Y Combinator CEO Garry Tan estimates that he spends $100,000 to $200,000 a year on AI tokens. Spread evenly over a year, that would be roughly $8,300 to $16,700 a month. The arithmetic is straightforward. The harder question is what that money buys.
Tan gave the estimate to Walter Isaacson for The Builders, in an interview recorded June 19 and released November 3, 2026. TechCrunch reported his remarks ahead of the episode. The figure describes Tan’s estimate at the time of recording. It isn’t a November invoice, a breakdown of what he spent on each tool, or a statement of Y Combinator’s AI budget.
One podcast appearance, two kinds of work
Tan said he used an agent to prepare for Isaacson’s interview. That’s a familiar assignment: research the host and anticipate questions. His public coding toolkit, gstack, illustrates a different use. It includes separate tools for reviewing an engineering plan, checking code and testing software. Someone can inspect those instructions without adopting Tan’s spending habits.
Consider a small change to a checkout page. One person might ask a model to review the changed file. Another might have an agent review the design, inspect the code, run a test and examine the failure before proposing a fix. Both people might describe the job as “review checkout,” but the second has asked the model to do considerably more work.
Tokens are the units used to measure the content a model processes or generates. Anthropic’s Claude Code documentation explains that larger working contexts consume more tokens and that each agent teammate maintains its own context. That helps explain how costs can grow even when the instruction typed by the user is short. It doesn’t tell us which tasks accounted for Tan’s spending.
Real Talk
The missing number is how much useful work survived review. Imagine spending $20 across ten attempts to fix a bug. If only one fix passes the tests and gets accepted, the model cost was $20 for that accepted fix. Dividing by ten gives you the cost per attempt, which answers a much less useful question. You still need to count the time someone spent reviewing all ten.
There’s a case for spending more when an extra check catches an error that would otherwise reach customers. There’s also a case for skipping it when a second agent repeats the first review and adds nothing. Tan’s estimate alone can’t distinguish between those outcomes.
💡 IskraCore Take
For a small team, the most useful part of Tan’s setup is the part he made public. Pick one recurring job and try one of the relevant review steps. For the next few runs, record what the model cost, how long review took and whether you used the result. If it consistently saves time or catches mistakes you care about, give it more work. That gives you a reason to set your own budget.