The CEO Who Lives in 2028
Garry Tan doesn't use AI agents as tools — he integrates them as invisible collaborators that operate in the background all day. Speaking in the recent episode of Walter Isaacson's podcast The Builders, the Y Combinator CEO revealed that he spends between $100,000 and $200,000 a year on AI model tokens, a figure he himself describes as a "crazy privilege that not everyone can afford."
The revelation builds on comments Tan first made in August 2026 during an appearance on the a16z podcast, where he coined the term "tokenmaxxing" to describe the strategy of loading hundreds of thousands of tokens into an agent at once, rather than rationing them. "You have to tune it all the way up," he said at that time. "Put 800 thousand or 1 million tokens into an agent context. When you do that, I think you basically get to 2028."
But it was in the Isaacson episode that the details of Tan's routine took on contours that go beyond a mere statement of intentions. The agent Tan uses, according to him, "read all of my email and all of my Dropbox. It read, like, all of my Apple Notes, even my therapy notes going back 20 years." What sounds like extreme digital surveillance in someone else's hands is, for Tan, the price of having a "Jiminy Cricket" always by your side, flagging problems before you even realize they exist.
The philosophy behind the bill
Tan's central argument is that startup founders have a structural advantage for absorbing the high cost of AI agents: the ability to turn expensive experiments into reusable processes. "A markdown file is an employee," he said on the a16z podcast. "It's an employee that will do the job perfectly every time, and it'll do it as many times as you want."
This reasoning places him at the forefront of a movement he himself named — "tokenmaxxing" — which advocates the indiscriminate use of tokens as a way to accelerate experimentation with autonomous artificial intelligence. For Tan, the right metric is not how much an engineer spends on tokens, but how much that engineer produces with them.
In practice, Y Combinator has already internalized this philosophy. According to Tan, the accelerator itself operates with an internal AI agent called Stray Light, which "has a personality and a corporate brain" — with access to the full context of every founder ever supported, every email received, and every application processed.
The debate: efficiency or volume?
Tan's stance is not unanimous. Praveen Neppalli Naga, CTO of Uber, argued in the week before the a16z podcast that the industry is "coming to the end of the so-called 'tokenmaxxing' era." For him, the next phase of AI adoption will not be defined by who spends the most tokens, but by who uses them most efficiently.
Scott Wu, CEO of Cognition, made a more moderate version of the same critique: the idea is "directionally correct," but some have gone too far. "People are like, 'We rank our engineers by how many tokens they're spending.' Well, let's try and rank people by how much output they're actually producing," he said.
The tension between efficiency and volume reflects a broader dispute about AI economics. On Tan's side, the assumption is that token costs will continue to fall while frontier model capability improves — making aggressive spending today an investment in learning how to build autonomous systems. On the other side, efficiency, not raw consumption, will be what separates winners from losers as AI becomes a standard business tool.
The future of personal agents
Tan believes that personal AI agents will eventually become mainstream, provided one problem is solved: the bill needs to come down. He also has reservations about the big tech model — "This is a huge company that controls the algorithm, and we all wish that they didn't at some level," he said about Meta. "So you could imagine a personal AI that you control, that you program."
What makes Tan's story relevant beyond the Silicon Valley startup ecosystem is that it reveals how far the adoption of autonomous agents has already moved beyond the prototype of the first-generation generative AI assistant. This is no longer about "a chatbot that answers questions" — it's about a system that operates continuously over a person's data and workflows, making decisions, anticipating problems, and executing complex tasks without direct human intervention.
Whether today's $100,000 to $200,000 cost per personal agent is a temporary price that will fall as AI infrastructure matures, or whether it represents the beginning of a permanent divide between those who can afford to operate in 2028 with today's technology and those who have to wait for the capability to reach the price they can pay remains an open question.
Sources: TechCrunch, Business Insider, Pivot News
✓ Independent sources cross-checked and verified before publishing
