China expands AI compute investment, models and chip supply
China's AI compute sector saw new model releases, chip production and investment, alongside local financing support and data-center construction activity.
• Alibaba open-sourced its Qwen3.8-27B model, a natively multimodal model supporting a 262K-token context window that can be extended to 1 million tokens through YaRN technology.
• Li Auto is reportedly exploring an in-house cloud inference chip based on the dataflow architecture of its autonomous-driving chip, although the project remains at an early stage.
• SiEngine said its self-developed 7nm automotive-grade Tiangong 100 AI accelerator has entered full mass production and is shipping in volume to automakers, Tier 1 suppliers and other partners.
• SMIC reported second-quarter revenue of $3.006 billion, up 36.1% year on year, and net profit attributable to shareholders of $479 million, up 261.7%; its co-CEO said average wafer selling prices rose 5.7% sequentially.
• Tencent sharply increased quarterly computing-power capital expenditure to a record 52.8 billion yuan, according to a report describing the spending as driven by a cautious approach to capacity.
• Guangzhou's Haizhu district introduced eight token-economy measures, including support for tokenized heterogeneous computing power and funding of up to 1.5 million yuan per eligible company annually, while a related Computing Power Token Loan offered early trial credit lines totaling 28 million yuan.
• A subsidiary of Shaanxi Construction Engineering won a 1.156 billion yuan EPC contract for the Yunjing AI Data Center in Jingtai County, Gansu province.
• Zhongji Innolight agreed to spend more than 1.747 billion yuan for a 10.47% stake in thermal-management company Jones Tech, citing rising power density as its 800G and 1.6T optical modules ramp up.