Regime change often emerges from costly, decentralized workarounds that, when scaled, outcompete failing institutions on maintenance grounds; paradoxically, higher adoption frictions slow transitions but make who adopts more revealing about the next regime.

Main Finding

The paper develops a formal theory of economic transition in which regimes persist so long as the inherited architecture remains the least‑maintenance viable option under a binding resource envelope. Regime failure is driven by a stabiliser→transmitter reversal (an anchor that once absorbed stress begins to propagate it). Decentralised agents respond by adopting costly, partial workarounds; if adoption and complementarities raise a workaround’s systemic capacity enough that it can be maintained within resources and requires less maintenance than the incumbent, the workaround becomes the new regime. The model yields four core results: inefficient regime persistence, endogenous transition via workaround scaling, under‑adoption relative to the social optimum, and a delay–information tradeoff (high adoption frictions delay transition but make observed adoption more informative about the future regime).

Key Points

• Conceptual definitions

  • Regime = stabilising architecture (institutions, standards, technologies, rules) that keeps a complex system viable under stress.
  • Anchor / stabiliser = component that absorbs systemic stress.
  • Stabiliser–transmitter reversal = when the anchor’s marginal role flips from damping volatility to transmitting/amplifying stress.
  • Workaround = decentralised, partial alternative domain agents adopt to reduce exposure to the failing anchor; costly at first but capable of building network effects and capacity.
  • Manager = system steward (state, central bank, platform, standard‑setter, etc.) constrained by a finite resource envelope R.

• Selection rule and viability

  • Manager chooses the least‑maintenance architecture among those that (a) meet a minimum stabilising capacity and (b) are maintainable within resource envelope R.
  • Viability requires µj(s, n_j) ≤ R (maintenance feasible) and Ωj(s, n_j) ≥ Ω̄ (capacity threshold).

• Adoption dynamics

  • Individual adoption payoff: U_i,d = θ_i Z_d(s) + q_d + η_d n_d − κ_d, where:
    • θ_i captures heterogeneity in exposure/adaptability,
    • Z_d(s) = bypass value of reducing dependence on incumbent under stress s,
    • q_d = latent domain quality,
    • η_d = network/complementarity effect,
    • κ_d = adoption friction (costs, legal uncertainty, organizational switching, etc.).
  • Workaround adoption evolves endogenously; adoption raises nd which feeds back into µ_d and Ω_d.

• Main mechanisms and implications

  • Persistence without optimality: regimes can persist not because of irrationality or capture but because, given R, no alternative has yet become less‑maintenance viable.
  • Transition is an emergent interaction: decentralised adoption + complementary investments build capacity; transition occurs when the manager’s least‑maintenance choice switches.
  • Under‑adoption: Private incentives under‑invest in adoption relative to social value because individual adopters do not internalise positive externalities that make a workaround systemically viable.
  • Delay–information tradeoff: Higher κ_d (friction) slows adoption and delays transition, but observed adoption under high κ_d is a stronger signal of latent domain quality (costly adoption screens out low‑quality domains).

• Formal structure and technical conditions

  • Results are stated under explicit monotonicity/single‑crossing and survival‑likelihood‑ratio conditions (to make first‑crossing and Bayesian inference claims precise).
  • The model defines a breakdown region (no architecture is viable for given R) and distinguishes it from smooth regime transition.

Data & Methods

• Type of paper: theoretical / formal model (no empirical dataset used).
• Time: discrete framework with a continuum of agents i ∈ [0,1].
• Objects and state variables:
  • Incumbent architecture A and set D of workaround domains.
  • Systemic stress s ≥ 0.
  • Maintenance µ_j(s, n_j) and stabilising capacity Ω_j(s, n_j) for architectures j ∈ {A} ∪ D.
  • Resource envelope R and minimum capacity Ω̄.
• Manager selection rule: choose j ∈ V(s, n) minimizing maintenance µ_j, where V(s, n) = {j : µ_j ≤ R and Ω_j ≥ Ω̄}.
• Adoption model: payoff as above; adoption fixed‑point (one domain) given by equation (4): n = 1 − F((κ − q − η n)/Z(s)), where F is distribution of θ.
• Analytical methods:
  • Comparative statics of R, κ, η, q, Z(s), and s to derive persistence, transition, and breakdown regions.
  • First‑crossing theorem for transition; single‑crossing for persistence after crossing.
  • Bayesian inference (survival‑likelihood ratio) to formalize the delay–information tradeoff: higher adoption thresholds produce stronger posterior evidence about domain quality when adoption occurs.
• Proofs and technical details: provided in appendices (existence of equilibrium, comparative statics, formal theorems, conditions).

