Garry Tan says he spends up to $200,000 a year on AI tokens

Y Combinator CEO Garry Tan estimates that he spends $100,000–$200,000 a year on AI tokens. In a new interview with Walter Isaacson for The Builders, he describes agents that help with coding, draw on deeply personal records and support YC’s operations. Yet he says personal agents remain too expensive to become mainstream.

TechCrunch covered the interview ahead of the episode’s scheduled Tuesday release. Tan’s spending figure is his own rough estimate. He pairs it with an acknowledgment that this level of use is a privilege, rather than something everyone can afford: he calls his budget “a crazy privilege that not everyone can do.”

An assistant with access to his private life

Some of Tan’s uses are straightforward work assistance. He uses agents on coding tasks and other projects, and had one research Isaacson’s podcast to help him prepare for the questions he might face. But his description of the agent’s access goes well beyond material assembled for a single assignment.

Tan says his agent has read his email, Dropbox and Apple Notes, including therapy notes going back 20 years. That is the personal context behind the kind of assistant he describes: something that knows him, helps with everyday activities and can suggest a next step rather than wait for every instruction.

He likens that assistance to having Jiminy Cricket on his shoulder. In his telling, the appeal is not simply getting an answer faster. It is having software that can run ahead, offer an idea and help him anticipate something he might otherwise miss. Those are Tan’s descriptions of the experience, not measured performance results.

YC’s agent draws on the organization’s records

At Y Combinator, Tan says an internal agent named Stray Light is helping rebuild how the organization operates. He describes it as having both a personality and a “company brain,” with context about every founder YC has backed, people who have emailed the organization and its applications.

Tan, an early OpenClaw adopter and now a power user, argues that agents should not be treated like traditional software. He says excessive concern about repeatability and safety can sacrifice what he values in working with them: behavior that can feel similar to interacting with a person.

So you could imagine a personal AI that you control, that you program.

Garry Tan, speaking to Walter Isaacson on The Builders, as quoted by TechCrunch

Lower prices, without surrendering control

For all his enthusiasm, Tan identifies cost as an obstacle to wider adoption. He believes personal agents will become mainstream, but says prices need to fall first. His assessment of the current economics is blunt: “It’s too expensive right now.” The six-figure budget he describes is not his prescription for ordinary users.

Price is not his only concern about the future he wants. Discussing Meta’s influence as a social media company, Tan criticizes the concentration of control over its algorithm. He contrasts that arrangement with the possibility of a personal AI that its user controls and programs.