Andrej Karpathy says he feels 'nervous' when he doesn't use up his AI token budget
Michael Macor/The San Francisco Chronicle via Getty Images
- Andrej Karpathy says he's focused on using up all of his AI tokens.
- He said he switched between tools like Codex and Claude to ensure he uses his entire budget.
- Tech leaders like Nvidia's Jensen Huang say heavy AI spending is becoming a workplace expectation.
Andrej Karpathy says he aims to use up his entire AI budget.
In an interview on the "No Priors" podcast, Karpathy — a former Tesla AI director and OpenAI cofounder — said he's shifted his mindset toward consuming every last AI token at his disposal.
"I feel nervous when I have subscription left over," he said on the pod, which was published on Friday. "That just means I haven't maximized my token throughput."
Tokens are the units AI companies like OpenAI and Anthropic use to price their models. Roughly speaking, a token can be as small as a short word or a part of a longer word; a common rule of thumb is that four characters equal one token.
For consumers and employees, tokens function like a budget: the more you use, the more work AI systems can perform.
Karpathy said that changes how he uses his AI. The constraint is no longer how quickly he can type a line of code — it's how many tokens he can deploy.
That shift has changed his goal. He now aims to "maximize subscriptions," he said, even switching between competing products as limits approach. "If you're running out of quota on Codex, you should switch to Claude."
Karpathy's comments come amid a broader rethink of how developers approach AI usage. Last week, Nvidia CEO Jensen Huang said on the "All-In" podcast that he expects employees earning $500,000 to use $250,000 worth of tokens.
"It is now one of the recruiting tools in Silicon Valley," Huang said. "How many tokens comes along with my job?"
Box CEO Aaron Levie echoed that sentiment, writing on X that the surge in AI token spending will "eventually hit the rest of knowledge work as well."
The shift suggests that access to compute is no longer the main constraint on AI output. Karpathy compared the feeling to his time as a Ph.D. student.
"You would feel nervous when your GPUs are not running," he said. "Now, it's not about flops — it's about tokens."