AINEWS
Samsung shipped roughly $12.7B more chips to China than to the US in H1 2026, and the Xi'an memory plant at the center of that flow runs on a US permit. That permit expires December 31, 2026.
Samsung shipped roughly $12.7 billion more chips to China than to the United States in the first half of 2026, and almost none of that gap exists without a single US government annual export license scheduled to expire December 31, 2026. The license covers one factory in Xi'an that builds the high-bandwidth memory, or HBM, the stacked DRAM chips that sit next to AI processors and move data fast enough to train and run large models. Without that plant, the China line in Samsung's books gets smaller (TechTimes, Aug 15, 2026).
HBM is the part that holds the numbers up. Samsung's sixth-generation version, HBM4, entered mass production in February 2026 under the JEDEC JESD270-4 standard. It doubles the per-stack interface width of the prior HBM3E generation to 2,048 bits and pushes throughput above 2 terabytes per second per stack. Nvidia's Vera Rubin GPU, the next US-frontier AI accelerator, pairs eight of those stacks per chip, which is why a single Korean memory plant now sits on the critical path for AI hardware on both sides of the Pacific (BusinessKorea, Aug 2026).
The H1 2026 split, drawn from Samsung's semiannual report filed on August 14, puts China at ₩88.6004 trillion (~$62.5 billion) and the US at ₩70.6466 trillion (~$49.8 billion), the first time the China line has run materially above the US line in the company's reporting. The US figure itself more than doubled from ₩33.4759 trillion (~$23.6 billion) a year earlier, so Samsung is not redirecting capacity away from America. It is adding China-bound memory output on top of US growth, not swapping one for the other (BusinessKorea; TechTimes).
The arrangement holds for three reasons. The Xi'an fab operates under a US annual export license granted to Samsung in late 2025, alongside parallel approvals for SK Hynix and TSMC, replacing the previous indefinite "validated end user" regime with a year-by-year renewal (Reuters, Dec 30, 2025). The US has now generalized that approach, and China fab oversight now runs on annual tool licences, not standing authorizations (Astute Group). Samsung's H1 R&D and capital spending crossed ₩55 trillion (~$38.8 billion), an order of magnitude that gives the company room to keep scaling Xi'an while it builds new lines in Korea and the US, and that scale is what makes the December decision consequential rather than symbolic.
Each plausible outcome of the December 31 decision does something different to the chip flow. A clean renewal leaves the Xi'an-to-China corridor intact, the HBM4 ramp continues, and the H1 2026 split becomes the new baseline that both Washington and Beijing plan around. A narrowed renewal, such as a cap on HBM4 stacks shipped from Xi'an or a quota tied to Vera Rubin-bound units, would slow the China line, push Chinese AI customers toward SK Hynix or domestic alternatives, and reshape the second-half numbers without breaking the underlying technology. A non-renewal would force Samsung to re-route Xi'an's output or absorb a write-down on the line, and would hand China a near-term memory-supply gap at exactly the moment its AI accelerator buildout is leaning hardest on advanced HBM.
The number to watch is not the H1 export total. It is whether the US Commerce Department treats the December 31 deadline as a routine renewal, a renegotiation point, or a pressure lever, and what conditions, if any, it attaches to the next twelve months of Xi'an's output. The same permit, in the same building, will keep both the US's Vera Rubin supply and China's competing AI stack running until at least the end of 2026. After that, the decision is political, not technical.