Regulating AI could change who wins AI race

A new economic study argues that existing regulatory thinking is poorly suited to the emerging structure of the AI industry, where powerful foundation model developers and application firms operate within a tightly interconnected supply chain. The research highlights how policy decisions affecting competition, computing resources, and pricing dynamics can shape not only corporate profits but also consumer welfare.

The study titled "The Economics of AI Supply Chain Regulation," published on arXiv, examines how regulatory interventions influence the rapidly expanding ecosystem of AI foundation models and downstream application developers. Using a game-theoretic economic model, the researchers analyze how competition policies and compute subsidies alter incentives across the AI supply chain and determine whether such policies ultimately benefit consumers, technology providers, and firms building AI-driven products.

The research discusses the concept of the AI supply chain, an economic system in which a small number of companies build large foundation models while thousands of downstream firms adapt them for specific applications using proprietary data. This layered market structure creates new forms of collaboration, competition, and regulatory risk.

Rise of the AI supply chain economy

AI development increasingly revolves around foundation models, massive neural networks trained on enormous datasets that can perform a wide range of tasks, including natural language processing and logical reasoning. These models form the technological backbone of modern AI products but are extremely expensive to develop and maintain. Training a model on the scale of GPT-4 can cost more than $100 million, making it impractical for most organizations to build their own.

Because of these high costs, the AI industry has evolved into a vertically structured supply chain. In this ecosystem, large technology companies build and maintain foundation models, while downstream firms adapt them for specialized applications such as legal research tools, medical assistants, or financial analytics systems.

The key mechanism that enables this structure is fine-tuning, a process through which downstream firms retrain a foundation model using domain-specific data. This step allows companies to customize general-purpose AI systems to meet the requirements of specific industries or tasks. While pretraining the model requires enormous computational resources, fine-tuning typically requires less computing power, making it accessible to smaller firms.

However, the process still involves significant costs. Downstream firms must preprocess large datasets to remove noise and inconsistencies before they can be used for training. At the same time, foundation model providers must supply the computational infrastructure needed for training and inference. As a result, both sides incur expenses that shape their strategic behavior in the market.

Providers charge two main types of fees. Firms pay a fine-tuning fee based on the amount of data used to retrain the model and an inference fee each time the model processes a query or generates output. These pricing structures create a two-sided revenue stream for model providers while tying the profitability of downstream firms directly to the infrastructure they rely on.

The researchers describe this structure as a unique form of co-creation. Foundation model providers supply the base models and computing infrastructure, while downstream companies contribute proprietary datasets and industry expertise. Together, they produce specialized AI services that are sold to consumers.

However, this collaborative structure also complicates regulatory oversight. Traditional competition policies were designed for markets where companies operate largely independently. In the AI supply chain, however, decisions made by upstream model providers can directly influence downstream competition, pricing strategies, and product quality.

Competition policies and consumer welfare

Regulators often attempt to increase competition by improving price transparency or by requiring firms to disclose accurate information about product performance.

The researchers identify two main categories of policy intervention: price competition policies and quality competition policies.

Policies promoting price competition focus on improving price transparency and making it easier for consumers to compare products. Examples include regulations requiring companies to display full pricing information upfront or banning hidden fees. In theory, such policies should lower prices and increase consumer welfare.

However, the study finds that the outcome is more complex in AI supply chains. When competition intensifies and firms reduce prices, their incentive to invest in model improvement can decline. Because improving AI products requires expensive data preprocessing and fine-tuning, lower profit margins may discourage companies from investing in quality improvements.

Under certain conditions, this dynamic can actually reduce consumer welfare. If compute costs and data preparation costs are relatively low, intensified price competition may lead firms to reduce investment in training data, lowering the quality of AI products available to consumers.

On the other hand, policies promoting quality competition, such as rules requiring accurate product claims or preventing companies from hiding negative reviews, consistently improve consumer outcomes. These policies encourage firms to invest more heavily in model improvements, increasing the quality of AI services while maintaining competitive pricing pressure.

The study asserts that quality-focused competition policies represent the most reliable tool for improving consumer welfare in AI markets. However, such policies come with trade-offs. While consumers benefit from higher quality products, the research shows that quality competition can reduce profits for downstream firms. Increased pressure to improve model performance forces companies to invest heavily in data processing and training infrastructure, eroding profit margins.

Compute subsidies and the economics of AI infrastructure

The study examines a second regulatory tool gaining popularity among governments: compute subsidies. Several countries and regional governments have begun subsidizing access to computing infrastructure used for AI training. For example, Chinese municipal governments have launched programs that subsidize a portion of the computing costs associated with training large AI models.

The researchers analyze how such subsidies affect the AI supply chain and find that they generally increase consumer welfare. By lowering the cost of computing resources, subsidies reduce the price that foundation model providers charge for fine-tuning. Lower prices encourage downstream firms to use more training data, improving the quality of AI products delivered to consumers.

However, the benefits depend heavily on cost conditions. If compute costs and data preprocessing costs remain high, the subsidies may become inefficient. In such cases, the public money spent on subsidies may exceed the additional consumer benefits generated by improved AI services.

The study also highlights the importance of subsidy design. Excessively large subsidies can lead firms to overinvest in fine-tuning, increasing public spending without producing proportional gains in consumer welfare. Policymakers therefore need to calibrate subsidy rates carefully.

When implemented under the right conditions, compute subsidies can produce a rare economic outcome in technology markets: a "win-win-win" scenario. In this situation, consumer welfare increases while both foundation model providers and downstream firms see higher profits.

Falling compute costs will reshape AI regulation

Advances in GPU technology are steadily reducing the cost of training and deploying AI models, a trend expected to continue for years. As computing becomes cheaper, the effectiveness of different regulatory policies will change. The researchers predict that price-competition policies, which may be effective today, could become less useful in the future. Meanwhile, compute subsidies may become increasingly beneficial as computing costs fall.

Lower compute costs also have uneven effects across the AI supply chain. Foundation model providers benefit because their infrastructure becomes cheaper to operate, allowing them to expand services and increase inference demand.

Downstream firms, however, may not always benefit. Cheaper computing intensifies competition in product quality, pushing firms to invest more heavily in training data and model improvements. In some cases, these additional investments can reduce profits even as the overall AI ecosystem grows.

For consumers, the outlook remains positive. Declining compute costs consistently increase consumer surplus by enabling higher-quality AI products at lower prices.