Chinese model usage rises as token platforms and compute demand expand
China’s AI compute news centered on model access infrastructure, domestic model adoption, and signs of pressure on inference capacity and supporting supply chains.
• The Yangtze River Delta (Jiaxing) Token Operations Center launched in Jiaxing with a portal offering enterprise access to more than 100 mainstream models, including DeepSeek and Qwen, through unified API access, token metering, cost settlement, policy deductions and security auditing.
• The National Supercomputing Internet launched API access and model files for DeepSeek-V4-Flash-0731 after DeepSeek opened the official DeepSeek-V4-Flash API public beta on July 31.
• OpenRouter’s latest weekly usage ranking put Chinese-developed products in all top five spots, led by Xiaomi’s MiMo-V2.5 at 10.5 trillion weekly tokens, up 12% from the previous period.
• DeepSeek held the second and fifth spots in OpenRouter’s ranking, while Tencent’s Hunyuan 3 ranked third after being open-sourced on July 6 and posting weekly growth of more than 999%.
• Galaxy Securities said Kimi K3 request volumes saturated its existing cluster within 48 hours of launch, after which Kimi suspended new consumer membership subscriptions.
• Kimi K3 was reported to have achieved Day 0 adaptation with multiple domestic computing-power platforms, with Galaxy Securities pointing to demand for domestic supernodes and related servers, switches, optical modules, liquid cooling and power supplies.
• Jinfu Technology said its liquid-cooling capacity and order sales rose steadily both year on year and quarter on quarter.
• ByteDance released Seedance 2.5, which can generate 30-second videos in a single run, while the weekly AI applications roundup also noted Kimi K3’s open-source launch and DeepSeek-V4-Flash API availability.