The Physical Stack of AI · The inference economy
You can state precisely what Altman and Amodei have claimed about the cost of intelligence, where the slogan "too cheap to meter" comes from, and which parts of that promise have actually arrived.
The phrase "too cheap to meter" is older than the chatbot industry — it was Lewis Strauss's 1954 pitch for civilian nuclear power, and history has been unkind to it. In 2025, Sam Altman picked it up for AI. In Three Observations, he wrote that "the cost to use a given level of AI falls about 10x every 12 months," and projected a near future in which intelligence is functionally free at the point of use. Dario Amodei's Machines of Loving Grace made a related, more careful version: a "country of geniuses in a datacenter" delivering compressed decades of scientific progress.
The numerical claim has substantially come true. GPT-4 cost ~\$30/Mtok output at launch. Three years later, GPT-5.5 ships at \$30/\$5 for better capabilities; Haiku-class outputs are at \$5/Mtok, and DeepSeek V4-Flash fractions of a cent. On a 10×-per-year curve, a 2025 Opus-tier task is heading toward Haiku pricing within 18 months.
Chapter contains 3 lessons.