Samsung’s earnings preview and TSMC’s June revenue keep AI hardware tight into H2, with China leaning harder on domestic chips
Kai Tanaka ·
Samsung Electronics signaled another strong quarter for memory in its preliminary results this week in Seoul, while Taiwan Semiconductor Manufacturing Co reported solid June revenue in Taipei. Together, the disclosures point to continued tight supply in high bandwidth memory and advanced packaging that underpin Nvidia and AMD accelerators, shaping deployment timelines for hyperscalers into the second half.
Why it matters now, procurement teams are locking second half allocations for HBM and TSMC’s CoWoS packaging, and this week’s company updates offered little evidence of spare capacity. That keeps delivery risk elevated for GPU clusters and AI infrastructure budgets in the United States, Europe, and Asia.
Korean memory and Taiwan packaging
Samsung’s preview highlighted improved profitability in memory, supported by premium DRAM and HBM shipments, according to the company’s filing. SK Hynix remains the leading HBM3 supplier to Nvidia, and both Korean firms are adding capacity, but supply remains gated by through silicon via processing and substrate availability that take time to ramp.
TSMC’s June sales extended a run of strong monthly prints tied to AI demand, based on the company’s revenue release. The foundry has expanded CoWoS throughput in Jhunan and is adding capacity in southern Taiwan with support from substrate partners such as Unimicron and Nan Ya PCB, yet industry orders for advanced packaging remain booked well ahead of lead times.
The packaging constraint sits alongside a maturing 5 nanometer and 3 nanometer node mix that continues to benefit AI accelerators and networking silicon. Suppliers in Taiwan’s backend ecosystem, including ASE and SPIL, report sustained utilization driven by AI and high end networking, according to recent disclosures and local trade press.
China’s AI compute split
In China, developers are scaling training and inference on domestic accelerators as US export controls limit access to the latest Nvidia parts. Alibaba Cloud, Baidu, ByteDance, and Tencent have expanded support for Huawei’s Ascend 910 series across public cloud instances and enterprise solutions, reflected in service documentation and recent procurement notices by state owned buyers.
Model providers continue to register systems with the Cyberspace Administration of China’s filing regime, with many listings indicating compatibility with Ascend for inference. Training remains a mix of domestic hardware and older export compliant Nvidia parts, which slows cycle times compared with deployments outside China.
For global technology teams, the week’s datapoints reinforce two planning assumptions, AI rollouts will be constrained by HBM and advanced packaging availability into the second half, and software stacks will continue to bifurcate between Nvidia based infrastructure abroad and Ascend based environments inside China.