TechGlobal.news

Korea chip exports strengthen on HBM as China moves to tighten AI model filings

Kai Tanaka ·

South Korea’s chip sector ended the week with firmer signals after customs data released Tuesday showed continued year over year growth in semiconductor exports led by high bandwidth memory, while in Beijing the Cyberspace Administration of China circulated draft updates to its generative AI model filing regime. The twin developments matter for global teams planning AI server procurement and compliance for China rollouts in the fourth quarter.

Korea’s export print extends a months long rebound anchored by HBM shipments that feed Nvidia and other accelerator platforms, reinforcing expectations of tight supply through year end. The data underscore how memory pricing and AI server build outs, rather than smartphones or PCs, remain the primary driver of Asia’s chip cycle into 2027.

South Korea’s HBM signal

SK Hynix and Samsung Electronics have both flagged continued expansion of HBM3E output and advanced packaging in Cheongju and Giheung, with additional outsourcing to partners to clear backlogs. Industry checks around advanced packaging, including CoWoS capacity at foundries in Taiwan and new lines in Japan, still indicate constrained availability that could stagger deliveries for some cloud customers.

Brokerage notes in Seoul this week highlighted allocation discipline for HBM tied to long term agreements with Nvidia and AMD, limiting spot availability for second tier server makers. Any incremental supply relief now depends on packaging throughput more than wafer starts, a bottleneck that has shaped bill of materials and deployment timelines across US and Chinese data centers.

China’s filing push on big models

The CAC’s draft, circulated for public comment according to state media, would raise disclosure for model registrants, including compute used for pre training and fine tuning, data source provenance, and benchmark evaluation methods. It would also clarify when internal or enterprise only systems require filing and reinforce security reviews for models embedded in public facing services.

The update targets the current cohort of Chinese model providers, notably Baidu, Alibaba Cloud, ByteDance, Tencent, Zhipu AI, Baichuan, 01.AI, and DeepSeek, that already submit filings before public release. Multinationals offering AI features in China through partners would need to align on documentation and content controls, with cloud providers adjusting fine tuning workflows to local rules.

For global technology teams, the message is to lock in HBM backed server capacity early and to budget around packaging constraints, while treating China deployments as a regulatory program with specific disclosures and testing artifacts. The combination sets the cadence for AI infrastructure through the next two quarters, linking hardware availability in Korea and Taiwan to compliance gating in China.