Implications for AI Economics

• Framing AI systems as stabilising architectures

  • Dominant AI infrastructures (large cloud compute providers, centralized data pipelines, widely used model stacks, commonly adopted tools/standards) function as stabilising anchors: they lower coordination costs, enable scale, and absorb systemic stress in digital economies.
  • These same anchors can undergo stabiliser→transmitter reversal: concentration of compute/data can propagate risk (cyberattacks, model behavior cascades, geopolitical choke points, energy bottlenecks, surveillance and regulatory shock transmission).

• What are AI workarounds?

  • Workarounds correspond to decentralized or alternative AI architectures: edge AI, federated learning, smaller open models, diversified chip and data supply chains, on‑prem or sovereign compute stacks, different governance/data‑sharing arrangements, robust verification and safety toolchains.
  • Initially costly and partial (lower performance, limited interoperability, higher integration costs), these can gain capacity through network effects, standards, and complementary investments (tooling, datasets, skills).

• Policy and market implications

  • Monitoring maintenance costs of incumbent AI architecture: public balance‑sheet equivalents such as subsidies/support to major providers, extraordinary operational assistance, energy allocations, surveillance/legal tolerances—rising maintenance can signal stabiliser stress.
  • Early‑warning indicators: clustering of workaround adoption in domains relevant to the failing anchor (e.g., spikes in edge deployments, open model forks, investment in alternative chip suppliers, enterprise moves to on‑prem/federated solutions) should be interpreted as revealing the domain in which a transition is forming.
  • Delay–information tradeoff relevant to AI policy design:
    • High adoption frictions (technical switching costs, regulatory barriers, certification burdens) slow diffusion of alternative AI architectures — delaying any beneficial decentralization.
    • However, observed adoption under high friction is a stronger indicator of a viable long‑term alternative (helps policymakers allocate scarce support more efficiently).
  • Subsidy and intervention tradeoffs:
    • There is a role for targeted subsidies to correct under‑adoption when social returns (systemic resilience, competition, reduced systemic risk) exceed private returns.
    • But subsidies risk “contaminating the signal”: artificially propping weak alternatives can produce false positives about viability; careful design (time‑limited, contingent on verifiable metrics, staged matching grants encouraging private co‑investment) is needed.
  • Resilience and diversification: Authorities should consider investing in alternative compute/data infrastructure and in reducing single points of failure (diversify supply chains, certify interoperable stacks) because alternative domains that achieve sufficient scale can become lower‑maintenance systemic architectures.

• Measurable indicators for AI early‑warning systems

  • Maintenance stress signals: rising public or private spending to prop incumbents; emergency regulatory exemptions for major providers; energy allocation anomalies; frequency and severity of outages/cascades.
  • Concentration metrics: market shares in cloud/TPU/GPU markets, model hosting concentration, data‑pipeline centrality.
  • Workaround clustering: rates of enterprise adoption of federated/edge AI, open‑model releases and uptake, investment flows into alternative chip fabs or sovereign compute projects, growth rates of developer ecosystems around alternative stacks.
  • Adoption friction proxies: switching costs (integration time, retraining cost), regulatory compliance burden, availability of compatible tooling, legal uncertainty indices.
  • Informational signal assessment: measure adoption occurrence conditional on friction level — adoption in high‑friction contexts should be given greater weight as evidence of latent domain quality.

• Research directions in AI economics suggested by the paper

  • Empirically test the delay–information tradeoff in AI: collect panel data on adoption of alternative AI approaches across firms/countries with varying switching costs and measure posterior predictive power for systemic transitions (e.g., market structure changes, regulation shifts).
  • Calibrate/extend the theoretical model to specific AI architectures: specify µ_j and Ω_j functions for cloud‑centric vs decentralized AI, include endogenous R (government resource mobilization), and strategic behaviour from large incumbents.
  • Welfare and policy design: quantify social externalities from workaround formation (positive resilience externalities vs negative fragmentation/externalities) to derive optimal subsidy/taxation schemes that preserve signal quality.
  • Dynamics of standards and interoperability: model multiple coexisting workaround domains and their compatibility to study tipping, coexistence, or fragmentation outcomes in AI ecosystems.

Summary takeaway for AI economists: treat dominant AI infrastructures as stabilising anchors whose failure may not be abrupt collapse but a reversal in role; monitor maintenance costs and decentralised workaround clustering; be mindful that friction both delays desirable transitions and helps reveal which alternatives are truly viable — policy should balance accelerating socially valuable adoption with preserving the informational value of costly adoption signals.