Chinese Tech Daily — 2026-07-08#

Top Story#

DeepSeek has reportedly initiated the development of its own AI inference chips, a project that began roughly a year ago. Unlike training accelerators, this bespoke silicon is designed specifically to handle user requests and generative workloads, aiming to reduce DeepSeek’s reliance on hardware from Nvidia and Huawei. By controlling the hardware layer, DeepSeek hopes to co-design its chips with its model architectures—such as its Mixture of Experts (MoE) structure—to drastically lower inference costs as its global token volume continues to scale.