AI could first concentrate markets then democratize them: the authors’ calibrated general-equilibrium model shows industry concentration rises at low AI adoption as large incumbents capture early gains, but falls once adoption becomes widespread (peak near ~15% adoption). Average markups increase when AI supply expands, while demand-driven adoption generates a non-monotonic markup response; a small (~3%) revenue subsidy to AI adopters is welfare-optimal in the U.S. calibration.

Main Finding

The paper develops a general-equilibrium model of heterogeneous firms and an endogenous AI sector to show that AI’s effect on market competition is non-monotonic. As AI diffuses, industry concentration first rises (incumbent large adopters expand) and then falls (extensive entry/adoption by smaller firms erodes incumbents). The model yields a hump-shaped relation between AI diffusion and concentration (turning point ≈ 15% adoption). Calibrated to U.S. data (current adoption ≈ 18%), the model predicts concentration will decline as AI continues to spread. The effect of AI on average markups depends on whether diffusion is supply- or demand-driven; and a small (≈3%) revenue subsidy for AI adopters maximizes welfare under the authors’ calibration.

Key Points

• Mechanism: Two opposing margins drive the non-monotonic outcome.
  • AI-usage (intensive) channel: lower AI prices and incumbent adoption let large users expand market share → increases concentration.
  • AI-adoption (extensive) channel: cheaper/more available AI induces smaller firms to adopt → reduces concentration.
• Fixed costs of AI adoption create scale economies that favor larger firms early in diffusion.
• Data matters: intermediate firms generate sellable data as a byproduct; larger firms produce more data, which can advantage them. The hump-shaped result remains even when firms can combine AI with their own data.
• Markups:
  • If diffusion is driven by AI supply improvements (cheaper AI), average markups tend to rise monotonically (lower marginal costs for adopters).
  • If diffusion is driven by demand-side factors (lower fixed adoption costs), average markups can be non-monotonic (rise at first, then fall when many small firms adopt).
• Calibration and quantitative results:
  • Hump turning point at ~15% adoption.
  • Current U.S. adoption proxy ≈ 18% → model predicts falling concentration going forward.
  • Welfare-maximizing revenue subsidy for AI adopters ≈ 3% (tradeoff between productivity gains and increased markup distortions).

Data & Methods

• Model:
  • General-equilibrium, continuum of intermediate-goods firms with idiosyncratic productivity shocks and variable markups (Kimball preference aggregator).
  • Firms choose AI adoption paying a fixed cost (scaled by productivity). Adopters use unskilled labor + an “AI bundle” (algorithm produced in AI sector + skilled labor); non-adopters use unskilled labor only.
  • AI producers form a differentiated-product sector: algorithms are produced using purchased data and energy. Energy supply is fixed, so AI energy demand endogenously affects energy prices. AI-sector entry is costly and endogenous.
  • Data produced as a byproduct of goods production and sold competitively to AI producers. Also analyzed: cases where firms can combine AI with their own data.
• Calibration:
  • Parameters chosen to match U.S. firm-level stylized facts on AI adoption and usage; quantitative exercises vary supply (AI productivity, data/energy supply) and demand (adoption fixed cost) drivers.
• Empirical evidence:
  • Proxy for firm-level AI adoption: share of AI-related job postings from Lightcast (textual keyword search including GenAI terms developed with an LLM), aggregated to 5-digit NAICS industry.
  • Linked to Compustat firm data to compute industry concentration (top firm share, HHI) and firm markups (sales − COGS log difference).
  • Regressions replicating Babina et al. (2024) style: long-differences (2010–2018 vs 2019–2025). Findings:
    • 2010–2018: Positive and significant correlation between AI job-share growth and increases in concentration.
    • 2019–2025: Correlation much weaker / statistically insignificant.
    • Markups have remained positively correlated with AI usage throughout 2010–2025.
• Identification / data caveats:
  • AI adoption is proxied by job-posting keywords (not direct investment or usage measures).
  • Industry-level analysis may mask within-industry heterogeneity; post-2019 attenuation may reflect changing adoption patterns or data limitations.

Implications for AI Economics

• Competition dynamics are path-dependent: early-stage diffusion can amplify incumbency effects, but broader diffusion can democratize productivity gains and lower concentration.
• Policy timing matters:
  • Antitrust and regulatory concerns should account for non-linear effects over the diffusion curve—interventions aimed at curbing concentration may be premature if adoption is still in the early intensive-margin stage.
  • A modest, targeted revenue subsidy to firms adopting AI can raise welfare by balancing productivity gains and markup distortions (paper-calibrated optimum ≈ 3%).
• Interpreting markup trends:
  • Observed persistent positive correlation between AI usage and markups (2010–2025) is consistent with supply-driven diffusion (cheaper/more productive AI), which tends to increase markups in the model.
• Data- and energy-related constraints matter: changes in data availability or energy supply can shift AI prices and thus the balance between intensive and extensive margins.
• Empirical strategy recommendations:
  • Improve firm-level AI adoption measures (beyond job-posting proxies) and identify supply vs demand drivers to validate mechanism distinctions.
  • Track adoption rates relative to the model’s turning point to anticipate concentration dynamics.
• Research extensions / limitations:
  • Model assumes a fixed energy endowment and competitive data market—altering these assumptions (e.g., endogenous data markets with frictions, platform ownership of data) could change quantitative results.
  • The current paper is an “incomplete draft”; robustness to alternative preference/markup specifications, multi-sector general equilibrium spillovers, and labor market dynamics (reallocation of skilled vs unskilled labor) are promising extensions.

Summary takeaway: AI can both intensify and weaken market concentration depending on where the economy is along the diffusion path. Policymakers should monitor adoption rates and drivers (supply vs demand) because the optimal regulatory or subsidy response differs across stages.