Frontier AI magnifies scientific productivity but cements winners: AI-enabled labs capture outsized credit and resources through self-reinforcing dynamics; giving baseline AI to all labs narrows that gap with only a limited loss in aggregate discovery, making access policy a governance lever over future scientific power.

Main Finding

Frontier-AI access increases total scientific discovery but concentrates scientific credit and resources in AI-enabled laboratories through a self-reinforcing feedback loop. Equalising interventions that give baseline AI tools to all labs can partially close the gap at a bounded loss in aggregate output; a constrained-welfare analysis identifies when such interventions are optimal. Concentration is strongest under winner-takes-most credit allocation and, with complete information, the weakest labs may rationally withdraw from frontier races. Framing frontier AI as scientific infrastructure shows that governance of AI deployment is a governance of future scientific power.

Key Points

• Model setup: laboratories compete for scientific credit in a proportional (Tullock) contest; a subset of labs gain access to frontier AI that increases their productivity.
• Aggregate effect: AI access raises overall discovery/output.
• Distributional effect: AI access concentrates credit and resources in enabled labs via positive feedback (more credit → more resources → higher chance of future access/wins).
• Policy trade-off: Providing baseline AI tools to all labs narrows inequality but can reduce aggregate output; the loss is bounded and may be socially acceptable depending on intervention cost.
• Welfare analysis: Incorporates social cost of interventions to identify constrained-optimal policies (when and how much to equalize).
• Winner-takes-most variant: When credit allocation is highly skewed, concentration intensifies and weaker labs may exit competition under complete information, further reducing diversity.
• Conceptual shift: Treat frontier AI as scientific infrastructure — access decisions shape long-run distribution of scientific capability and power.

Data & Methods

• Approach: Theoretical, game-theoretic model using a Tullock contest framework (probability of earning credit is proportional to effort/ability share).
• Agents: Research laboratories competing for priority/credit; heterogeneous labs in productivity/cost.
• Treatment: A subset obtains AI tools that raise productivity; alternative model variant assumes winner-takes-most credit allocation.
• Analysis techniques:
  • Equilibrium characterization of effort, credit shares, and resource allocation.
  • Comparative statics to show how AI access changes aggregate output and concentration.
  • Policy analysis modeling interventions that provide baseline AI to all labs, with an explicit social cost parameter.
  • Welfare optimization under feasibility constraints to derive constrained-optimal policy.
  • Examination of complete-information outcomes including potential withdrawal of weaker labs.
• Data: No empirical data — results are theoretical/analytical and qualitative; robustness checks include varying contest success functions and credit allocation rules.

Implications for AI Economics

• Access is an input: Frontier-AI should be considered core scientific infrastructure; unequal access is a central market failure with long-term distributional consequences.
• Policy levers matter: Public or philanthropic provision of baseline AI capability can reduce concentration and preserve broader participation, but policymakers must weigh the aggregate-output trade-offs and intervention costs.
• Funding and credit design: How scientific credit or prizes are allocated (proportional vs winner-takes-all) materially affects concentration — funders can mitigate concentration by adjusting reward structures.
• Institutional strategy: Labs and countries lacking access face not only short-run productivity losses but long-run erosion of scientific capacity (possible exit), arguing for capacity-building and international cooperation.
• Competition and systemic risk: Extreme concentration reduces diversity of approaches and increases systemic reliance on a small set of AI-enabled institutions — relevant for innovation resilience and epistemic robustness.
• Further empirical work: Theoretical predictions call for empirical assessment of AI access, credit flows, entry/exit behavior of labs, and calibration of welfare trade-offs to guide concrete policy.