TechGlobal.news

Korea’s HBM capacity keeps AI servers tight, China accelerates shift to Huawei chips

Kai Tanaka ·

South Korea’s SK Hynix and Samsung stayed at the center of the AI hardware story this week, with guidance and supplier chatter pointing to tight high-bandwidth memory supply into 2027. At the same time, China’s largest AI platforms increased deployments on Huawei’s Ascend processors and in-house silicon, extending a split in compute stacks that global teams must navigate.

The linkage matters for cost and timelines. Nvidia and AMD pricing remains sensitive to HBM availability, and Chinese model builders are optimizing away from U.S. GPUs, which affects software frameworks, toolchains, and cross-border interoperability.

Korea’s HBM bottleneck

SK Hynix remains Nvidia’s primary HBM partner, and Samsung is pushing node conversions and through silicon via capacity in Cheonan and Pyeongtaek to close the gap. Market researcher TrendForce estimated in 2024 that SK Hynix held more than half of the HBM market, with Samsung and Micron splitting the remainder, a concentration that still shapes GPU supply.

Samsung’s memory chief Lee Jung-bae has flagged HBM3E ramp and hybrid bonding as priorities, while SK Hynix CEO Kwak Noh-Jung has focused on yield discipline to protect margins. The transition to HBM4, needed for next generation accelerators, adds tooling and materials constraints that limit near term relief even as both firms add cleanroom space.

Server builders say qualification cycles for new HBM stacks remain lengthy, which keeps delivery risk elevated for 2026 capacity plans. That sustains bargaining power for Korean suppliers and leaves hyperscalers and OEMs managing allocation alongside Nvidia and AMD roadmaps.

China leans into domestic accelerators

Baidu, Alibaba, and ByteDance have expanded training and inference on Huawei’s Ascend 910B and on their own chips, reinforcing a pivot that began after tighter U.S. export rules. Baidu’s Ernie services run across Baidu Kunlun and Ascend, Alibaba Cloud’s Qwen models are optimized for Hanguang and Ascend, and ByteDance’s Doubao stack has added Ascend support to diversify compute.

The Cyberspace Administration of China continues to require filings for publicly accessible generative services, and procurement guidance from MIIT and other agencies favors domestic hardware in state linked deployments. That is pulling more developers into Huawei’s MindSpore and Baidu’s PaddlePaddle, alongside CUDA alternatives, which complicates cross border MLOps for multinationals.

For global technology teams, the takeaway is bifurcation with bottlenecks. AI infrastructure is splitting by geography and vendor stack, and memory supply remains the limiting factor for GPU throughput, so capacity decisions in Seoul and procurement choices in Beijing will keep setting the pace into 2027.