Market Power in Artificial Intelligence (Gans, 2024/2026)

Summary

"Market Power in Artificial Intelligence," by Joshua S. Gans (Rotman School of Management, University of Toronto; NBER), circulated as NBER Working Paper 32270 in March 2024 and later published in the Annual Review of Economics (Vol. 18, 2026, first posted online 6 May 2026; DOI 10.1146/annurev-economics-051624-061832). Gans is a leading industrial-organization economist of AI — co-author, with Ajay Agrawal and Avi Goldfarb, of Prediction Machines and Power and Prediction, the standard "AI as cheap prediction" economic framing this survey builds on. This paper is a genuinely new anchor for the wiki's AI-rents research cluster: the existing pages (Korinek, Stiglitz, and collaborators) focus on redistribution and tax design given AI-driven rents; this is the competition-policy question underneath that literature — do AI markets naturally concentrate, and why?

The Framework: Three Distinct Markets

Gans's central methodological move is to separate "the provision of AI" into three markets that behave differently and call for different competition analysis:

• Training data — the data used to build a prediction algorithm in the first place. Gans treats this as the component that drives entry: an incumbent with a large training-data repository has a durable advantage in building superior models, a barrier new entrants must somehow clear. Crucially, Gans notes that data is non-rival — it can in principle be shared with entrants at zero marginal cost — so market power here is not a resource-scarcity story but an incentive story: incumbents choose not to share or sell their training data, even at a price, because doing so would erode their own competitive position.
• Input data — the ongoing, real-time data an algorithm needs to generate each individual prediction (e.g., current demand conditions). Gans treats this as the component that drives within-market competitiveness rather than entry: because input data is generated continuously, whether it can be shared or traded in an arms-length market determines whether rival firms' prediction quality converges or diverges over time.
• AI predictions themselves — markets where the product is the prediction, the closest current real-world example being platforms selling predicted match quality between advertisers and consumers (an application of Bergemann & Bonatti's 2015 economics-of-data framework). Prediction-quality differences here translate directly into advertising-revenue differences between platforms with different underlying data.

The Headline Finding

Across all three markets, Gans's conclusion converges on a single structural variable: whether functioning markets exist for trading data across firm boundaries. Where data can be bought and sold between firms, market power in AI provision is transient — rivals can buy their way to competitive parity. Where no such market exists — because incumbents refuse to sell, or because no institutional mechanism for data trading has developed — the training-data and input-data advantages compound, and market power persists. The paper's own conclusion is explicit that this is unresolved territory: "Precisely how market power might impact the various markets that constitute the provision of AI is an open question," and Gans flags an unaddressed complication — firms integrated across two or more of the three markets (training data, input data, predictions) may compound their advantage through feedback loops even as that same integration requirement makes entry harder for any rival who would need to compete in all three markets simultaneously to challenge them.

Relation to the Georgist Case

This paper supplies the competition-economics half of a question the wiki's rentier economy narrative and markup-rent literature leave open: when AI-sector profits look like rents rather than returns to genuine innovation, what specifically makes them persist rather than compete away? Gans's answer — non-rival data that incumbents strategically withhold from trade — is structurally close to the Georgist land case in one respect and sharply different in another. The similarity: in both cases, a fixed or slow-to-replicate input (land; a training-data repository) grants a durable advantage independent of ongoing effort. The difference, which Gans's own framework makes precise and which this wiki should not blur: land is truly non-producible and rival in use, while training data is non-rival and, in principle, producible or licensable — Gans's whole policy-relevant finding is that the market-power outcome depends on an institutional choice (does a data-trading market exist?) rather than on a physical scarcity constant. That makes the AI case more tractable by policy than the land case in one sense (data markets could in principle be built where land supply cannot be expanded) and less settled in another (no consensus data-trading-market design yet exists, whereas the land case has centuries of assessment and taxation practice to draw on).

Nuances and Limits

• Working paper, not a magnitude study. This is a theoretical/survey paper organizing the relevant economic models (Hagiu & Wright 2023; Bergemann & Bonatti 2015, among others), not an empirical measurement of how much AI-sector profit is actually rent versus genuine returns to R&D and scale. It should be read alongside, not as a substitute for, empirical markup studies like De Loecker, Eeckhout & Unger.
• No land- or resource-rent framing in the original. Gans's paper is entirely silent on land, location, or the Georgist rent tradition — the connection drawn in this page's "Relation to the Georgist Case" section is this wiki's own synthesis, not Gans's argument, and should be read as such (D-claim, interpretive).
• Author disclosures. Gans discloses paid speaking, consulting (including antitrust and IP consulting via Charles River Associates), book royalties on the AI-economics trilogy this paper builds on, and equity/advisory relationships with AI startups — noted in the paper's own front matter. This does not by itself undermine the survey's economics, but is worth recording given the paper bears on live antitrust policy questions.

Bears On

• Research: Korinek & Vipra: Concentrating Intelligence — the redistribution/tax-design half of the AI-rents question; this paper supplies the competition-policy half of the same underlying concern.
• Research: De Loecker, Eeckhout & Unger: The Rise of Market Power — the empirical markup-rent evidence this paper's theoretical framework could, in principle, be tested against for the AI sector specifically.
• Guide: Portal: The Rent Frontier — the wiki's index of contested, non-land rent domains this paper's three-market AI-competition framework extends.

See Also

• Portal: The Rent Frontier
• Korinek & Vipra: Concentrating Intelligence
• De Loecker, Eeckhout & Unger: The Rise of Market Power
• Jan Eeckhout
• Narrative: The Rentier Economy

Sources

1. Joshua S. Gans (2024), "Market Power in Artificial Intelligence," NBER Working Paper 32270 (March 2024); published as Joshua S. Gans (2026), "Market Power in Artificial Intelligence," Annual Review of Economics 18. NBER PDF · Annual Reviews (paywalled) — the NBER working-paper PDF (the free, open-access version; the Annual Reviews version of record is paywalled and returned a 403 to direct fetch) was downloaded and read in full 2026-08-16; used for the abstract, the three-market framework (training data, input data, predictions) and its entry-vs-competitiveness distinction, the non-rival-but-withheld characterization of training data, the data-trading-markets headline finding, the multi-market-integration complication, and the verbatim conclusion quotation. Author disclosures (consulting, royalties, equity/advisory relationships) are as stated in the paper's own front matter.