One Generation Behind, on Purpose: Samsung and SK Hynix Upgrade Their China Fabs as the Licenses Expire Every December

Reuters reported on Oct 21, 2026 that the upgrades are already underway and that waiting for Washington has become expensive. Six days earlier, Solidigm — SK hynix's NAND unit — asked US authorities to approve sending more American chipmaking equipment to Dalian as a condition of continuing its US fab build-out. At the same time SK hynix said it was not seeking any capacity expansion at Dalian Mod B, the unused 2023 building next to the live 192-layer line. Both statements can be true: the request is for tools, the denial is about wafers. That gap is where the whole regime now lives.

On Sept 2, 2025, the Commerce Department's Bureau of Industry and Security struck Intel Semiconductor (Dalian), Samsung China Semiconductor and SK hynix's Chinese entities off the validated end-user list, ending the standing authorisations that had let them ship US tools without case-by-case approval. In their place came annual licences for 2026, granted on Dec 30, 2025, covering Samsung's Xi'an NAND fab and SK hynix's Wuxi DRAM and Dalian NAND plants. Those licences permit operation and maintenance of existing capacity; EUV was never available. What they do not settle is the following year. SK hynix's one-year Dalian authorisation expires on Dec 31, 2026, and as of mid-October 2026 Washington had not issued a 2027 authorisation.

The arithmetic explains the discipline. Roughly 40–45% of SK hynix's NAND output and 30–35% of Samsung's comes from China, and the Wuxi DRAM fab alone supplies nearly half of SK hynix's DRAM capacity. Those are not marginal lines that can be idled in a shortage — they are part of why the two groups can post record profit at all. Samsung's Device Solutions division earned ₩89.2 trillion of operating profit in the second quarter, and memory revenue and profit were an all-time high. Running the China line one generation behind is the cheapest way to stay inside the licence while keeping the line good enough to defend share against YMTC and CXMT.

This is where the decoupling thesis breaks. Applied Materials shipped $8.5 billion into China in fiscal 2025, 30.1% of revenue; Lam Research shipped $7.9 billion, 33.8% of fiscal 2026 revenue; KLA shipped $4.05 billion, 29.8%. None of those numbers fell. KLA's China revenue was flat year on year — its China share of revenue dropped from 42.8% to 29.8% because Korea grew 26.2% on HBM and advanced DRAM investment, not because China stopped buying. Export controls have not closed China's tap; they have put it on an annual meter.

The intuitive read — China builds memory, China takes the pricing power back — is backwards for now. China is short high-bandwidth memory, and short memory is expensive memory: on Sept 10, 2026, Reuters reported Huawei raised the indicated price of its Ascend 950DT accelerator card above 250,000 yuan, a 20% increase, because HBM supply was tightening. Chinese accelerator vendors are repricing because the memory underneath them is scarce. Meanwhile CXMT is growing where supply already exists — monthly DRAM capacity near 320,000 wafers rising to about 420,000 by 2027, and a 9.5% share of global DRAM revenue in 2Q26. CXMT is a share donor at the commodity end, not a price setter at the HBM end.

The point of those numbers is capacity to act. SK hynix generated ₩127.2 trillion of operating cash flow over the trailing twelve months against ₩36.0 trillion of capital expenditure, leaving ₩91.2 trillion of free cash flow — ample to equip a Dalian line while still funding Korea. Samsung's profit engine is even more concentrated: Device Solutions' ₩89.2 trillion of operating profit was 99.7% of group operating profit. Neither group is upgrading China because it is short of capital; both are upgrading because they cannot afford to let a third of their memory output fall behind.