Y Combinator's CEO to Founders: Don't Skimp on AI Tokens — It's a Ticket to the Future, Today

Garry Tan, CEO of startup accelerator Y Combinator, has a blunt message for founders worried about burning through too many AI tokens (the unit of measure for the cost of querying an AI model and getting a response back): "Don't skimp on fuel." Speaking on the a16z podcast, Tan said founders should give themselves total freedom to use AI agents — even when it comes at significant cost.

"You have to run it at full throttle. Instead of second-guessing whether to load a million tokens or 800,000, just load it," he said of using AI agents. "When you do that, you start living in 2028" — meaning founders can access today the kind of AI capability most people won't have until years from now.

Tan acknowledges the payoff isn't cheap: "Running agents at full capacity costs $50,000 to $100,000 a year. That's a huge amount of money. But for a CEO or a founder, it genuinely makes sense," he said. He argues founders can turn that costly experimentation into something that pays off over time: once an agent successfully completes a task, that task should be turned into a "skill" — a reusable set of instructions the AI can then execute automatically, again and again. "A markdown file is an employee," he said. "It's an employee who nails the task perfectly every time, as many times as you want."

Tan isn't alone in this view — but it's far from a consensus one: plenty of other voices in Silicon Valley consider "tokenmaxxing" — deliberately running up token spend — pointless. Last week, Uber's chief technology officer, Praveen Neppalli Naga, said the company was seeing "interesting data" on its AI spending and called it "another sign that the tokenmaxxing era is ending." "The next phase won't be defined by who burns the most tokens, but by how efficiently people use them," he wrote on X.

Cognition CEO Scott Wu, whose company builds AI coding tools, made a similar point on a podcast back in June, saying companies that encourage tokenmaxxing have "lost the plot": "It's not the wrong direction overall, but some people have taken it too far in places. People say, 'We rank our engineers by how many tokens they burn.' Let's rank people by the results they deliver instead," Wu said.