Chinese Models Now Process More Tokens Than American Ones. The Price War Has Begun.

140 trillion. China’s daily AI token usage, March 2026. Up from 100 billion in early 2024. 1,000-fold in two years. On OpenRouter: 61% of top-ten token consumption is Chinese models.

I sat through HumanX panels on “open vs. closed models” without once hearing that the Chinese open-source ecosystem already won the volume game.

The Xiaomi Ghost

Mid-March: anonymous model appeared. Industry consensus: DeepSeek V4. It was Xiaomi — a phone and EV company. When a consumer electronics manufacturer produces frontier-competitive models that fool the industry, the model layer is commodity.

When Chinese models run at 1/6 to 1/4 the cost, it propagates through the stack. Different utilization economics. Different infrastructure requirements. US export controls forced efficiency innovations on constrained hardware — primarily Huawei Ascend accelerators — producing models that are more commercially competitive, not less.

The controls may have accelerated exactly what they were designed to prevent. And the workforce implications: fine-tuned open-source models on a single B300 or Vera Rubin GPU for a few thousand dollars a month create AI doppelgangers making elite knowledge workers 10x productive. Chinese-style efficiency, not American-style frontier pricing, is where mass adoption points.

China’s 15th Five-Year Plan: AI as foundational priority. $1.4 trillion target by 2030. 70% sector adoption by 2027. 700+ registered AI services. This is state-directed industrial policy at unprecedented scale.

While I was at HumanX, I took the chance of Visiting Zuck’s house in SF, where he created Facebook.