Y Combinator CEO Garry Tan spends somewhere between $100,000 and $200,000 a year on AI tokens. He said so on Walter Isaacson’s podcast The Builders, as TechCrunch AI reports. That’s more than a lot of companies pay a full-time engineer. Tan says agents now help with “almost anything that you do in your day-to-day life.”
The number is the headline. What it tells us about where personal AI is heading matters more.
What Tan Actually Does With It
Tan isn’t paying for a fancy chatbot. He’s paying for an assistant that knows nearly everything about him. According to TechCrunch AI, his agent has read “all of my email and all of my Dropbox,” his Apple Notes, and “even my therapy notes going back 20 years.”
He uses it for:
- Research and prep, including getting ready for Isaacson’s interview
- Coding, backed by gstack, his public set of 23 Claude Code skills on GitHub
- Life admin, with the agent acting like “a Jiminy Cricket on your shoulder” that will “run ahead and say, ‘What about this?'”
YC runs on the same idea. Its internal agent, Stray Light, has a personality and what Tan calls a “company brain,” with full context on every founder YC has backed, every email, and every application.
The Bigger Shift: Context Is the Product
What stands out here is that Tan’s edge doesn’t come from a better model. It comes from context. His agent is useful because it has 20 years of his life to work with. Stray Light is useful because it holds YC’s whole institutional memory.
That’s where the industry is going. Model quality is converging fast. Whoever controls your context, meaning your email, files, notes, and history, controls the value. So expect the fight among OpenAI, Anthropic, Google, and Apple to move from benchmark scores to who gets access to your data and keeps it.
Tan also pushes back on a common enterprise habit. He warns that people get so focused on “repeatability and safety that they actually trade off all of the beautiful parts of working with an AI agent.” Plenty of security teams would argue the opposite, especially when the agent has read your therapy notes. Both sides have a point. A locked-down agent is often a useless one. An agent with no guardrails can leak your life.
The Cost Problem (and the Control Problem)
Tan is blunt about the catch: “It’s too expensive right now.” He calls his spending “a crazy privilege that not everyone can do.”
He also doesn’t think free agents from Big Tech are the fix. Pointing to Meta, he said, “This is a huge company that controls the algorithm, and we all wish that they didn’t at some level.” His alternative is “a personal AI that you control, that you program.”
Remember who’s talking, though. Tan runs the world’s best-known startup accelerator and has a clear stake in a future full of independent, agent-first startups. His enthusiasm is real, and so is his incentive. His forecast that personal agents go mainstream sounds plausible, but it hangs on one condition he named himself: cheaper inference.
The Future Cast: 2027 to 2029
Here’s how this probably plays out over the next one to three years:
- Token prices keep falling, but usage climbs faster. Per-token costs have dropped steeply year after year. Agents that run all the time and read everything will soak up those savings, so total bills may not shrink much for heavy users.
- Personal agents split into two camps. Cheap, ad- or platform-subsidized agents from Big Tech on one side. Paid, user-controlled agents (open-source stacks, local models, bring-your-own-key setups) on the other.
- “Company brains” become standard. Stray Light is an early version of what most firms will build: an agent with full institutional memory and its own access controls.
What You Should Do Now
- Audit your context. Figure out which data would make an agent truly useful to you or your team, and where it lives.
- Track agent spend like payroll. If agents do real work, budget for them as real labor, with a cost per task.
- Decide on control early. Choose whether your agent layer runs on a platform you rent or a stack you own. Switching later gets harder once years of context are locked in.
- Set guardrails without killing usefulness. Scope access per task instead of giving the agent everything or nothing.
Tan’s $200K bill looks extreme today. In a few years, a smaller version of it could just be a normal line on the budget. The full interview details are in the original TechCrunch AI piece.