Y Combinator’s Garry Tan uses AI agents so much, he spends up to $200K a year on tokens

When Garry Tan talks about 'cyber psychosis,' most people probably think he is joking. But when you look at the receipt, the joke starts to look more like a business strategy. The Y Combinator president recently admitted on Walter Isaacson’s podcast that he is spending upwards of $200,000 a year just on API tokens to fuel his personal AI agents. That is not a typo. That is the salary of a senior engineer at a mid-sized startup, all spent on inference calls.

For those of us building in the trenches, this number is jarring. Most founders are trying to optimize their prompt engineering to shave cents off their monthly bill. Tan is leaning into the opposite direction, essentially brute-forcing his way into the future. It raises a massive question for the rest of the ecosystem: Is this the new cost of staying relevant, or is this just what happens when a venture capitalist gets a new toy?

The Founder as a Super-User

Tan’s argument is that he is staying close to the metal. By spending his own time and a significant amount of capital on these agents, he is acting as his own guinea pig. He claims he barely sleeps because he is so obsessed with what these agents can do. In the world of crypto and AI, we call this 'eating your own dog food,' but at a $200k price point, it’s more like eating at a Michelin-star restaurant every night of the week.

From a builder's perspective, there is something to be said for this level of immersion. If you are leading the world's most influential startup accelerator, you cannot afford to be an armchair theorist. You have to know exactly where the hallucinations start, where the latency kills the experience, and where the current models fall flat. Tan is buying data points that most people can only read about in research papers.

The Economics of 'Cyber Psychosis'

Let’s break down that $200,000. If we assume he is using top-tier models like GPT-4o or Claude 3.5 Sonnet, that is millions upon millions of tokens every single month. This isn't just asking a chatbot to summarize an email. This is likely an army of autonomous agents running in loops, scraping data, synthesizing reports, and perhaps even writing code for internal tools. It is a level of activity that suggests he has replaced entire workflows with automated systems.

However, for the average founder, this model is unsustainable. We are currently in a period where venture capital is getting tighter and efficiency is being rewarded over pure growth. If a seed-stage founder told their board they were spending $16,000 a month on personal LLM experimentation, they would be laughed out of the room. Tan has the luxury of being his own board, but the lesson for builders isn't to copy the spend—it's to copy the obsession.

Why the Spend Matters for Builders

If you are building an AI startup right now, you should be looking at Tan’s behavior as a signal for market demand. If a busy executive is willing to drop six figures on token costs, it means there is immense value locked behind these models that hasn't been productized yet. We are still in the 'raw material' phase of AI. The builders who will win are the ones who can take that $200k worth of utility and package it into a $200-a-month subscription that actually works for the rest of us.

• Efficiency is the next frontier: As token costs eventually drop, the winners will be those who have mastered the orchestration of these agents.
• The feedback loop: Tan is getting immediate feedback on what works. Founders need to find ways to get that same feedback without the massive price tag.
• Personalized agents: The fact that Tan is building these for himself suggests that off-the-shelf solutions aren't meeting the needs of high-level power users yet.

The Skeptic’s View

Now, let's be honest. There is a bit of theater here. Mentioning a $200k spend on a podcast hosted by Walter Isaacson is a great way to signal that YC is the epicenter of the AI boom. It’s branding as much as it is R&D. We should be careful not to mistake a high bill for high productivity. It is very easy to spend a lot of money on LLMs by running inefficient loops or asking models to perform tasks they aren't ready for yet.

For builders, the trap is thinking you need to spend like Garry to get results like Garry. You don't. The real advantage isn't the tokens; it's the fact that he's actually using the tools instead of just talking about them. In the crypto world, we saw this with 'yield farming'—the people who spent the most weren't always the ones who understood the tech the best; they just had the most capital to risk.

The goal for a founder shouldn't be to maximize token spend, but to maximize the insight gained per token.

We are seeing a shift where the 'builder' identity is becoming synonymous with 'agent operator.' If you aren't currently experimenting with how agents can automate your own life, you are going to find it very hard to build tools that automate others'. You don't need $200k, but you do need the curiosity that leads to that kind of investment.

What This Means for the Future of Work

Tan’s 'cyber psychosis' is a look at a future where we all have a personal cloud of agents working for us. But it also highlights a growing divide. If the best tools require a $200k-a-year overhead, we are going to see a massive gap between the 'AI-enabled' elite and everyone else. As a builder, your job is to bridge that gap. The real opportunity isn't in helping VCs spend more on OpenAI; it's in making those same capabilities accessible to the person who can only afford $20 a month.

Ultimately, Garry Tan’s token bill is a loud, expensive signal that we are still in the very early innings. When the leaders of the industry are this obsessed with the raw tech, it means the finished products are still a long way off. That is good news for founders. There is still plenty of room to build, optimize, and disrupt before this becomes a commodity market.

The Takeaway

Don't get distracted by the dollar amount. The real takeaway is that one of the most successful investors in history is spending his nights building agents because he sees a level of utility there that the general public hasn't grasped yet. Stay skeptical of the hype, but don't ignore the sweat equity (and actual equity) being poured into this space. Build for utility, watch your margins, and keep your agents running—even if your bill is a lot smaller than Garry's.

Read the original at TechCrunch Startups →