Asia-Pacific wrap: China leans on domestic AI chips as US curbs persist, Korea channels capex to HBM

Asia-Pacific tech markets ended the week focused on China’s constrained AI compute and Korea’s expanding memory supply, issues that set the pace for global AI deployment. In China, major platforms continued to retool for domestic accelerators as tightened United States export controls keep Nvidia-class GPUs scarce. In Korea, Samsung Electronics and SK Hynix kept steering capital and engineering toward high bandwidth memory and advanced packaging.

The stakes are immediate for hyperscalers and enterprise buyers, because training and inference costs pivot on available compute and HBM throughput. Any change in Chinese demand composition or Korean output resets accelerator procurement and delivery schedules in the United States and Europe.

China’s AI compute workarounds Baidu, Alibaba Cloud, Tencent, and ByteDance have been adapting software stacks to Huawei’s Ascend NPUs and to in-house and local options such as Baidu’s Kunlun, aiming to keep model training and inference onshore. Baidu’s PaddlePaddle and Alibaba’s ModelScope have emphasized heterogeneous hardware support, including Ascend back ends and mixed CPU plus NPU pipelines, to mitigate the risk of single-vendor lock-in.

Regulatory gating remains in place, with content and safety rules enforced through China’s model registration regime under the Cyberspace Administration. That keeps large foundation models from Baidu, Zhipu AI, Baichuan, and SenseTime focused on enterprise and compliant consumer use, which in turn channels demand toward inference clusters built on domestic accelerators.

Korea’s HBM and packaging push SK Hynix remains a key supplier of HBM3 and HBM3E for AI accelerators, while Samsung is working to win broader qualifications with top GPU vendors. Both are investing in through-silicon via and hybrid bonding lines, along with advanced packaging that links HBM stacks to compute dies, as they execute on government backed semiconductor cluster plans around Yongin and Pyeongtaek.

Downstream constraints are still in focus, including substrate availability and advanced packaging bottlenecks that affect end-to-end throughput. Japanese suppliers such as Tokyo Electron, Disco, and Ajinomoto continue to factor into lead times for tools and build-up films used across HBM and AI server manufacturing.

For global technology teams, the weekend takeaway is practical. Treat model portability across non Nvidia hardware as a first-class requirement in China-facing products, and assume HBM availability, not just GPU counts, will drive delivery timelines for AI services. Procurement strategies that bundle accelerators with guaranteed HBM and packaging slots, and software stacks engineered for Ascend and other NPUs where required, will have an execution edge.