The token is the new kilowatt-hour

A new unit of account is crystallizing around AI. Every prompt, every agent task, every model call is now tallied in tokens, the small chunks of text and data that models read and write, and that count is starting to look less like a billing trick and more like the meter the next phase of the economy will be read on. The kilowatt-hour analogy is a useful interpretive frame, not a proven structural equivalence — token markets lack electricity's regulatory scaffolding and uniform pricing conventions.

Consumers still pay flat subscriptions, but the businesses running real AI workloads are already discovering what an electric bill feels like when the meter runs around the clock.

That pressure is reshaping behavior inside the firms that adopted AI fastest. Engineers at Uber and Amazon have moved from tokenmaxxing, where every workflow was redesigned to lean harder on the model, to tokenminimizing, a phrase coined by a writer at The Information, where guardrails cap how many tokens any one process can burn. The shift is the most legible sign yet that the unit has matured from curiosity to line item.

Economists are now reading the meter too. As WAER's reporting on tokens as the AI kilowatt-hour notes, a new working paper by Nicola Borri, Aleh Tsyvinski, and Yukun Liu uses 380 trillion AI tokens as a dataset, tying aggregate AI consumption to company stock prices. OpenAI's CEO, Sam Altman, has already framed what the meter is for — the next competitive axis, he has suggested, will not be who has the best model. It will be who controls the meter, who pays the bill, and who can be measured by it.