Competition in Generative AI: Evidence from Consumer Demand (NBER w35953)
A · Strong evidence An NBER working paper by Harvard and Georgia State economists using card transaction data on 174 million US consumers and a structural demand model; strong data and method, not yet peer reviewed, with disclosed ties to the data provider's advisory council.
Summary based on the full text·Original: about 75 minutes to read
In short
NBER Working Paper 35953 (October 2026) by Mark L. Egan, Dominic Russel and Claire Shi (Harvard) and Vincent Yao (Georgia State). Using de-identified card transactions from January 2021 to April 2026, it studies who pays for consumer AI services from Anthropic, Google, OpenAI and xAI, and how competitive that market is. Paid adoption is still small and skewed to younger, higher-income people. Retention is low, multihoming is falling and switching rises when rivals release better models. A structural model finds consumers price-sensitive and the services increasingly close substitutes, so providers are unlikely to hold lasting pricing power unless model quality diverges.
Key numbers
- 2.9%US consumers paying for at least one major generative AI service in April 2026 (OpenAI 2.4%, Anthropic 0.5%)
- 45% vs 30%12-month retention of first-time paying subscribers for OpenAI against Anthropic; about 70% for Netflix and Spotify
- -4.6average own-price elasticity of demand for paid AI services
- 0.89estimated correlation between consumer preferences for OpenAI and Anthropic products
- about 25%rise in rival-to-provider switching after Anthropic or OpenAI releases an improved model
As reported by the source.
English·Web page·opens nber.org
What it reports
- Nearly all paying users are on basic plans of $30 a month or less.
- People earning over $200,000 are 1.7 times as likely to pay as those under $50,000; under-55s are 1.8 times as likely as those 55 and older.
- About 45% of Anthropic payers also paid OpenAI in June 2024; by April 2026 it was under 30%.
- Annual retention ranges from 20% to 45% across major providers.
- Estimated markups range from 19% to 59% of price, implying 41-81% of revenue covers marginal running costs; with learning-by-doing, marginal cost is roughly 70-80% of price.
- Consumer surplus among adopters is about $35 a quarter beyond what they spend; for Anthropic, prices would be 30% higher if it were not substitutable with OpenAI.
- Removing Anthropic from the market lets OpenAI raise prices; Anthropic's presence is worth $5.05 per quarter per AI user in welfare.
How they got it
Card data covering 174 million consumers and 114 billion transactions, with transaction descriptors to identify AI provider payments and demographics by age, income and region. Model capability is measured with the Epoch Capabilities Index. The demand model follows Gentzkow (2007), letting consumers choose bundles and separating correlated tastes from complementarity, estimated by maximum likelihood; supply is a Bertrand-Nash pricing game, with a dynamic learning-by-doing extension.
Limits
- Paid consumer subscriptions only; free use and enterprise contracts are excluded.
- Adoption is identified from card transaction descriptors, so payments not visible as a charge from a major provider are not counted.
- Markup estimates exclude fixed costs and R&D, so they bound profitability from above.
- Not peer reviewed; two authors sit on the data provider's advisory council.
Why it matters for companies
For buyers, this is evidence that consumer-grade AI services are close substitutes and that switching is common and cheap, which supports multi-vendor strategies and short commitments. Low retention and high price sensitivity limit what providers can charge. Their pricing power rests on staying ahead on model quality, so firms should expect continuing price pressure rather than lock-in.
Summary based on the full text. Read the original for the full argument and data.Read the original on nber.org