The AI system is recursively manufacturing the dependencies required to justify its own continued expansion.
This is more consequential than the claim that artificial intelligence is a speculative bubble. Investment does not simply respond to demand. It changes the environment in which future demand arises. Capital builds computational and physical infrastructure; organisations reorganise work around the resulting capabilities; those changes create dependencies on models, compute and supporting systems; and those dependencies increase the value of continued provision. Expansion changes the conditions against which further expansion is judged.
This is a familiar system dynamic. Arthur (1989) showed that technologies subject to increasing returns can become progressively favoured as adoption itself generates advantages for further adoption. Pierson (2000) extended the logic of increasing returns to institutional development: sequence matters because earlier choices alter the costs and probabilities of later ones. Path dependence does not mean that change becomes impossible. It means that the probability landscape is changed by the path already taken.
AI now has several of the conditions under which such reinforcement can occur. Compute capacity enables applications; applications encourage organisational adaptation; adaptation creates skills, interfaces, workflows and expectations organised around continued access to AI; those dependencies support anticipated demand; anticipated demand supports further investment in compute and infrastructure. Adoption becomes one of the causes of further adoption. The loop need not have been designed for this to occur.
Unruh’s (2000) account of carbon lock-in provides a useful parallel without implying that AI and fossil-energy systems are equivalent. His central argument was that technologies, organisations and institutions can co-evolve through path-dependent increasing returns until their mutual dependencies form a persistent techno-institutional complex. Infrastructure matters because it is not merely a passive substrate. Once built, it changes the relative costs and possibilities of what can happen next.
AI expansion is increasingly physical in precisely this sense. Data centres require electricity generation and transmission, cooling, land and semiconductor supply. Current development is already being displaced geographically by constraints on power, land and grid connections. These investments persist on different timescales from the models that motivate them. Software can change quickly; data centres, transmission infrastructure, financing arrangements, organisational structures and altered labour practices generally cannot change at the same rate.
That difference in persistence is central. The system does not have to remain technologically stable to become structurally difficult to reverse. A model can disappear while infrastructure remains. A provider can fail while workflows built around its services remain. Expectations can change while debts, contracts and electricity investments persist. The faster one layer changes relative to another, the greater the possibility that commitments made under one technological configuration survive into another.
This also changes what it would mean for an AI bubble to burst. There need not be one trigger or a clean return to previous conditions. If expected returns fall, investment can contract; reduced investment can alter infrastructure expansion and prices; those changes can affect adoption and valuations; changing valuations can further alter investment. The same reinforcing structure that accelerates expansion can transmit contraction through the dependencies accumulated during growth.
None of this requires AI to fail technically. A technology can remain useful while the financial expectations surrounding its expansion fail. Nor does a collapse in valuations erase the infrastructure, institutions and dependencies produced during the expansion. Path-dependent systems inherit their histories. What survives a correction becomes part of the initial conditions of whatever develops next.
The problem, then, is larger than whether AI is currently overvalued. AI expansion is reorganising the systems on which its own future demand depends. The more infrastructure, institutions and everyday practices are configured around it, the less future demand can be treated as independent evidence that the original investment was justified. Some of that demand will exist because previous investment helped create the conditions under which continued AI use became useful, economical or difficult to avoid.
The question is therefore not only whether artificial intelligence can justify the system being assembled around it, but how much of that system will eventually require artificial intelligence because we assembled it that way.
References
Arthur, W. B. (1989). Competing Technologies, Increasing Returns, and Lock-In by Historical Events. The Economic Journal, 99(394), 116–131. DOI: 10.2307/2234208
Pierson, P. (2000). Increasing Returns, Path Dependence, and the Study of Politics. American Political Science Review, 94(2), 251–267. DOI: 10.2307/2586011
Unruh, G. C. (2000). Understanding Carbon Lock-In. Energy Policy, 28(12), 817–830. DOI: 10.1016/S0301-4215(00)00070-7
Daedelus Kite 