Y Combinator's Garry Tan Reveals $200K Annual Token Spend on AI Agents

On a recent episode of Walter Isaacson's podcast The Builders, Garry Tan put a price tag on his heavy reliance on AI agents, estimating an annual token bill between $100,000 and $200,000. He described using a personal suite of 23 Claude Code skills — published as the open‑source project gstack — to handle research, coding, email triage, and even lifelong personal notes.

Tan's agents ingest his entire digital footprint: emails, Dropbox files, Apple Notes, and therapy journals spanning two decades. He likens the experience to a constant “Jiminy Cricket” that anticipates needs, suggests alternatives, and executes tasks with a degree of autonomy that feels more like a human collaborator than traditional software.

Y Combinator itself runs an internal agent named Stray Light, which carries a personality and a “company brain” containing every founder interaction, application, and email in the accelerator's history. This agent is reshaping operational workflows across the organization.

Why it matters for GPU infrastructure

Tan stressed that the current token expense is a “crazy privilege” and that mainstream adoption hinges on lowering compute costs. As demand for autonomous agents grows, the pressure on GPU clusters to deliver cheaper inference will intensify, making efficient hardware provisioning and token‑optimized models a strategic priority for cloud providers like AiGpu.