The so-what: paperwork loosened, the market still closed

Investors watching Asian chip names should not mistake procedural relief for demand recovery. On January 15, 2026, the US Bureau of Industry and Security (BIS) shifted its license posture on H200- and MI325X-class accelerators bound for China and Macau from a presumption of denial to case-by-case review 1 2. That reads like a thaw. Yet by mid-year the practical result was the opposite of a reopening: Nvidia's share of China's AI-accelerator market has cratered toward single digits, with Bernstein projecting roughly 8% in 2026 versus about 40% the year prior, while Huawei is expected to pass 50% 3. The policy machinery got friendlier; the addressable market got smaller. For anyone modelling Asian equipment makers, foundries, or Nvidia's China total addressable market, the operative variable is no longer the license letter but the buyer's decision to walk.

What actually changed in the rulebook

The January rule is conditional, not permissive. To win case-by-case approval, exporters must certify that filling a China order will not delay any US end-user order, that aggregate China and Macau shipments stay at or below 50% of what the exporter ships to US customers, and that foundry capacity is not diverted from more advanced US-bound chips 2. Each shipment needs performance certification from an independent US-headquartered lab, and remote-access safeguards apply to cloud scenarios 2. Eligibility is bounded by hard technical thresholds, broadly total processing performance below 21,000 and DRAM bandwidth under 6,500 GB/s, which keeps Nvidia's Blackwell generation outside the window entirely 2. Alongside the rule, the White House layered a 25% tariff on advanced AI chips meeting the same thresholds under Section 232 authority, announced January 14 2. So the headline of resumed sales came wrapped in a volume cap, a tax, and a certification burden.

Then came the enforcement tightening. On June 1, 2026, BIS confirmed that the licensing requirement follows a company's ultimate parent, not its postal address: any firm headquartered in China, or ultimately owned by a China- or Macau-based parent, needs a license to receive advanced chips even if the buying subsidiary sits in Singapore, Malaysia, or elsewhere 4. That plugged the extraterritorial loophole Chinese buyers had exploited after the Biden-era diffusion framework was scrapped in 2025 4. The Entity List continued to expand in parallel, with additions such as GMC Semiconductor and Jicun Semiconductor cited for diverting US-origin items to already-listed parties 5. For compliance teams across Asia, the change is meaningful: know-your-customer screening now has to trace ownership up the corporate tree, not just check a shipping label.

Fabs and equipment makers: managed access, capped ambition

The fab-tooling story ran the other way, toward pragmatic accommodation. For 2026, Commerce granted annual export licenses letting TSMC, Samsung, and SK Hynix keep shipping US-controlled tools into their existing China fabs, replacing the older validated-end-user waiver regime that expired at end-2025 6 7. The catch is that these annual quotas explicitly bar using the tools to expand capacity or upgrade to more advanced nodes 7. That is a deliberate design: keep mature-node and memory output flowing to avoid a supply shock, while freezing technological advancement inside China. For SK Hynix and Samsung, whose China plants remain material to global DRAM and NAND, the annual-license model reduces the tail risk of an abrupt shutdown but caps the strategic upside of those assets 6.

Upstream equipment names carry a different exposure. US officials pressed ASML in April and again publicly in June 2026 over concerns that advanced tooling or components had reached restricted Chinese parties, including support flagged around Huawei-linked partners 8. The expanded Foreign Direct Product Rule reaches equipment from foreign-headquartered suppliers such as ASML, Tokyo Electron, and SEMES, meaning Dutch, Japanese, and Korean toolmakers increasingly operate inside the perimeter of US extraterritorial reach 5 8. The investor read: allied toolmakers face a slow, structural compression of their China revenue, less a cliff than a managed decline with rising compliance friction.

The decoupling read for investors

The sharpest signal is a demand-side one. A Bloomberg survey published July 7, 2026 found Chinese executives expect domestic products to account for 46% of their AI-accelerator budgets over the next twelve months, up from 30% currently 3. Beijing is reinforcing the shift with a reported multi-year, two-trillion-yuan data-center build-out that targets at least 80% domestic core technology 3. When buyers switch by choice and by mandate at once, decoupling stops being a policy overhang and becomes a structural feature of the stack. The bifurcation is deepest at the software layer, splitting between Nvidia's CUDA and Huawei's CANN, where switching costs cut both ways and lock in whichever ecosystem wins each customer 3.

Positioning implications: first, treat Nvidia's China line as structurally impaired regardless of license relief, and stress-test TAM models that still assume a recoverable China segment. Second, favor Asian memory and foundry names whose China assets are now cash-flow-stable but growth-capped, and price the strategic ceiling accordingly. Third, watch allied toolmakers for the slow squeeze rather than a single shock, with headline risk clustered around FDPR enforcement actions. The thaw in paperwork is real, but the durable trade is the one that assumes the two chip ecosystems keep drifting apart.