Only 2.9% of US Consumers Pay for Generative AI, Harvard Study Finds

Just 2.9% of US consumers paid for at least one major generative AI service as of April 2026 — roughly 6 million people, about two-thirds the subscriber base of YouTube TV. That's the headline finding of a Harvard Business School working paper published in October 2026, built on 114 billion transactions processed by fintech firm Fiserv from early 2021 through April 2026.

The paper, "Competition in Generative AI: Evidence from Consumer Demand," comes from HBS professor Mark Egan, HBS doctoral students Dominic Russel and Claire Shi, and Georgia State University professor Vincent Yao. It may be the most detailed picture yet of who actually pays for generative AI — and how long they stay. The answer: very few people, and not for long.

What does the data cover?

The research tracks consumer credit- and debit-card payments to four companies: OpenAI, Anthropic, Google, and xAI. Enterprise contracts are excluded, which skews the revenue picture. Consumer subscriptions made up about 60% of OpenAI's revenue at the start of 2026, but only around 20% of Anthropic's historical revenue, where enterprise clients dominate.

The cutoff is April 2026, so the paper reflects nothing after that date.

Who actually pays?

Paying subscribers skew sharply affluent, young, and urban. They tend to earn more than $200,000 a year — more than twice the US median income. People under 55 and residents of coastal or major metropolitan areas are significantly more likely to subscribe.

In short: the consumer AI market is small, rich, and concentrated.

How bad is the churn?

Bad, by subscription standards. After 12 months, roughly 46% of OpenAI subscribers remained, compared with 35% for xAI, 28% for Anthropic, and 24% for Google. The New York Times retained about 53% of digital subscribers at one year; Spotify and Netflix held roughly 67% and 66%.

The researchers built a "multihoming index" for AI. Americans routinely pay for multiple media services — Netflix and Disney+, for example — but willingness to hold multiple AI subscriptions started high and has been declining across all four providers.

Are the services interchangeable?

Increasingly, yes. The share of Anthropic users who also paid for OpenAI fell from 44.3% in mid-2024 to 28.3% in April 2026. Over the same period, the share of OpenAI users also paying for Anthropic rose from 2.1% to 5.6%. The researchers read this as the two services becoming more interchangeable: as tools combine coding, writing, and other functions in a single product, consumers see less reason to keep several subscriptions.

Model releases drive switching. The GPT-4o and o1 launches in 2024 saw the highest rates of Anthropic-only subscribers adding or switching to OpenAI — around 7.9% and 7.4% respectively.

How much pricing power do providers have?

Not much, at least on this evidence. Most subscribers pay at the low end. The share paying less than $35 a month was 97% for OpenAI, 94% for Google, 88% for xAI, and 76% for Anthropic. Anthropic had a longer tail: 10% paid $35–$99, 9% paid $100–$199, and 5% paid $200 or more.

The researchers' demand model estimates that a 10% price increase would cut demand by 46% — a modeled response, not an observed reaction to a real price change. They also estimate that consumers in the top quartile of preference for ChatGPT would pay $37 more per month than those in the bottom quartile.

The model says consumers currently capture a large share of the surplus: paid subscribers gain about $35 per quarter in extra value. "Absent a substantial increase in AI model quality differentiation, growing substitutability among services will constrain providers' pricing power," the authors write.

The paper also runs a counterfactual: if Anthropic exited the market, OpenAI prices would rise 2%, xAI's 1.1%, and Google AI prices would fall 4.4%. OpenAI's market share would dip 0.7%, Google's would drop 7%, and xAI's would rise 17.1%.

What are the caveats?

Egan and Yao are unpaid members of Fiserv's Small Business Index advisory council; Yao also serves as a paid consultant to Fiserv. The paper is a working paper, not yet peer-reviewed.

The findings land as Anthropic prepares for a potential public offering and OpenAI works toward a possible 2027 IPO. The consumer revenue base those offerings would rely on, the research suggests, is small, concentrated, and quick to leave.