From Tokenmaxxing to Token Budgeting

|

,

A recent NPR article by Greg Rosalsky argues that AI tokens may become a standard economic measure, much as the kilowatt-hour became the standard measure for electricity. Along the way he tracks how companies went from maximizing their AI use to reining it in. He also highlights the AI Premium study, published as a working paper by the National Bureau of Economic Research (NBER).

Below are the key points:

  • What a token is. Rosalsky describes tokens as a meter running in the background. Every time a model reads a prompt or writes an answer, that work is counted in tokens. They are small chunks of text and other data. More work usually means more tokens. OpenAI chief executive Sam Altman has said he expects intelligence to be sold on a meter, like electricity or water.
  • Consumers pay flat, businesses pay by the token. Most individuals still buy a monthly subscription, Rosalsky reports. Businesses and developers increasingly pay per token instead. As their AI use grew, many found the bills grew faster.
  • From tokenmaxxing to guardrails. For a while, tech workers used AI freely. Some called it tokenmaxxing. Rosalsky reports that companies including Uber and Amazon have since capped usage to bring costs down. A writer at The Information dubbed the correction tokenminimizing.
  • Tokens leave a trail. This is the part economists care about. Token data is a record of AI use as it happens. Rosalsky notes that researchers could watch AI spread in near real time, rather than waiting on surveys, earnings calls, and company announcements.
  • What the study did. Economists Nicola Borri, Aleh Tsyvinski, and Yukun Liu analyzed 380 trillion tokens. The data came from OpenRouter and covered January 2024 to April 2026. OpenRouter is a single service that gives developers access to hundreds of models at once, which is why it sees so much traffic. The data is anonymized. It spans more than 400 models and roughly 2 percent of global monthly AI use. From it, the economists built a measure of overall AI consumption they call the AI Factor.
  • The AI premium. They then asked a simple question. As AI use rises, whose share prices rise with it? Companies most sensitive to that measure went on to earn higher returns. The paper puts the gap at about 0.64 percentage points a week. It was strongest for heavy use of frontier models, and absent for casual use.
  • Not just tech. The pattern reached well past Silicon Valley. Rosalsky lists airlines, cruise lines, utilities, manufacturers, retailers, banks, and even waste management. The effect was strongest in the United States and Europe. It was absent in China and other emerging markets. “The story of AI is no longer just a Silicon Valley story,” Tsyvinski told him. Financial markets, he added, already see Main Street being affected.
  • A signal about work, too. The paper goes further than the article here. It finds market-implied AI exposure is more positive for nonroutine interactive work. It is more negative for analytical, scientific, and operations-control skills. The authors also report early evidence of a rising agentic economy.
  • Plenty of caveats. The paper has not been peer reviewed. OpenRouter users are likely heavy, cost-conscious users, not typical ones. Stock prices reflect what investors expect, and investors are often wrong. If markets are efficient, much of that optimism is already priced in. Rosalsky is blunt that none of this is financial advice.

-maxxing glossary

Some commonly used AI-related -maxxing lexicon glossaries that circulate informally in AI and tech discussions include:

  • Tokenmaxxing. Using AI as much as possible, with little attention to cost.
  • Tokenminimizing. The correction. Cutting token spending and putting limits in place.
  • Measurement-maxxing. A humorous term for the recent surge in interest in accurately measuring AI.
  • LLMmaxxing. Reaching for a language model for nearly every thinking task.
  • Promptmaxxing. Getting very good at prompting to improve results.
  • Contextmaxxing. Feeding the model rich, well-organized context.
  • Benchmarkmaxxing. Tuning for benchmark scores rather than real capability.
  • Agentmaxxing. Building workflows around agents to automate as much as possible.
  • Copilotmaxxing. Keeping an AI assistant running constantly to raise output.

The through-line in Rosalsky’s reporting is a shift in how AI gets counted. Tokens started as a billing unit. They are becoming a measuring unit. That matters more than any single finding about share prices, because past technological revolutions were largely studied after the fact. This one can be watched as it happens. Rosalsky ends on the point that better data is rarely a bad place to start.

Acknowledgement: This summary was generated using AI writing tools as part of my experimental learning process. For more information, check out the original articles listed below.

Related Resource Links