The Economics of AI Foundation Models: Openness, Competition, and Governance
summary
The gist
The strategic choice of model “openness” has become a defining issue for the foundation model (FM) ecosystem While this choice is intensely debated…
In short
The paper analyzes how foundation model developers should choose their level of openness based on a 'data flywheel effect' (k). It shows three strategic regimes: Harvest (high openness/price), Defend (restricted openness for intermediate k), and Dominate (low price/high openness for very strong k). This reveals an 'openness trap' where transparency mandates can harm competition.
Key concepts
- Data Flywheel Effect (k)
- This measures how much a model's performance improves over time due to ecosystem feedback. A strong effect means better models lead to more users, which leads to better data, creating a virtuous cycle of improvement and market dominance.
- Openness Trap
- This is a policy paradox where regulations forcing transparency can backfire. If a developer is in an intermediate market strength, forced openness removes their strategic flexibility, causing them to reduce investment and lower overall consumer welfare.
- Harvest Strategy
- When the data flywheel effect is weak, the incumbent maximizes short-term profit by choosing maximum openness combined with a high license price. This strategy focuses on capturing immediate revenue before the market shifts or competitors catch up.
- Defend Strategy
- For intermediate flywheel strengths, incumbents strategically restrict openness. This move is designed to intentionally hinder potential entrants' ability to learn and improve their models, securing a future advantage through strategic opacity.
Terminology used across episodes
This episode discusses
- The Economics of AI Foundation Models: Openness, Competition, and Governance · Paper Radio
- The 2024 Foundation Model Transparency Index
- Human-AI Interactions and Societal Pitfalls · Paper Radio
- Retrieval-Augmented Generation for Large Language Models: A Survey
- On the Societal Impact of Open Foundation Models
- Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models
- Modeling the Economic Impacts of AI Openness Regulation
- The Model Openness Framework: Promoting Completeness and Openness for Reproducibility, Transparency, and Usability in Artificial Intelligence
- Generative AI and Organizational Structure in the Knowledge Economy
The paper
The Economics of AI Foundation Models: Openness, Competition, and Governance · Read on arXiv
School of Business, University of Connecticut · Faculty of Business, The Hong Kong Polytechnic University
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We're looking at the paper: "The Economics of AI Foundation Models".
Title and authors: Ruth: So we’re looking at how this paper develops that game-theoretic model to analyze how openness shapes competition in the AI value chain, and what they find about the incumbent's strategy. It seems they’ve found three distinct strategic regimes for an incumbent developer based on the strength of that data flywheel effect.
Warren: That sounds like a really useful framework for understanding real-world behavior because it moves away from just saying "open is good" or "closed is good." They’re defining when each strategy makes economic sense.
Iris: They break it down into three specific strategies: the Harvest strategy, the Defend strategy, and the Dominate strategy. Each one corresponds to a different strength of that data flywheel effect, let's call it k.
Ruth: When k is weak or very strong, like in those extremes, the incumbent prefers a higher level of openness because they can leverage that advantage effectively regardless of whether they are maximizing short-term gain or long-term market share.
Warren: But for that intermediate range of the data flywheel effect strength, the incumbent strategically restricts openness. That’s where they introduce what they call an “openness trap.”
Iris: The openness trap is a critical policy paradox because transparency mandates, which are often pushed by regulators, can actually backfire by removing firms’ strategic flexibility and lowering their investment.
Ruth: It means that if you try to force a developer into total openness when the data flywheel effect is intermediate, they might end up pivoting to a Harvest strategy instead of what they were aiming for.
The paper's summary: Ruth: To summarize, the paper constructs this game-theoretic framework where the incumbent chooses both a license price and an openness level in the first period, and how that interacts with the deployer’s ability to fine-tune in the next period.
Warren: The deployer’s main job is to take that general model and specialize it for end users through fine-tuning, and the cost of that fine-tuning changes depending on how open the model is.
Iris: Specifically, higher openness, which they call eta one, reduces the deployer’s fine-tuning cost because it makes the architecture more transparent and encourages knowledge sharing within the ecosystem.
Ruth: And that leads directly into their three canonical strategies based on k: when k is low, they harvest by maximizing openness and price for short-term profit; when k is intermediate, they defend by restricting openness to stop entrants from learning; and when k is very high, they dominate with a low price and high openness.
Warren: It really shows that the incumbent's rational choice isn't fixed; it’s dynamic based on how much advantage they actually have in terms of that data flywheel effect.
Iris: And this structure helps us see why vertical integration or simple subsidies might not work as simply as people assume, because their effectiveness is entirely conditional on how strong that data flywheel effect k is.
The paper's improvements: Ruth: The authors suggest some important improvements to the existing literature, mainly by formalizing the economics of strategic openness within a game-theoretic model rather than relying on simpler competition models.
Warren: They argue that existing models of platform competition or open-source software just don't capture the unique dynamics specific to AI value chains, so they needed this new formalization. It’s about filling a gap in the research.
Iris: They also point out that when policymakers mandate full transparency, their analysis shows a regulatory paradox—the openness trap—which implies that governance needs to be nuanced rather than one-size-fits-all.
Ruth: For incumbents, the improvement is recognizing that openness and pricing are tools to be co-optimized based on k; if the flywheel is weak, you maximize short-term revenue; if it’s strong, you accelerate lock-in with lower prices and greater openness.
Warren: And for policymakers, it’s a caution against blunt regulation. They suggest that effective governance might involve things like private model registration to provide oversight without stripping firms of their ability to compete strategically.
Iris: For downstream deployers, the paper offers a clearer test: they should evaluate whether vertical integration is actually efficiency-enhancing by checking if k is strong enough to ensure incumbent victory in the second period. If k is weak, they should avoid integration entirely because it could just lock them out of more efficient entrants.
Conclusion: Ruth: So wrapping up, the paper on The Economics of AI Foundation Models: Openness, Competition, and Governance shows that the strategic choice of openness is deeply tied to the data flywheel effect. It forces us to rethink how we analyze competition in this new value chain.
Warren: Essentially, it gives incumbents a playbook: if your advantage is weak or very strong, lean into openness; if it’s in the middle, play defensively by restricting learning opportunities for competitors.
Iris: And the main implication for regulators is that they have to be careful with mandates because forcing too much openness can actually hurt consumer welfare by causing firms to stop investing in fine-tuning.
Ruth: It really highlights that we can’t just look at the technology itself; we have to model how incentives—like licensing fees and access rules—shape behavior under different levels of competitive advantage.
Warren: It suggests that the "best" strategy isn't a single universal approach but depends entirely on where you land on that strength of the data flywheel effect, k.
Iris: That’s what it gets us thinking about how to structure governance in this space without inadvertently creating those traps for the developers themselves.
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- 2501.15307-Influence Functions: A Unified Differentiation-Based Framework
- The day in conversation for 2026-11-19