Policies that spur quality competition in AI supply chains reliably boost consumer surplus, while price-competition rules and compute subsidies help only in particular cost environments; unexpectedly, price competition or compute subsidies can produce win-win-win outcomes for providers, downstream firms and consumers under the model's assumptions.

Main Finding

A game-theoretic model of an AI supply chain (one foundation-model provider supplying compute/fine-tuning services to two competing downstream firms) shows that: - Policies promoting quality competition always raise consumer surplus. - Policies promoting price competition increase consumer surplus only when compute or data-preprocessing costs are high. - Compute subsidies increase consumer surplus only when compute or preprocessing costs are low. - Price-competition policies and compute subsidies can be complementary, and under those policies both the provider and downstream firms can also gain (a win-win-win). By contrast, pro-quality policies raise the provider’s profits but reduce downstream firms’ profits. - As compute costs fall, pro-price-competitive policies lose effectiveness while compute subsidies may become effective.

Key Points

• Context: Foundation models create vertical AI supply chains where providers supply base models, compute, and fine-tuning/inference services; downstream firms co-create domain-specific products using proprietary data and provider infrastructure.
• Market structure: One upstream provider and two symmetric competing downstream firms (stylized vertical supply chain).
• Policy interventions analyzed:
  • Pro-price-competitive policies (e.g., measures that intensify downstream price competition or limit downstream market power).
  • Pro-quality-competitive policies (e.g., measures that encourage competition on quality, disclosure/standards that boost quality-driven differentiation).
  • Compute subsidies (direct subsidies to reduce cost of compute/fine-tuning for downstream firms).
• Complementarity and regime dependence:
  • Price competition and compute subsidies are not universally welfare-improving; their effectiveness depends on underlying cost structures (compute and data-preprocessing costs).
  • When those costs are high, encouraging price competition reduces markups and benefits consumers; when costs are low, reducing costs further (via subsidies) helps lower prices/raise quality.
• Distributional effects:
  • Pro-price policies or compute subsidies can create a Pareto-improving outcome (consumers, provider, and downstream firms all better off) in some parameter regions.
  • Pro-quality policies unambiguously benefit consumers and the upstream provider but harm downstream firms’ profits (they face tougher quality competition or greater investment burdens).
• Dynamic implication: Declining compute costs (a realistic trend) alter which policy levers are effective over time.

Data & Methods

• Method: Analytical game-theoretic model (theory paper), no empirical dataset. Key elements likely include:
  • Vertical model with one upstream foundation-model provider and two downstream firms.
  • Downstream firms purchase compute/fine-tuning services and use proprietary data to invest in product quality; they then compete (price/quality) in the downstream market.
  • Provider sets fees for compute/inference/fine-tuning; downstream firms choose quality investments and prices; equilibrium derived via backward induction.
  • Consumer surplus and profits (provider + downstream firms) computed under equilibrium outcomes.
  • Comparative statics analyze how equilibrium and welfare respond to policy interventions and changes in cost parameters (compute and data-preprocessing costs).
• Assumptions and simplifications:
  • Symmetric downstream firms, one upstream provider, static interaction (no multi-period investment/dynamic competition).
  • Quality production depends on joint upstream compute and downstream data/preprocessing; costs captured through parameter(s) for compute and preprocessing.
  • Policies modeled as exogenous changes to market structure (e.g., intensifying price competition), subsidies to compute, or measures affecting quality competition.
• Robustness: Results are analytical and hinge on parameter regimes; qualitative insights depend on the model’s structural assumptions.

Implications for AI Economics

• Policy design must be cost-aware: Regulators should assess the prevailing cost structure (compute and data-preprocessing costs) when choosing interventions. A one-size-fits-all approach (e.g., always promoting price competition or always subsidizing compute) can be suboptimal.
• Complementary mixes work best: Combining pro-price measures and compute subsidies can be mutually reinforcing in some regimes; policymakers can target the combination that fits the current technological cost environment.
• Promote quality competition to protect consumers: Policies that raise quality competition reliably increase consumer surplus, but they reallocate surplus toward upstream providers and away from downstream firms—regulators should anticipate and potentially mitigate these distributional effects (e.g., targeted support for downstream SMEs).
• Antitrust and subsidy instruments both matter: Traditional antitrust tools that intensify downstream price rivalry and industrial-policy tools (compute subsidies, credits) have distinct and conditional effects; both should be on the policy toolkit and calibrated to industry parameters.
• Dynamic regulatory strategy: As compute costs decline over time, the optimal policy mix should adapt — pro-price interventions may be less needed, while targeted compute subsidies may gain or lose traction depending on residual preprocessing costs and data frictions.
• Research and monitoring priorities: Empirical measurement of compute and data-preprocessing cost levels, margins across supply-chain layers, and the distribution of bargaining power will be crucial for implementing the model’s policy recommendations in practice.
• Limitations: Theoretical, static, and simplified (one provider, two symmetric downstream firms). Real-world supply chains include multiple providers, asymmetric firms, dynamic investment and entry, regulatory compliance costs, and data-sharing arrangements that can modify the qualitative results. Further empirical and richer-theory work is warranted to operationalize policy guidance.