Y Combinator CEO Garry Tan has a message for founders and CEOs concerned about the cost of using large numbers of AI tokens: “Burn, baby, burn.” Speaking on an episode of the A16z podcast this week, Tan argued that founders should spend heavily on AI agents, even when doing so can result in high annual costs.Tan said founders should give themselves permission to use large token budgets when working with AI agents, arguing that doing so can provide access to a level of AI capability that he believes will otherwise become available only in the future.“You have to tune it all the way up. You’re just, like, let me load a million tokens or 800,000 tokens in,” he said about using AI agents. “When you do that, I think that you basically get to live in 2028.”
Garry Tan backs heavy AI token spending
Tan acknowledged that using AI agents at full capacity can be expensive. He estimated that the cost could reach tens of thousands of dollars per year, but said the spending can make sense for founders and CEOs experimenting with AI-driven workflows.“It costs, I don’t know, 50 or $100,000 a year to use the agents at full strength,” he said. “It just costs like a crazy amount. But for a CEO or for a founder, it actually makes a lot of sense to do that.”His approach is to use AI agents for tasks first, then turn successful workflows into repeatable processes. Tan said founders can document the steps an agent follows and use those instructions to automate the same work in the future.“A markdown file is an employee,” he said, referring to a written set of instructions that tells an AI agent how to perform a particular job. “It’s an employee that will do the job perfectly every single time, and it’ll do it as many times as you want.”
AI tokenmaxxing divides Silicon Valley
Tan’s approach contrasts with criticism of “tokenmaxxing,” a term used for deliberately allowing AI systems or agents to consume large numbers of tokens instead of prioritising lower usage and costs.Uber’s tech chief Praveen Neppalli Naga recently argued that companies may be moving away from this approach as they focus more closely on the efficiency of AI spending.“The next phase, whatever we call it, will not be characterised by who spends the most tokens, but about how people use them as efficiently as possible,” he said in a post on X.Cognition CEO Scott Wu also questioned the way some companies have approached tokenmaxxing. Speaking on a podcast in June, Wu said companies could lose sight of the relationship between token consumption and actual work produced.“It is directionally correct, but I think there are definitely some places where people have gotten carried away,” Wu said. “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.”
AI agents become part of founder workflows
Tan’s argument focuses on using higher AI token limits as an investment in experimentation rather than treating token consumption only as a cost to minimise. He recommends first allowing agents to use the resources needed to complete a task, then converting successful processes into reusable instructions.The approach reflects a broader discussion among technology companies about how much they should spend on AI inference as increasingly capable models and agents are used for software development, research and other business tasks.
