Chinese clouds tilt to Huawei Ascend as Korean HBM supply stays tight into 2026

China’s largest cloud providers are centering new AI capacity on domestic accelerators, led by Huawei Ascend, because United States export controls keep Nvidia’s latest Blackwell class GPUs out of the mainland. At the same time in South Korea, tight supply of high bandwidth memory from SK Hynix and Samsung remains the key bottleneck that will guide AI server shipments into 2026.

The immediate driver is policy. The US Commerce Department’s Bureau of Industry and Security expanded chip restrictions in 2023 and 2024, closing prior carve outs for downgraded Nvidia parts, and Chinese platforms need predictable compute for model training and inference.

China AI compute shifts to domestic silicon

Alibaba Cloud, Baidu, Tencent, and ByteDance have been rolling out services tuned for Ascend based clusters, along with software stacks such as Huawei’s MindSpore and CANN, and Baidu’s PaddlePaddle. That reduces reliance on CUDA and shifts developer tooling, which affects model portability and ecosystem choices for multinational teams operating in China.

Beijing continues to align funding and compliance. The Cyberspace Administration of China’s generative AI rules require security assessments for public model releases, and the Ministry of Industry and Information Technology has pushed compute infrastructure build outs through provincial projects. The mix favors domestic accelerators for government procurement and state linked workloads.

Cloud workloads on Ascend are paired with open source frameworks that interoperate with PyTorch, but performance parity varies by model and operator. That matters for global developers who must support diverging back ends across regions to keep latency, cost, and compliance in balance.

Korean HBM supply sets AI server pace

SK Hynix holds more than half of the HBM market by shipments, with HBM3E qualifications anchoring allocations to Nvidia and other GPU vendors. Samsung is ramping its own HBM3E output and is working with customers on reliability and power targets to win a larger share in 2026 build plans.

Korea’s Ministry of Trade, Industry and Energy continues to back capacity through tax credits under the K Chips framework and through support for advanced packaging clusters. Packaging remains a choke point since GPU makers rely on limited outsourced capacity for 2.5D and 3D integration, including TSMC’s CoWoS lines for many high end accelerators.

For global technology teams, the weekend takeaway is practical. Plan for China specific AI stacks that prioritize Ascend compatibility and budget for HBM constrained server lead times that push delivery schedules and pricing across regions, since memory output, not only GPU tape outs, will set the cadence of AI infrastructure through next year.