The global LLM usage leaderboard has refreshed, and the gap at the top keeps widening. Cailianshe reported on November 9, drawing on OpenRouter's latest numbers, that total global calls to large AI models hit 210 trillion tokens for the week of November 2-8, up 8.25% week over week. Chinese models accounted for 116.18 trillion tokens of that, a 16.71% jump. US models logged 13.98 trillion, down 32.4%. That makes it 28 consecutive weeks of Chinese LLM weekly traffic ahead of the US.

The chart's upper ranks keep concentrating in Chinese hands. Four of the global top five are Chinese models. DeepSeek V4.1 Flash 1017 held No. 1 for a third straight week on 56 trillion tokens, up 17%. Tencent's Hunyuan Hy4, released November 4, debuted straight into fourth place. Tencent told Securities Times the Hy family has entered the ranks of the strongest open-source models, with weights to be open-sourced soon.

Illustration of a global LLM usage leaderboard
The weekly token leaderboard: usage traffic as a public scoreboard

Traffic is surging, and inference-side economics are finally coming into focus. DeepInfer, DeepSeek's open-source inference framework, discloses $38 of revenue per GPU-hour on production clusters, per Zhejiang Securities data cited by 21st Century Business Herald. Deduct GPU rent and the inference gross margin lands at about 82%. The picture isn't uniformly rosy, though. The SiliconData token price index has slid to $0.94 per million tokens, and B300 rental rates are up 51.9% since May, as next-generation GPU supply still trails demand.

Public markets voted earlier than most. Mainland mutual funds added to large-model, server and software leaders in Q3 2026 while trimming compute-leasing positions, according to China Merchants Securities. The top ten computer-sector holdings by added value came to 7.871 billion yuan, with Zhipu sitting at the top. Zhejiang Securities argues enterprise AI competition now hinges on delivery and efficiency rather than raw capability. Anthropic's month-12 customer retention stands at 87%, and expanding existing customers into more tasks and teams has become the main growth engine.

One caveat deserves mention. The leaderboard tracks API traffic on OpenRouter alone, so private deployments and on-device usage don't show up, and it diverges from true market share. Still, 28 straight weeks pointing the same direction is a stable observation window. Token traffic is turning into the public scoreboard of Chinese LLM competitiveness.