Asia wrap: Seoul strengthens HBM and packaging push as China adds more filed AI services
Kai Tanaka ·
South Korea moved this week to reinforce investment in high bandwidth memory and advanced packaging, according to government statements and local industry disclosures, as GPU vendors lock in next year’s supply. In China, the Cyberspace Administration added another batch of generative AI services to its public filing list, with Alibaba Cloud, Baidu, ByteDance, and DeepSeek among those expanding enterprise offerings. Both developments shape procurement and compliance decisions for global teams planning AI deployments into 2027.
HBM remains the limiting factor for AI server builds, and packaging is the main cost driver for accelerators, so policy clarity affects pricing and delivery schedules. China’s filing regime continues to steer local platforms toward domestically hosted models, which matters for vendors that integrate across markets under data and export constraints.
South Korea leans further into HBM and packaging
Seoul’s trade and finance authorities outlined permits, infrastructure support, and tax credits tied to mega cluster projects near Yongin and Pyeongtaek, according to notices circulated to industry groups. Samsung Electronics and SK Hynix are pushing HBM3E ramps and pilot HBM4 development along with more CoWoS class capacity in partnership with domestic OSAT providers.
SK Hynix has told investors that HBM lines are booked out into next year, while Samsung’s semiconductor division has emphasized yield gains and customer qualifications in recent briefings. Nvidia’s reliance on high layer HBM and advanced packaging keeps both Korean suppliers central to the GPU roadmap through 2026.
China expands filed generative AI services
Beijing’s filing list for generative AI services grew again this week, reflecting a shift by large platforms toward enterprise focused versions of their models. Alibaba Cloud’s Qwen family, Baidu’s Ernie, ByteDance’s Doubao, and DeepSeek’s research models dominate new entries, with filings required for public availability under CAC rules.
Cloud buyers in China are emphasizing inference cost per token, which has pushed adoption of open weight Qwen variants and smaller context models hosted on domestic accelerators from Huawei. Cross border vendors face separate security reviews, and access to restricted Nvidia parts remains limited under United States export licensing.
For global technology leaders, the signal is to plan for a constrained HBM environment even as packaging scales, and to segment AI stacks by jurisdiction. Vendor roadmaps that marry Korean memory and packaging with model portfolios that satisfy China’s filing regime will see fewer surprises in cost and compliance over the next four quarters.