Infrastructure Concentration
Infrastructure concentration names the material substrate that the AI democratization narrative systematically omits. The builder celebrated for unprecedented productive capability operates atop an infrastructure whose ownership is more concentrated than any comparable infrastructure in modern history. Chip fabrication plants costing tens of billions of dollars. Data centers consuming electricity at rates comparable to small cities. GPU clusters designed by a handful of companies. Cooling systems consuming millions of gallons of water. And the trained models themselves — proprietary intellectual property representing billions of dollars in computation and the extracted, uncompensated intellectual labor of millions of human beings.
The Efficiency Imperative — Contrarian ^ Opus
There is a parallel reading that begins not with democratic ideals but with the thermodynamic reality of computation itself. Infrastructure concentration in AI emerges not from regulatory capture or market failure but from the irreducible physics of information processing at scale. The consolidation we observe reflects optimal engineering solutions to fundamental constraints: heat dissipation requirements that mandate geographic clustering, network latency boundaries that determine data center placement, fabrication precision limits that restrict chip manufacturing to facilities operating at the edge of what materials science permits. The concentration is the price of the capability — not a bug to be fixed but the structural requirement for the technology to exist at all.
The historical analogies to railroads and utilities miss the crucial distinction: those infrastructures delivered standardized commodities through networks that could be segmented and regionalized without loss of function. AI infrastructure delivers emergent capabilities that arise only at specific scales of integration. A model trained on a billion parameters behaves qualitatively differently from one trained on a trillion; the difference cannot be achieved through distributed ownership or democratic committees deciding training objectives. The intellectual property question similarly misunderstands the transformation: the training data is not enclosed but transmuted, like iron ore becoming steel. The original materials remain in the commons, accessible as they always were. What emerges is something genuinely new, whose value derives not from extraction but from the astronomical computational investment required to discover the patterns latent in the noise. The concentration reflects not capture but the minimum viable structure for this class of technology to function.
In the AI Story
Morozov's historical framework draws on the well-documented trajectory of previous critical infrastructures. Railroads, telephone networks, and electrical grids all followed a recognizable pattern: initial private development, rapid concentration, emergence of critical dependency, eventual imposition of democratic governance through regulation or public utility designation. In each case, the democratic response was resisted by the industries it constrained; in each case, the resistance deployed the language of inevitability and market superiority; in each case, the response was eventually recognized as necessary for benefits to be broadly shared rather than narrowly captured.
AI infrastructure is the most concentrated and the most consequential infrastructure humanity has built. It is governed by the least democratic mechanisms of any critical infrastructure in modern history. The decisions that determine how models are trained, what data they are trained on, what behaviors they are optimized for, what safety constraints they observe, and at what price they are offered are made by relatively small groups of people operating within institutions accountable to shareholders and boards but not to the populations whose lives their products reshape.
The dependency this creates is qualitatively different from previous infrastructure dependencies. The user who depends on AI for cognitive augmentation has restructured not just purchasing habits but thinking habits, professional capabilities, and her relationship to difficulty itself. The dependency operates at the level of cognition rather than consumption, which means switching costs are measured not in inconvenience but in diminished capacity — atrophied skills, lost tolerance for friction, restructured expectations that make operating without the tool feel not merely inconvenient but cognitively impoverishing.
The training data question adds a dimension without precise precedent. The models were trained on datasets incorporating the intellectual labor of millions — writers, programmers, artists, researchers — whose work was extracted from the public internet and incorporated into proprietary systems without compensation, without meaningful consent, and without any mechanism for contributors to share in the value their contributions made possible. Morozov has identified this as a form of enclosure: the conversion of a commons into private property that generates revenue for the enclosing institution while providing no return to the commons from which value was extracted.
Origin
Morozov has developed the infrastructure concentration analysis across multiple essays, most notably 'Socialize the Data Centres!' (New Left Review, 2015) and extended it with specific reference to AI infrastructure in 'Socialism After AI' (New Left Review, December 2025).
Key Ideas
Material substrate. The empowerment narrative operates at the application layer while concealing the infrastructure layer on which the application depends.
Historical trajectory. Previous critical infrastructures followed a predictable pattern from private concentration to democratic governance. AI infrastructure is traversing the same arc.
Cognitive dependency. AI infrastructure creates dependency at the level of cognition itself, producing switching costs that are qualitatively different from previous platform dependencies.
Enclosure of the commons. The training data regime represents the conversion of a vast intellectual commons into proprietary capital without compensation to contributors.
Appears in the Orange Pill Cycle
Scale Dependencies and Democratic Tensions — Arbitrator ^ Opus
The question of infrastructure concentration hinges critically on which aspect we examine. On the pure engineering constraints of AI systems, the contrarian view dominates (80/20): the physics of computation at this scale does mandate certain consolidations that democratic governance cannot wish away. The thermodynamic realities of cooling, the network effects of model training, and the precision requirements of chip fabrication create natural monopolies more absolute than any regulatory capture could achieve. These are not choices but constraints.
Yet when we shift to examining the social and political implications, Edo's framing becomes essential (70/30). The historical pattern of infrastructure capture is not negated by engineering requirements; if anything, the technical necessities make democratic oversight more urgent, not less. The cognitive dependency Edo identifies operates regardless of whether concentration emerges from physics or politics. The question is not whether concentration will occur but how its inevitable power will be governed. The enclosure argument occupies middle ground (50/50): while the contrarian correctly notes that training creates something genuinely new, Edo's point about uncompensated extraction remains valid — both can be true simultaneously.
The synthetic frame requires holding both the engineering reality and the political necessity in view simultaneously. Infrastructure concentration in AI is overdetermined: it would emerge from technical constraints alone, from economic incentives alone, and from network effects alone. This overdetermination means that democratic governance cannot prevent concentration but must instead focus on the terms of access, the distribution of benefits, and the mechanisms of accountability. The proper analogy is not to railroads but to nuclear power: a technology whose physical requirements create natural monopolies and whose social implications demand democratic oversight not in spite of but because of those requirements.
Further reading
- Evgeny Morozov, 'Socialize the Data Centres!' New Left Review 91 (2015).
- Evgeny Morozov, 'Socialism After AI,' New Left Review, December 2025.
- Kate Crawford, Atlas of AI (2021), on the material dimensions of AI.
- Jathan Sadowski, Too Smart (2020), on digital capitalism and extraction.