Chinese open-weight models are accelerating their push into global markets as commercialization matures. OpenRouter data cited by Southern Metropolis Daily shows that in the week of August 24-30, among the top nine models by Token consumption, China and the US held six and three seats respectively, with combined Token volumes of 55.16 trillion and 17.07 trillion - and China's lead has widened since May-June.
The shift is driven by price-performance. UBS analysts note that as frontier-model costs climb, overseas users have grown ROI-conscious; China's models combine systematic price advantage with rapidly closing capability gaps. On Artificial Analysis, Kimi K3, Zhipu GLM-5.3, Alibaba Qwen 3.8 and DeepSeek V4 Pro sit just behind the top closed US models, yet Kimi K3 costs about $0.84 per task versus $3.69 for Claude Fable 5.1 - roughly a 4.4x gap.
Foreign firms are taking notice: US food-delivery giant DoorDash, Airbnb and Germany's Siemens have begun shifting back-office AI to Chinese models. Moonshot is negotiating revenue-share deals with Microsoft, Amazon and Google to host Kimi K3, seeking a cut as high as 30%, while Zhipu has started distribution cooperation with AWS and is talking with two large Western cloud providers.
To balance open release with monetization, vendors are tightening licences: Qwen 3.8-Max and Kimi K3 require separate permission for MaaS or AI-assistant operators above certain revenue thresholds. Per Xinhua, overseas AI firms including Harvey (built on Kimi K3) and Thinking Machines (using DeepSeek-V3 routes) now treat Chinese open-weight models as a base for their own products - turning China's models from exported products into exported technology.
Image: Efficiency of AI-related computer chips. Source: Wikimedia Commons (CC BY-SA 4.0).
