25 Times Is Not 25 Percent

Cathie Wood put the number bluntly: the amount of text processed by artificial intelligence grew 25-fold in one year. Many people heard 25 percent. Those are not remotely the same thing. Moving from 100 to 125 marks a strong year. Moving from 100 to 2,500 changes the scale entirely. That gap is the white noise surrounding this story.

A token is not a cryptocurrency. It is a small piece of text. Models break prompts and responses into these fragments, with a word often taking one token or slightly more. A short answer can consume hundreds. An agent that reads a contract, compares four reports and prepares a memo can use hundreds of thousands. Tokens have become the unit for measuring, billing and scaling machine work.

Wood's figure comes from a single gateway, OpenRouter, which carries traffic from hundreds of language models. It does not capture the entire market, since Google, OpenAI, xAI and Anthropic handle large volumes independently. But it offers a visible stream. The striking change is not simply that people write more prompts. Machines now assign work to one another.

Using OpenRouter data, Peter Walker identified Feb. 6, 2026, as the last day when people used more tokens than agents. Since then, agent consumption has risen about 14 times, from half a trillion to 7.3 trillion tokens a week. Human use has grown less than threefold. An agent does not get tired. It opens another session, calls another model and keeps working after the user closes the computer. Each cycle adds to a total that can reproduce itself.

There is an important qualification: about 70% of agent tokens come from cached prompts, which cost much less. The bill therefore does not rise as quickly as the raw count. But that does not deflate the story. It makes the expansion easier to sustain. Brett Winton linked the volume to money, arguing that if intelligence production scales this quickly, data centers could afford more expensive capital and still earn returns. In Mexico, the effect could appear both at work and in finance, as companies begin to compete with computing infrastructure that investors may treat as a lower-risk asset.

“

We're going to have these models at a company be outputting more tokens per day than all of humanity put together, and then 10 times that, and then 100 times that.