Skip to main content
Back

Tsinghua Researchers Prove AI Regulation Backfires When Compute Costs Are Low

A game-theoretic model reveals when compute subsidies, price-competition rules, and quality-competition rules actually help consumers in AI markets, and when they backfire.

Paper, Full Access
  • Game-theoretic paper models one foundation model provider serving two competing downstream firms in the AI supply chain. Read paper
  • Pro-price-competition rules help consumers only when compute and data preprocessing costs are high.
  • Compute subsidies help consumers only when those same costs are low, making the two policies complements.
  • Pro-quality-competition rules always raise consumer surplus but shrink downstream firm profits.
  • Pro-price-competition and compute subsidies can deliver a win-win-win across consumers, providers, and downstream firms.
  • As GPU costs fall, price-competition rules may lose effect while compute subsidies gain it.

Regulators are scrambling to figure out how to police markets built on foundation models, where a handful of upstream providers like OpenAI sell fine-tuning and inference to countless downstream firms building domain-specific apps. A new economics paper takes a rigorous crack at the question, modeling the AI supply chain as a Stackelberg-style game and asking a blunt question: which regulatory levers actually raise consumer welfare, and which ones quietly hurt it?

The paper from researchers at Tsinghua and UT Dallas concludes that the answer depends heavily on the cost structure of compute and data preparation, and that the intuitive policies can be exactly the wrong ones under common conditions.

Why the usual regulatory playbook breaks

Traditional antitrust intuition says lowering prices in a downstream market helps consumers. But in AI, downstream firms and the model provider are locked in a co-creation loop: downstream firms upload proprietary data to the provider's infrastructure for fine-tuning, incurring data preprocessing costs, while the provider bears compute costs and charges a fine-tuning fee plus an inference fee based on usage. Squeeze downstream prices too hard and firms invest less data into fine-tuning, quality drops, and consumers can end up worse off.

The authors formalize this with a three-player setup: a foundation model provider and two competing downstream firms, where each firm's product quality equals a base foundation model quality plus the volume of data it uses for fine-tuning. Compute costs scale quadratically with data volume because large-scale fine-tuning runs on distributed GPU clusters where communication overhead grows non-linearly as more nodes are added. Consumer demand for each firm depends linearly on both its own price and quality relative to its rival.