Regulating AI supply chains — foundation model providers selling to downstream application firms — seems like a single optimization problem: maximize consumer surplus. Choose the right policy lever and pull.
The paper (arXiv:2603.12630, March 2026) models this supply chain game-theoretically and finds that the right lever depends on the cost regime, and two levers that seem redundant are actually complementary. Pro-price-competitive policies (preventing monopoly pricing) boost consumer surplus only when compute or data preprocessing costs are high. Compute subsidies (reducing costs directly) work only when costs are already low. Each policy is effective precisely where the other fails.
The mechanism is the interaction between cost structure and competitive dynamics. When costs are high, the bottleneck is market power — firms can charge monopoly prices because alternatives are expensive. Price competition breaks the monopoly. When costs are low, the bottleneck is investment incentive — firms won't improve quality without subsidy because margins are thin. The subsidy funds the improvement. Applying the wrong lever to the wrong regime is neutral at best and counterproductive at worst.
The structural lesson: complementary policies are not redundant policies applied simultaneously — they are policies that cover each other's dead zones. The cost landscape has two regimes, and each regime responds to a different intervention. A single policy that works everywhere does not exist because the mechanism that creates the surplus is different in each regime. The policymaker's job is not to find the best lever but to read the regime correctly.