The So What

China's Ministry of Commerce is drafting export-control language that would bar Huawei, Alibaba, ByteDance and other domestic chip designers from sending their most advanced designs to TSMC and other foreign foundries 1 2 3. Read as pure industrial policy, it looks like Beijing finishing the decoupling Washington started. Read against what SMIC can actually deliver this year, it looks more like Beijing choosing to ration its own AI buildout rather than let TSMC keep doing it for free.

What Is Actually Being Proposed

The measures under review would fold into the next revision of China's catalogue of technologies restricted from export, and they go beyond foundry access 1 2. Regulators are also weighing curbs on transferring advanced AI models and the training data behind them, plus tighter scrutiny of Chinese firms' overseas technology acquisitions 3. All of it is still at the industry-feedback stage, nothing has been formally enacted, but the direction is unambiguous: keep frontier AI development, from training data to the silicon it runs on, inside Chinese borders 1 2 3.

The foundry piece is the one investors should watch first. Chinese fabless designers have spent years having their most advanced dies etched at TSMC on nodes SMIC cannot yet match at volume, even as the finished chips were assembled into systems and deployed domestically. A formal ban would close that gap on Beijing's own timeline rather than Washington's.

Why the Timing Reads as Leverage, Not Just Policy

The proposal surfaced the same window Washington opened its own review of how Chinese AI firms reach Nvidia chips through offshore subsidiaries and shell entities, after a run of Chinese model breakthroughs made clear existing restrictions were being routed around 4. Chinese buyers including ByteDance, Alibaba and Tencent have placed orders for more than 400,000 Nvidia H200 units this year, chips Nvidia's own finance chief has said have generated no recognized China revenue because delivery keeps getting tangled in the compliance fight 4. A Chinese ban on its own firms using TSMC lands as a mirror-image pressure point: if Washington can choke offshore chip access, Beijing can choke offshore chip fabrication.

The Capacity Math Beijing Would Be Betting On

This is where the policy runs into physics. SMIC's advanced-node capacity, 7-nanometer and below, sat near 45,000 wafer starts per month at the end of 2025, and industry trackers expect it to climb to roughly 60,000 by the end of 2026 and 80,000 in 2027, with early 5-nanometer pilot runs already underway for partners including Huawei and Alibaba 6. That wafer capacity alone could, in principle, bake die for well over a million Huawei Ascend-class accelerators a year 5 6.

The constraint sits one layer downstream, in memory. High-bandwidth memory is what turns a finished logic die into a usable AI accelerator, and China's only qualified domestic HBM supplier, CXMT, is projected to produce enough stacks next year for something like 250,000 to 300,000 Ascend-class chips, even as SMIC's die output runs far ahead of that number 5 6. Analysts tracking the ramp believe China's stockpile of foreign HBM, built up before controls tightened, ran out in late 2025, meaning every chip built from here depends on CXMT's memory line keeping pace with SMIC's wafers, which it currently does not 5 6.

Who Would Actually Get Squeezed

Huawei is targeting roughly 750,000 Ascend chip shipments in 2026, and ByteDance and Alibaba are reportedly lining up orders for Huawei's chips as a domestic hedge against tightening access to Nvidia 7. Layer a formal TSMC ban on top of that demand and the queue does not shrink, it just becomes captive. Every designer currently splitting orders between TSMC and SMIC would be pushed entirely onto a supply chain that, by its own trackers' estimates, can fully serve well under a third of the demand it would suddenly own 5 6 7. The near-term effect is not more Chinese AI chips reaching the market, it is longer wait times and tighter rationing among Huawei, Alibaba and ByteDance for the CXMT memory that gates all of them.

TSMC, meanwhile, has less exposure to lose than the policy debate implies. Chinese decoupling pressure from Washington has already pushed TSMC's revenue mix toward North American customers, who accounted for the bulk of sales in the company's most recent quarter, with AI accelerator demand from Nvidia and AMD doing the heavy lifting 8. A Chinese self-imposed ban would formalize a shift that United States export rules had already been driving for three years.

What to Watch

Three signals will show whether this becomes real: whether the catalogue revision actually gets published rather than staying in consultation, whether CXMT's HBM output tracks its 2026 projections or slips the way China's foreign-HBM stockpile already did, and whether Huawei's next-generation Ascend 950-series ramp gives Beijing enough domestic output to make the ban politically affordable. Until CXMT closes the memory gap, a formal TSMC ban is a stronger statement of intent than an operational reality, and the investors most exposed are not TSMC shareholders but the Chinese AI developers who would be first in line for a supply chain that cannot yet fill its own orders.