One AI lab could soon have more AI workers than there are people on Earth

Imagine one company going from 10 million workers this year to 100 million "next year to a billion the year after that," as podcaster Dwarkesh Patel put it. He says it is plausible that around 2030 a single lab will hold more working "equivalents than there are people on Earth." For now, he adds, AI systems cannot yet handle entire jobs by themselves or work as independently as people, so those numbers measure potential computing capacity, not actual employees.

The math comes in two parts. The amount of raw computing power at the frontier, measured in basic calculations, is growing four to five times a year. At the same time, the amount of computing needed to reach a given level of skill is falling about three times a year, because models and chips get more efficient. Multiply more power by less need, and the effective population of AI workers a leading lab can support grows about ten times a year.

That growth favors the winner, Dwarkesh Patel argues. He says training an AI system is expensive once but cheap to copy, so the cost can be spread across billions of uses. When computing power is scarce, he says, the lab that is slightly ahead can also use that scarce power more profitably and charge more for it. He raises another possibility, that if models learn from deployment, a widely used model could get much more real world data, which would widen the lead further.

He stresses that this pull toward centralization holds even without machines that rapidly improve themselves, a process researchers call recursive self improvement. With that kind of self improvement, he says, growth could jump to a hundred or a thousand times a year.

Semiconductor analyst Dylan Patel agrees that every force points toward centralization. Dwarkesh Patel gives the outcome as most of the world's working minds ending up inside just two labs.