflexivity at work. Their stock strength draws passive flows, which pushes them higher, which draws more flows. Reflexive loops work until they do not; small disappointments turn into large drawdowns because every portfolio owns the same factor at the same time. These are not flawed companies. They are superb businesses priced for more-than-superb execution. Concentration is a silent leverage. You do not see it in debt ratios. You see it when correlations jump to one during a surprise.
The industry’s own base rates should sober the forecasts. An MIT survey found that the vast majority of large-company generative AI pilots stall before production. That does not mean there will be no productivity boom. It means we still lack repeatable, defensible use cases at the scale implied by today’s capex. History says general-purpose technologies diffuse slowly. They deliver high returns, but after a long period of reorganization, retraining, and process change. In the interim, boards fund experiments that create option value, not earnings. Investors are valuing the options as if they were near-the-money. The math of geometric returns applies: volatility in expected payoff destroys compound outcomes. High variance with slow adoption is a drain, not a boon, until the learning crystallizes.
One defense of bubbles is that they leave useful infrastructure. The fiber glut of 2000 later powered Web 2.0 at bargain prices. Railroads bankrupted investors but reorganized into the backbone of commerce. That logic works when the assets are long-lived and general-purpose. AI’s infrastructure is different. Data centers and GPUs depreciate fast. Nodes go obsolete with each architectural jump. Power and cooling are scarce and local. If model economics compress and inference shifts closer to devices, stranded capacity is not just possible; it is likely. The binding constraint is not capital, it is energy. Water, grid interconnects, transformer lead times, and permitting are the stubborn facts. This is an engineering problem before it is an equity story. If a bubble builds assets in the wrong places or with the wrong half-life, it leaves kindling, not cathedrals.
From a game-theory lens, today’s AI race is a Red Queen game. Every big player must run faster just to stay in place. There is no stable coalition to slow the spend because the payoff to defection is high and verification is hard. That structure breeds overshoot. Vendor ecosystems benefit in the short run, but cyclicality will be brutal. The last dollar of supply usually arrives right as demand growth normalizes. The industry then learns about operating leverage on the downside. That is the cycle we saw in memory, in networking, and in smartphones. It is not different this time in that respect; the only unknown is the amplitude. Anyone invoking picks-and-shovels immunity forgets that even shovel makers have inventories and debt.
For the bubble to be net-beneficial, failures must be small, information-rich, and contained. Funding should be equity-heavy at every layer, not just at the top. Asset build-out should favor flexibility: modular data centers, reconfigurable capacity, and contracts that match duration risk with revenue reality. Energy pricing must be real, not cross-subsidized by regulated ratepayers who cannot opt out. The ecosystem should be broad, not a tight cluster of megacaps whose luck or misstep steers global indices. And investors should be honest about base rates. If 95 percent of pilots fail, valuations should reflect that attrition and the compounding delay to cash conversion. An antifragile bubble does not pretend away variance; it uses it without betting the system on a single path.
The wrong question is Is it a bubble. It is. The right question is Where is the fragility. Identify the places where capital is indiscriminate, where assets age faster than the debt behind them, where expectations are indexed to a flawless S-curve. Check the quiet linkages: index concentration, vendor financing, power constraints, regulatory bottlenecks, and geopolitical chokepoints in chips and tools. Stress the scenarios everyone is waving away: a year of flat inference demand, export controls that deepen, grid delays that push go-lives to 2027. Do the probability math on what a 10 percent miss means when everyone is levered to the same story without calling it leverage. Bubbles can be good if they harden the system. They are dangerous when they make it brittle while telling you it is safer than ever.
