AI usage is measured in “tokens”, which are small chunks of text that AI systems use to understand, analyse and generate language. On average, one token is roughly equivalent to three-quarters of an English word.

There have been astonishing increases in AI use in 2026, and the increase in token use has vastly outweighed the declining “per token cost”.

The world’s top companies that are using AI had deployed “token leaderboards” to try to get their employees to create as much gain in productivity as possible from the most advanced models.

Trillions, and sometimes thousands of trillions of tokens (quadrillions) have been used over the past few months primarily for “agentic use”, which is for agents that can perform tasks automatically. The problem here is that incredible bills were racked up, so much so that many of those companies have now begun to ration the use of these models.

This is important. It might show that there are limits to how much work could be automated. The virtual workers might be more expensive than the human ones. It depends on the task.

One factor to watch out for however, is that many companies, including Western ones, are diverting to much cheaper forms of AI, derived from Chinese models that are provided freely on to the market.

So there is much uncertainty here, but some clear trends emerging.