Y Combinator's Garry Tan Says He Spends Up to $200K a Year on AI Tokens

Garry Tan of Y Combinator says he spends as much as $200,000 a year on AI model tokens, a figure that puts a price tag on the kind of heavy agent use most people only read about. TechCrunch reported the comment on November 2, attributing it to an appearance on The Builders, the podcast hosted by Walter Isaacson.

Tan jokingly described suffering from “cyber psychosis” and said he sometimes barely sleeps because he is so excited to work with his AI agents. The self-deprecation is light, but the underlying intensity is not: a $200,000 annual token bill implies near-continuous, high-volume use of frontier models.

A second detail travelled further than the number itself: Tan let an AI agent read 20 years of his therapy notes. Hindustan Times put that in its headline, a reminder of how quickly agent convenience turns into a question about who gets to hold a person's most sensitive records.

The token bill matters because it previews the cost structure of the agent era. When a model stops answering questions and starts carrying out tasks, call volume, context length and runtime all multiply, and spending shifts from flat subscriptions toward cloud-like metered usage.

For founders, that is a useful benchmark. Heavy agent use is no longer a free productivity toy but an operating expense that belongs in a budget, and Tan's case shows there are power users willing to pay, as long as the hours saved are worth more than the tokens burned.

For most individuals the figure is obviously extreme. It still points to a broader question: if AI is priced per token, who can afford sustained agent use, and does metered pricing concentrate agent capability among those who can?

Privacy is the other half of the story. Handing years of therapy notes to a model sends extraordinarily sensitive material into a third-party system, raising questions about retention, training use and access controls. Tan appears comfortable with that trade; it is not advice for everyone else.

What to watch next is whether Tan names the specific tools and models behind that bill, and whether falling inference prices turn a $200,000-a-year habit into something ordinary teams can contemplate.