In AI hardware the binding constraint is packaging and optics, not chips

The headline is a ban, the constraint is the calendar

8 October was the first session after China's national holiday, and two laser-chip makers closed the day at the 20% limit-down. EFY's compilation shows losses of 5.4% to 6.5% continuing on 9 October, while the large module makers fell between 2.14% and 2.5%. The trigger was a Morgan Stanley note dated 1 October, which said the US Federal Communications Commission (FCC) could restrict Chinese makers of optical transceivers in October. An optical transceiver is the part that converts data moving between servers and accelerators in a data centre into light signals.

Headlines read this as the footsteps of an import ban, but the detail of the note says something else. The curb is most likely to target 3.2-terabit products, a next generation not yet on the market. One Chinese maker has said commercial production of these products will not begin before 2028–2029. Washington's real option is to restrict what comes two years from now, not today, because its own data centres would pay the price of restricting today's supply.

No substitute in optics

Counterpoint's August estimate shows Chinese makers holding about two-thirds of global optical transceiver unit supply. The market leader alone has 27% of the data-centre transceiver market; the largest US rival has 17%. That same Chinese leader earned more than 60% of its revenue in the United States in the first quarter of 2026, and its customers include America's largest AI companies. According to Counterpoint, Western makers cannot replace this volume within two years.

That is why the likely rule takes the form of a content requirement, not a ban. According to the Morgan Stanley note, Washington could allow Chinese-made modules whose bill of materials is at least 65% US-sourced. The requirement targets not the module but the laser chip inside it; the fact that upstream laser-chip suppliers, not module makers, fell hardest on 8 October points the same way. One Chinese maker called the 65% figure market speculation but did not rule out building a plant in the United States.

Political pressure is also building. On 25 September four senators introduced a bill banning Chinese-made optical modules from the federal government's classified systems, naming two Chinese makers. Under a draft reported by Reuters in August, the administration planned to ban new Chinese-made transceiver models and announce the rule by the end of 2026. The rules the FCC finalised in September, however, did not name any product.

In packaging, the dependence sits in Taiwan

The other end of the interconnect layer is packaging. AI accelerators link the compute die and high-bandwidth memory inside a single package via a silicon interposer; TSMC builds this structure with a method called CoWoS. According to The Register, when TSMC began producing Nvidia's Blackwell chips in Arizona in 2025, they were shipped back to Taiwan for final assembly. A chip made in the United States cannot be used until it is packaged.

The deal GlobalFoundries announced on 8 October is the first piece aimed at breaking that loop: its Malta fab will produce interposers for TSMC's CoWoS ecosystem for five years, in a deal worth $2 billion. Volume production starts in the first half of 2028. Output at Amkor's Arizona packaging plant is planned for early 2028, and TSMC's own Arizona packaging plant for 2029. End-to-end AI chip production in the United States will therefore be possible in 2028 at the earliest.

Capacity in Taiwan is meanwhile growing fast. According to an estimate cited by TrendForce, TSMC's CoWoS capacity will rise from about 130,000 wafers a month at end-2026 to 260,000 by end-2028, with most of the expansion at the AP7 site in Taiwan. The record quarterly revenue of NT$1.49 trillion TSMC reported on 8 October shows demand is filling that capacity. The US share will grow in absolute terms but remain small in proportion until 2028.

Why now, and what comes next

The two steps landing in the same week is no coincidence. On 23 September the US–China truce was extended from 10 November to 10 January 2027; whether the suspension of China's rare-earth controls is also covered by the extension could not be verified. Within this three-month window, Washington wants to reduce dependence on China without breaking the truce. Targeting the next optical generation and building a domestic packaging source both shape the post-2028 landscape without cutting today's supply.

The first test is the FCC's October–November agenda: if the rule targets only 3.2-terabit products and upstream components, this reading is confirmed. The second test is the investment timetables of GlobalFoundries and Amkor; if the 2028 target slips, US dependence on Taiwanese packaging extends to 2029. Türkiye's direct exposure is limited, but Turkish data-centre investment draws on the same optical and accelerator supply chain and will see cost increases in import prices.