Garry Tan spends $200,000 a year on AI tokens. Just like that — a sum most small businesses can’t afford to lose, yet here it is, burned through like jet fuel. The Y Combinator CEO isn’t buying software licenses or hiring interns. He’s paying for the raw computational power that keeps his digital workhorses running.

It’s a staggering figure. It highlights a shift no one predicted 3 years ago. The cost of intelligence is dropping, but the cost of using it aggressively is skyrocketing. Tan admits to “cyber psychosis.” He barely sleeps.

He told Walter Isaacson on his podcast, *The Builders*, set to publish Tuesday, that he’s obsessed. This isn’t just a tech flex. It’s a warning.

The Price of Obsession

Tan’s spending habits have become a benchmark for the new wave of AI-native executives. He doesn’t treat these agents as tools. He treats them as employees. The distinction matters because it changes how companies budget for technology.

A $200K bill is not a line item for payroll. It’s a line item for experimentation, for failure, and for the sheer volume of trial and error. Most enterprises still view AI as a cost-saving measure. Tan views it as a capacity multiplier.

The math works out if you have the revenue to back it up. Y Combinator has that revenue. The startup it incubates often don’t.

The detail that landed hardest in the podcast was the sleep deprivation. Tan spoke about the hyper-stimulation of managing autonomous agents that never stop. It’s a problem specific to this era of tech. Previous generations of founders had the luxury of silence.

They could step away from the desk. Tan can’t. His agents are running experiments, generating code, and analyzing data while he’s trying to rest. The psychological toll is real and documented in early adopter circles.

Nobody in the room looked convinced this is a sustainable long-term strategy. It feels like a sprint, not a marathon.

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

What’s Inside the $200K Bill

Token costs have plummeted in raw terms, but usage has exploded exponentially. Tan’s expenditure reflects a trajectory that few traditional IT departments are prepared for. We’re not talking about one chatbot interface. We’re talking about hundreds of parallel instances, each consuming memory and processing power.

The granularity of pricing has changed. You pay for every word generated, every image rendered, and every calculation made. It’s additive. It’s relentless.

The bill accumulates silently in the background while the user is away. Comparing this to 2015 software spending reveals a stark difference. Back then, you paid once for a license and then had to take a support call to fix a bug. Now, you pay continuously for the agent’s “thought process.” A single complex problem might require an agent to simulate 50 different scenarios. Each simulation costs money.

If the agent fails, you pay again to fix the error. Tan’s approach is brute force. He throws compute at problems until one works. It’s inefficient in human terms.

It’s efficient in machine terms. This is the new economics of innovation.

India’s AI Ambition Meets Token Economics

For Indian readers, this story lands with a different weight. India is positioning itself as the next major AI hub, with heavy government backing for data centers and talent pipelines. But the cost of inference is a hidden barrier. If a US founder can burn $200K a year without blinking, what does that mean for a Bengaluru startup operating on leaner margins?

The exchange rate alone makes this expensive. A 200,000 USD bill is roughly 1.65 crore INR. That’s a venture round of Series A funding gone in one year on compute. It’s a sobering calculation for founders in the global South.

However, there’s an opportunity in efficiency. Indian engineers are historically excellent at optimizing code and reducing overhead. The challenge is shifting from optimizing for cost to optimizing for output. If Indian startups can build agents that achieve 80% of Tan’s results for 20% of the price, they have a massive competitive edge.

The global market is hungry for cheaper, faster intelligence. This isn’t just about copying Tan’s spend. It’s about building the infrastructure that allows cheaper, more efficient token consumption. The policy angle is also shifting.

India’s AI mission needs to account for the operational costs of AI, not just the R&D costs. This is where the rubber meets the road.

The Psychology of Sleepless Innovation

Tan’s admission of “cyber psychosis” is more than a quip. It points to a deeper cultural shift in tech leadership. The idea that you must be always-on, always-tethered to your digital workers, is taking hold. It’s a dangerous precedent.

Burnout was already a staple of the startup world. Now it has a new fuel source. The agents don’t sleep, so the founder feels they can’t. It creates a feedback loop of anxiety and excitement.

The dopamine hit of watching an agent solve a problem at 3 AM is powerful. It’s addictive. And that’s a problem for boardrooms and hospitals alike.

Historically, we’ve seen this pattern with other disruptive techs. Early internet pioneers were obsessed with uptime. Blockchain developers have talked about “node fatigue.” It’s a familiar refrain. But the scale has changed.

In previous eras, you could turn off the computer. Now, the work continues without you. You check in, see what they’ve done, and then you’re pulled back in. The line between work and life has dissolved.

It’s not just about hours worked. It’s about cognitive load. The brain doesn’t rest when the mind is still processing news of what your machines are doing. This is a human cost we’re only starting to measure.

Market Implications for AI Spenders

The $200K figure is a signal to investors. It suggests that AI operational budgets will become a major line item in startup financials. VCs need to ask harder questions about token efficiency. They need to know if their portfolio companies are burning cash on inefficient prompt engineering or if they’re genuinely scaling value.

It’s a new metric for due diligence. “Token burn rate” might become as important as “burn rate” for cash. The distinction is subtle but vital. One is failure.

The other is growth. Or is it? Maybe it’s just expensive trial and error.

For public companies, this creates a tension. Shareholders hate uncontrolled spending. They want predictable costs. AI is unpredictable by design.

Agents make mistakes. They retry. They explore dead ends. All of that costs money.

That’s a harder sell than it sounds to a CFO who managed a fixed budget for 20 years. We’re moving from a world of fixed costs to a world of variable, behavioral costs. It’s a shift that requires new financial models. And new human tolerance for waste.

Because waste is where learning happens. And learning is where the future is. Tan uses a high volume of autonomous AI agents for daily operations, research, and code generation. These agents run continuously and require significant computational resources, leading to high token consumption. The cost reflects an aggressive strategy to leverage AI for maximum output, even if it means a hefty monthly bill.

Frequently Asked Questions

Why does Garry Tan spend so much on AI tokens?

What are AI tokens and why are they expensive?

Tokens are the units of text or data processed by large language models. They are expensive because they require substantial GPU power to generate. The cost is tied to the complexity of the model and the length of the output. As models get smarter, the energy consumption per token can vary, but usage volume is the primary driver of spend.

How does this affect Indian tech startups?

Indian startups face higher relative costs due to currency exchange rates and leaner funding structures. This creates a pressure to optimize for efficiency rather than scale. The opportunity lies in developing cost-effective AI solutions that can compete with US-heavy spending models on a tighter budget.

Is $200K a year for AI tokens normal?

No, it is not normal for most companies. This is an outlier case driven by a founder’s personal obsession and the massive scale of his operations. Most enterprises and startups spend a fraction of this amount, focusing on specific use cases rather than general-purpose agent swarms.

What is “cyber psychosis” in this context?

“Cyber psychosis” is Tan’s colloquial term for the mental state of being overwhelmed by the constant flow of information from his AI agents. It refers to the lack of sleep and the hyper-focused, anxious state that comes from managing autonomous systems that never stop working. Source: TechCrunch

Source: TechCrunch