Limited Access Orders
In their 2009 collaboration Violence and Social Orders, North, John Wallis, and Barry Weingast developed a taxonomy distinguishing limited access orders from open access orders. Limited access orders are social arrangements in which a dominant coalition controls access to valuable resources — political, economic, or informational — and uses that control to generate rents. These orders are stable because the rents give coalition members an incentive to maintain the restriction and cooperate rather than fight. They are also economically inferior to open access orders, in which competition is broad, entry is unrestricted, and the creative destruction that drives long-term growth is permitted to operate. The AI economy is exhibiting structural features that raise the risk of limited-access formation. The computational costs of training frontier models, data requirements, and engineering expertise create natural barriers to entry. Platform dynamics create winner-take-most outcomes. The prestige hierarchy of the AI industry further concentrates talent and resources.
In the AI Story
The risk is not hypothetical. A handful of organizations — Anthropic, OpenAI, Google DeepMind, Meta's FAIR, a few Chinese national champions — control the frontier of AI capability. The computational costs of training runs at the frontier have reached hundreds of millions to billions of dollars per model, creating capital barriers that exclude nearly all potential entrants. The training data required is owned or accessed primarily by the largest incumbents. The specialized engineering talent is concentrated in a small number of organizations that can afford to compensate it at the levels the market demands.
Platform dynamics reinforce the concentration. The value of an AI service increases with the number of users (more data for training, more feedback for refinement, more scale for cost amortization). The data network effect — unique to AI platforms — converts usage into quality improvement that compounds rather than erodes. Early leaders have structural advantages that compound over time, producing the winner-take-most dynamics characteristic of limited-access emergence.
The informal norms of the AI industry further concentrate resources. The prestige hierarchy places frontier model builders at the apex. Elite training programs — PhD work at a handful of universities, postdoctoral positions at the frontier labs — produce credentials that open doors unavailable to anyone outside the network. The cultural narratives about AI development emphasize the unique capabilities of the frontier actors, reinforcing the perception that serious AI work can only happen within the dominant coalition.
Inclusive institutional design in the AI era requires deliberate countermeasures against these concentrating forces. Antitrust enforcement calibrated to AI market dynamics. Open-source requirements preventing complete privatization of AI capability. Interoperability standards reducing switching costs and preventing platform lock-in. Public investment in AI research maintaining competitive alternatives to corporate development. Educational institutions distributing AI literacy broadly rather than concentrating it in elite programs.
Origin
The framework emerged from North, Wallis, and Weingast's work on the transition from pre-modern to modern political and economic orders. They argued that the transition from limited to open access orders in a small number of Western societies since 1800 was the most consequential institutional development of the modern era — and that most societies remained limited access orders, governed by dominant coalitions whose management of violence required extracting rents through restricted access to economic and political opportunity.
The application to AI extends the framework from nation-state scale to industry-structure scale, drawing on parallel analyses of platform monopolies by Lina Khan, Tim Wu, and other new-school antitrust scholars, as well as on the network-economics framework of Carl Shapiro and Hal Varian.
Key Ideas
Rent-seeking produces stability. Limited access orders are stable because the dominant coalition has incentives to maintain the restriction — the rents fund cooperation within the coalition.
Open access is historically rare. Most societies throughout history have been limited access orders. The modern open-access arrangements are historical anomalies requiring deliberate institutional design.
Natural barriers compound institutional ones. AI's computational, data, and talent requirements create barriers that reinforce whatever institutional concentration policy permits.
Platform dynamics accelerate concentration. Network effects and data network effects convert incumbency advantages into structural lock-in that intensifies over time.
Open access requires active maintenance. Without deliberate countermeasures — antitrust enforcement, interoperability standards, open-source requirements, public research investment — market dynamics produce limited access by default.
Debates & Critiques
Debates center on whether AI concentration is truly a limited-access phenomenon or merely the temporary dominance of first movers in a rapidly evolving market. Optimists argue that open-source models (Meta's Llama, Mistral, various Chinese models) and continuing innovation will erode frontier-lab advantages. Pessimists argue that the capital requirements, compounding data advantages, and platform dynamics are producing permanent structural concentration that will require active regulatory intervention to address.
Appears in the Orange Pill Cycle
Further reading
- North, Wallis, and Weingast, Violence and Social Orders (Cambridge University Press, 2009)
- Lina Khan, 'Amazon's Antitrust Paradox' (Yale Law Journal, 2017)
- Carl Shapiro and Hal Varian, Information Rules (Harvard Business School Press, 1999)
- Tim Wu, The Curse of Bigness (Columbia Global Reports, 2018)