Korea’s HBM choke point still defines AI hardware, while China’s model makers lean into domestic chips

The most consequential Asia Pacific tech story this week was unchanged, Korean high bandwidth memory supply remains the tightest link in the AI hardware chain. SK Hynix and Samsung set the pace for HBM3E output that feeds Nvidia class systems, while TSMC’s CoWoS packaging limits continue to cap how fast those parts reach data centers.

The focus matters because AI compute procurement for 2026 delivery depends on these three vendors. TrendForce has estimated SK Hynix held a majority share of the HBM market in 2023, with Samsung pushing to close the gap as qualifications expand, and Micron targeting niche wins.

Korean press and suppliers have kept attention on two transitions, HBM3E volumes through 2025 and early HBM4 planning tied to next generation AI platforms. Nvidia’s Blackwell architecture uses HBM3E and relies on TSMC for advanced packaging, so any slippage in memory stacking or CoWoS module output constrains system shipments.

SK Hynix chief executive Kwak Noh Jung has emphasized prioritizing high margin HBM, while Samsung’s memory head Lee Jung Bae has pledged to grow share through higher layer stacks and better power efficiency. Both makers are investing in additional cleanroom space in Icheon, Cheongju, Hwaseong, and Pyeongtaek, with suppliers reporting longer lead times for through silicon via tools and metrology.

In Japan, TSMC’s Japan Advanced Semiconductor Manufacturing unit in Kumamoto is ramping mature logic and specialty processes that support AI adjacent components such as power management and image sensors. TSMC has also been expanding CoWoS capacity in Taiwan, and any spillover packaging activity to Japan would ease bottlenecks that affect Nvidia and other accelerator vendors.

In China, model providers Baidu, Alibaba Cloud, and ByteDance continue to ship services tuned for domestically available accelerators from Huawei’s Ascend line, according to product notes and developer documentation. The Cyberspace Administration of China still requires filing for public model releases, and U.S. controls restrict access to Nvidia’s highest end GPUs, which keeps inference efficiency and quantization a priority for Chinese platforms.

For global teams planning spend into the September quarter, the through line is clear. AI model roadmaps are increasingly gated by Korean HBM and TSMC packaging, and China facing firms are optimizing for non Nvidia silicon. Procurement, model deployment choices, and regional data center siting will reflect those constraints rather than headline model benchmarks.