Asia tech week: HBM supply still the hinge as China’s enterprise AI shifts to domestic chips
Kai Tanaka ·
SK Hynix and Samsung remained the focal point for high bandwidth memory supply into the weekend, while Baidu, Alibaba Cloud, and ByteDance advanced enterprise AI deployments tailored for domestically available accelerators in China. The twin pressures, packaging bottlenecks and tightened United States export controls on Nvidia class parts, continue to set global planning assumptions for model training and inference.
Why it matters now is straightforward. HBM and advanced packaging such as TSMC’s CoWoS remain the scarce links that gate AI server output through 2026, and China’s shift to Huawei Ascend and in house silicon is changing the demand mix for memory, networking, and software stacks.
Korean memory and packaging set the pace
SK Hynix has led HBM shipments and is ramping HBM3E output, including at its Cheongju facilities, according to company disclosures. Samsung Electronics memory chief Lee Jung bae has flagged increased HBM and packaging capacity tied to customer qualifications, aligning with demand from Nvidia and other accelerator vendors.
Packaging capacity is still tight. TSMC told investors it would expand CoWoS capacity materially in 2024, and supply remains closely managed alongside substrate and test constraints. Korea’s industry ministry, MOTIE, is backing long horizon investments in the K semiconductor cluster around Yongin to secure power, water, and supplier depth, which are prerequisites for sustained HBM and logic growth.
China’s enterprise AI tilts to domestic accelerators
China’s Cyberspace Administration requires model filings, which has pushed large platforms to focus on compliant, enterprise ready stacks such as Baidu’s Ernie, Alibaba Cloud’s Qwen, and ByteDance’s Doubao. With United States rules curbing shipments of prior China specific Nvidia parts, Chinese buyers have shifted deployments toward Huawei’s Ascend line and in house options such as Baidu’s Kunlun for targeted workloads.
That shift is rippling through the bill of materials. Software teams are optimizing inference for Ascend and for domestic interconnects, and buyers are recalibrating memory bandwidth and capacity needs per accelerator. Cloud and on premises deployments are proceeding with virtual private cloud and on site options to meet data governance requirements.
For global tech teams, the takeaway is operational. AI roadmaps will live or die on access to HBM and advanced packaging in Korea and Taiwan, and on compatibility with the silicon that is actually shipping in China. Procurement, model architectures, and latency budgets need to reflect those realities now, not at the next hardware cycle.