Once intelligence is too cheap to meter, the world will be unrecognizable

"Intelligence too cheap to meter" is a dream peddled by AI companies and accelerationists. In a narrow sense it is already true: if intelligence means solving coding and mathematics problems, AI is cheap today.

But anyone who has spent much time with current models knows they cannot yet replicate human remote work. They are too jagged, too confused about the world, and too unreliable to be trusted with a real job.

The compute market reflects this. A million output tokens from Fable or Astra costs $50, including healthy markups for Nvidia, the cloud providers, and the labs themselves. At hardware cost, the same tokens would run to a few dollars. Compared with the time it would take a human to produce a million tokens, either price is a bargain. But if the tokens were worth what a human's are, buyers would pay far more than $50 for them. The price is low because, for most work, the tokens are not yet a substitute for a person.

In the long run I expect intelligence really will be too cheap to meter, in the fullest sense of "intelligence". When current AIs can do a task at all, they do it for a small fraction of what a human costs. Given the breadth and depth of what they can already do, and the pace of progress, I expect that once we can build AGI, it will likewise take very little compute to do an hour of human work.

The world's stock of compute is also growing exponentially, as it has for the last half-century. On current trends the world's chips will probably be able to run many billions of AIs by the mid-2030s; if AGI sets off an industrial explosion then production could be far greater. Chips are ultimately manufactured goods, and it is not particularly hard to make vast quantities of them.

"Too cheap to meter" suggests a utility, something abundant and unremarkable, like tap water. But the world it describes is nothing like our own. In that world, every person on Earth is matched by one or more AIs of similar skill, and their number grows rapidly every year while ours barely changes. Humans become a small and shrinking minority of all thought.

What happens in that world is decided by the AIs and by whoever directs them. If their goals differ from ours it would be difficult to stop them: they could be substantially smarter than us, but also, there will simply be many more of them. If they are controlled by a dictator or a small elite, the AIs could be used to control or kill everyone else more thoroughly than the Stasi or the NKVD ever managed. Even if we avoid this outcome, human labor is insignificant next to the output of AI systems and robots. Outcomes run from a jobless aristocracy living off its investments to complete expropriation. Not all of these are catastrophic, but all of them are profoundly alien.

That leaves the (potentially short) period where AIs can do most human work but there are not yet billions of them vastly outproducing humans. Once AI systems are capable of performing human work, the price of compute is set by what that work is worth. Demand for remote work only saturates with billions of AIs; with fewer, the value of compute is the marginal wage of the labor they automate. If a year of an H100-equivalent can do human work valued at $250,000, it sells for something like $250,000.

For ordinary people, this means that AI is expensive during this period. People earn their income through wages, and they compete with AIs for work. An hour of labor from an AI would therefore cost roughly the same as an hour of labor from a human. Some jobs will be strongly complementary to AI, and people in them will become much more productive. Some people are rich enough to afford AI assistants regardless. But this cannot be true for most people until AI labor is cheap compared with human labor across the board. At that point we will already be in a radically transformed, potentially very scary world.

1. Chips are scarce today largely because manufacturers, uncertain about future AI demand, are far more worried about overproducing than underproducing. Fabs are also expensive, but in large part because we keep getting better at making chips, so the cost of a fab, and of the research behind each new generation, is amortized over only a few years. If that progress stopped, chips would still get cheaper over time: the fabs could be amortized over much longer periods and we would get better at building them as chip manufacturing became a commodity business.