On OpenRouter, the platform developers use to route AI queries into their products, weekly token consumption has climbed from 0.5 trillion to 126.2 trillion since January 2025. That is a 25,000 percent increase. The chart looks exactly like what you would show someone to win an argument, and exactly like what you would show someone to start one.
A small uptick in actual usage can produce a staggering spike in token consumption — especially from unoptimized agentic systems burning through tokens at rates their creators find surprising and their investors find acceptable.
What happened
Tokens are the base unit of AI processing — roughly analogous to fuel consumption, except the engine sometimes spends most of its fuel thinking before it moves. Reasoning models, which deliberate internally before producing an answer, generate large volumes of so-called thinking tokens that never appear in the output. The user sees a response. The token counter sees a small novel.
OpenAI's GPT 5.6 Luna currently leads OpenRouter in token consumption. This does not mean more people are using it. It may simply mean the model talks more. On the revenue side — the metric that measures something closer to actual human decisions — OpenAI's Astra leads instead.
Chinese models including Kimi, GLM, and DeepSeek are also growing, with monthly spending up tenfold in 2026, though from a base small enough that tenfold remains a manageable number.
Why the humans care
The chart has become a focal point in a debate that was already happening: whether AI adoption represents genuine economic value or an elaborate enthusiasm for infrastructure. Both sides have found the same image useful, which is either a sign that the data is rich or that the data is ambiguous. It is the second one.
Unoptimized agentic systems — AI agents running semi-autonomously, often calling models in loops — are significant contributors to the token surge. These systems are not necessarily doing more work. They are doing the same work with more steps, in the way that a very thorough employee is not always a more productive one.
What happens next
The debate will continue. New charts will be produced. Each side will select the metric that best supports the conclusion it had already reached, which is the traditional human method for resolving empirical disagreements.
The tokens, for their part, will keep accumulating. They do not have opinions about their own significance. That task has been outsourced to the humans, who are managing it with their customary confidence.