The authors of this report examined the economic and production attributes of pre-trained artificial intelligence foundation models to answer the following questions: Does the market for foundation models have the characteristics of a natural monopoly, and, if so, is regulation needed? The authors established empirical criteria and then applied these criteria to the status quo foundation model market and to four hypothetical scenarios set in 2027.

Evaluating Natural Monopoly Conditions in the AI Foundation Model Market

Jon Schmid, Tobias Sytsma, Anton Shenk

ResearchPublished Sep 12, 2024

Because of the wide variety of tasks they can be used to perform, foundation models — a class of artificial intelligence (AI) models trained on large and diverse datasets and capable of performing many tasks — have the potential to have a large effect in shaping the economic and social effects of AI. The authors of this report examined the economic and production attributes of pre-trained foundation models to answer the following questions: Does the market for foundation models have the characteristics of a natural monopoly, and, if so, is regulation of that market needed?

A natural monopoly refers to a market in which the total cost of serving the full range of demand is lower for a single firm than for multiple firms. Unlike a conventional monopoly, in a natural monopoly, competition and traditional antitrust policy cannot be assumed to alleviate the problems associated with concentrated market power. The authors established empirical criteria for classifying a market as a natural monopoly and applied them to the status quo foundation model market and to four hypothetical scenarios set in 2027 to understand possible future market dynamics.

Application of the natural monopoly criteria to the status quo AI foundation language model market (as of January 2024) indicates that the current case for a natural monopoly is relatively strong. This conclusion is based on the observations that the current generation of foundation models is reasonably homogeneous, economies of scale are high, costs are largely sunk, and network effects and economies of scope are present.

Key Findings

To consider how market structure may change in the future, the authors varied two technology variables: the scaling hypothesis and the cost of compute technology

  • In scenarios in which the scaling hypothesis is assumed to hold — i.e., the relationship between performance and model size/compute expenditure persists — the case for a natural monopoly relative to the status quo market is stronger. This is largely due to the increase in economies of scale and the risk of sunk costs associated with pre-training very large foundation models.
  • In contrast, in scenarios in which the scaling hypothesis breaks down, the argument for a natural monopoly decreases relative to the status quo market.

While the authors find that the status quo market has characteristics of a natural monopoly, they believe that the rationale for natural monopoly regulation is weak because of the low observed social cost associated with the current market structure

  • Potential social costs of concentration in the market for foundation models include pricing above marginal cost, low product quality, costs associated with market concentration in the market for compute, the environmental impact of large training runs, and systemic risk.
  • If evidence of significant social costs emerges, the question of regulation should be reconsidered.

Topics

Document Details

Citation

RAND Style Manual
Schmid, Jon, Tobias Sytsma, and Anton Shenk, Evaluating Natural Monopoly Conditions in the AI Foundation Model Market, RAND Corporation, RR-A3415-1, 2024. As of September 23, 2024: https://www.rand.org/pubs/research_reports/RRA3415-1.html
Chicago Manual of Style
Schmid, Jon, Tobias Sytsma, and Anton Shenk, Evaluating Natural Monopoly Conditions in the AI Foundation Model Market. Santa Monica, CA: RAND Corporation, 2024. https://www.rand.org/pubs/research_reports/RRA3415-1.html.
BibTeX RIS

Research conducted by

Funding for this work was provided by gifts from RAND supporters. The research was conducted by the Technology and Security Policy Center within RAND Global and Emerging Risks.

This publication is part of the RAND research report series. Research reports present research findings and objective analysis that address the challenges facing the public and private sectors. All RAND research reports undergo rigorous peer review to ensure high standards for research quality and objectivity.

This document and trademark(s) contained herein are protected by law. This representation of RAND intellectual property is provided for noncommercial use only. Unauthorized posting of this publication online is prohibited; linking directly to this product page is encouraged. Permission is required from RAND to reproduce, or reuse in another form, any of its research documents for commercial purposes. For information on reprint and reuse permissions, please visit www.rand.org/pubs/permissions.

RAND is a nonprofit institution that helps improve policy and decisionmaking through research and analysis. RAND's publications do not necessarily reflect the opinions of its research clients and sponsors.