Satya Nadella calls the reflex of running the most powerful, expensive AI on every problem 'addictive,' and admits he is guilty of it. The fix he pitches (developer led swarms of AI coding agents) will burn more of those models, not fewer.
Satya Nadella, the CEO of Microsoft, says he is a "token-maxer," and that the habit is a problem. The admission lands harder than the warning that preceded it, because the company he runs is best positioned to profit when the world's developers and knowledge workers keep reaching for the most expensive AI on the shelf.
"I am like a token-maxer too. So it is addictive," Nadella said in a recent YouTube interview titled "Satya Nadella on A.I. Jobs: Humans Will Do the Glue Work", as reported by The Decoder. "Token-maxing" is the reflex of using the most powerful and priciest AI model for every task, even when a smaller, cheaper one would do.
The economic rule Nadella offers in the same interview is plain enough to carry to any knowledge worker's tool stack. "The hard truth is that the marginal cost of productivity improvement has to match the marginal cost of the token," he said, per The Decoder. In plain language, each extra dollar of AI spend should buy a real extra dollar of useful work, not just the feeling of having used the smartest model available.
That is the warning. The cure Nadella proposes, however, runs on the same tokens he is warning against. He sketches a future in which developers oversee "hundreds or thousands" of AI coding agents rather than writing code themselves. To keep up, he coins a job description he calls "cognitive coverage": the work of deeply understanding the code that AI agents have produced, so a human can actually verify and steer what the swarm is doing.
"I have a repo full of code written by agents. I am cognitively understanding what happened," Nadella said, per The Decoder. In a workplace built around hundreds or thousands of agents, that work does not shrink the bill. It grows it, because every verification pass and every steering prompt is another call to a frontier model.
This is the tension worth sitting with. The CEO of the company best positioned to monetize frontier tokens is publicly naming the reflex to use them as an addiction, while pitching a roadmap that will burn those tokens at higher, not lower, volume. Nadella is not lying in either direction. He is naming a real productivity trap, and then describing the future his own company is selling into it. The reader does not have to resolve that contradiction. They can run the math on their own stack.
The test Nadella himself offers is portable. Does the productivity gain from the model you reached for actually match the cost of the model? For a developer staring at a swarm of agents, the question gets harder, not easier, because the cost of a wrong answer is multiplied across every agent they sign off on.