JD cuts inference latency as Tencent open-sources translation model
China's AI compute updates spanned domestic chip supply, inference optimization, model services and plans for optical and satellite infrastructure.
• JD Ads and Huawei Kunpeng reduced end-to-end inference latency by 15% through interconnect, shared-memory and communications changes, validated on 5% of traffic in core use cases.
• Cambricon Chairman Chen Tianshi denied that the company had communicated a target of more than 1 million chip shipments next year.
• CXMT reportedly turned down Apple, prioritizing domestic demand for memory chips.
• Tencent open-sourced Hy-SimulMT for real-time Chinese–English speech translation, reporting a three-second response time.
• Kimi apologized for K3.1 Pro rate limits and unnecessary quota consumption, announcing a one-time subscriber reset and compensation equivalent to 25% of monthly allowances.
• Optowide said it is shipping passive optical components for 800G and 1.6T optical modules in volume.
• Chinese researchers completed a real-time bidirectional multicore-fiber transmission test at 3.2 Pb/s, demonstrating potential applications in links between computing hubs and backbone networks.
• Shanghai's 2027–2030 satellite internet plan targets peak AI computing capacity exceeding 30 POPS per satellite and support for space-based inference.