Abstract
Recently, rumors regarding North American restrictions on Chinese high-speed optical modules have intensified again. However, at the administrative regulatory level, mainstream Chinese data center optical module manufacturers such as Zhongji InnoLight, Eoptolink, Accelink, and Dongshan Precision have not yet been listed on the Entity List that triggers commercial supply cutoffs. Relevant customs tariff codes have also obtained exemptions in this year's semiconductor tariffs and the new Section 301 tariffs. Behind the repeated policy deliberations is the stark reality of effective capacity in the inaugural year of 1.6T commercialization—even though North American hyperscale cloud vendors are fully supporting the US-based representative company AOI through warrants and billion-dollar three-year long-term agreements, its Q42026 order fulfillment rate for 1.6T modules remains only between 11.1% and 17.5%, with monthly effective output still far below a fraction of China's top two leaders.
The drastic gap from nominal capacity to actual delivery is primarily constrained by the quota-based allocation of 3nm optical digital signal processing (DSP) chips, the test cadence for 224G ultra-high-frequency mixed insertion, and delays in the delivery of grid transformers in the United States. Digging deeper into financials and upstream materials, North American cloud giants generally adopt a rolling pull model with zero prepayment and no guaranteed minimum volume for pluggable module procurement. Chinese leading module manufacturers advance billions of RMB in cash to lock in future wafer and chip quotas, absorbing manufacturing costs with high yields of 85% to 90%. Meanwhile, export license constraints on indium phosphide substrates, silicon photonics reduction solutions led by Chinese vendors, and the revenue dependence of US chip makers on Chinese buyers collectively form a bidirectional interlocking structure that is difficult to sever unilaterally.
Looking through the Q3 2026 production ledgers and regulatory timelines, the compromise sentiment of "abandoning bans on current products and retreating to setting limits at the 3.2T speed tier" essentially aligns with the concentrated capacity ramp-up nodes of overseas 6-inch indium phosphide lines and Co-Packaged Optics (CPO) platforms in 2028–2029. Facing the trend where security reviews over the next two years will penetrate from physical country of origin at customs to parent company control and component sourcing, Chinese optical communication enterprises must leverage the 1.6T silicon photonics volume window to define new optical engine products downstream and tackle high-speed lasers and core electrical chips upstream. Only then can they transform current delivery barriers into long-term industrial discourse power.
Introduction
In early November 2026, news that "Microsoft pressured the White House to halt the ban on Chinese optical modules, with Washington possibly retreating to limit speeds at 3.2T" rapidly circulated in tech and capital markets. Some market briefs even quoted a vivid warning: expensive AI chips have already been purchased; without Chinese optical modules, these chips would sit idle and gather dust in server rooms. However, checking the Federal Communications Commission (FCC) electronic docket up to November verifies that Microsoft never submitted any public opposition letter, responding to Reuters inquiries only with "decline to comment." The named warning that "purchased AI chips literally have to wait empty-handed for optical transceivers" actually came from John Miller, General Counsel of the Information Technology Industry Council (ITI), an organization representing Silicon Valley tech giants. On the institutional execution side, although the FCC implemented new rules closing the "logical hardware component loophole" effective October 13 this year, none of the mainstream Chinese data center optical module vendors—including Zhongji InnoLight, Eoptolink, Accelink, and Dongshan Precision (Source Photonics)—have been included in the "Entity List" that truly triggers commercial supply cutoffs. Furthermore, the customs tariff code for optical modules (8517.62.00) has consistently received exemptions in the intensively implemented Section 232 semiconductor tariffs and the July version of Section 301 tariffs.
On one hand, Congress continues to name entities in bills and leak ban proposals; on the other, the administrative regulatory layer repeatedly hesitates regarding current 800G and 1.6T commercial deliveries. Following the dense disclosure of Q3 2026 production and financial ledgers, what emerges is not just the surface of lobbying games, but the real wall North American hyperscale cloud vendors hit when trying to foster domestic alternatives—while computing giants have ample capital expenditure, they cannot buy scalable optical interconnect delivery capabilities replaceable in the short term.
I. Holding Hundreds of Millions in Long-Term Orders, Why Is US Domestic Production Line Fulfillment Below 20%?
In macro narratives, it seems North American cloud giants need only allocate a small portion of their $730 billion annual capital expenditure to support local manufacturers with US wafer fabs and compliance labels to quickly establish a parallel supply chain. AOI (Applied Optoelectronics), the most aggressive US-based optical module manufacturer expanding locally, provides a real sample to test this hypothesis: To build a domestic safety cushion for the 1.6T era, Amazon signed warrant-linked agreements with them early on, and another top North American cloud vendor signed a three-year exclusive framework agreement for 1.6T modules in October this year, promising to assign at least 10% of their total 1.6T demand to AOI provided supply targets are met, with an expected total value of $1 billion to $1.5 billion. As of early November, AOI had backlog confirmed orders for 1.6T modules worth $400 million to $450 million; however, during the Q3 earnings call on November 5, the company guided actual deliverable 1.6T revenue for Q4 to only $50 million to $70 million, resulting in a single-quarter order fulfillment rate as low as 11.1% to 17.5%. Converted at an initial factory average price of $850 to $1,200, AOI's Q4 monthly effective 1.6T output is only about 14,000 to 27,000 units, less than 9% of its publicly claimed "nominal capacity of 330,000 to 500,000 units/month for high-speed mixed lines." In contrast, looking at the global supply landscape, the Chinese top-two dualists (Zhongji InnoLight and Eoptolink), with over 98% of revenue focused purely on optical modules, achieved combined Q4 monthly effective 1.6T outputs of 500,000 to 1.38 million units (with Zhongji InnoLight at approximately 350,000 to 900,000 units and Eoptolink at approximately 150,000 to 480,000 units). Zhongji InnoLight's single Thailand factory alone has a monthly capacity at the 500,000-unit level. Even the US photonic powerhouses Coherent and Lumentum, excluding external chip sales and Optical Circuit Switches (OCS), are constrained by rigid flow cycles from internal wafers to modules lasting a quarter (with work-in-progress accounting for 60%), resulting in Q4 monthly effective 1.6T supplies of only 110,000 to 320,000 units.
The reason nominal capacities in the tens of millions shrink drastically at the 1.6T delivery site is that transitioning production lines from 800G (8×100G) to 1.6T (8×200G) is not simply swapping shells for mixed use, but requires passing two hard thresholds: upstream quotas and high-frequency testing. On the material side, 1.6T pluggable modules rigidly depend on TSMC 3nm optical DSP chips monopolized by Broadcom (Sian3) and Marvell (Ara), with lead times of 9 to 18 months. During capacity shortages, chip manufacturers naturally prioritize the top two global module leaders with the largest shipment bases and strongest bargaining and coordination capabilities. Smaller catch-up players like AOI (data center modules account for only 58% of revenue, with the remaining 41% being cable TV equipment) immediately hit chip quota ceilings if customer demand switches rapidly to 1.6T. On the manufacturing verification side, 1.6T modules operate at ultra-high baud rates of 106.25 to 113.4 GBd. Existing 400G/800G bit error rate testers are not backward compatible upward. Every module must complete sub-layer link training and cascaded forward error correction stress tests mandated by the IEEE P802.3dj standard on test benches above 65GHz before shipping, significantly extending temperature cycling burn-in dwell times. Additionally, cloud giants require modules from different suppliers to achieve "zero abnormal disconnection" mixed interoperability when plugged into the same 102.4T switch. Each time a new second-source chip supplier is introduced, joint acceptance testing in customer laboratories takes several months.
More difficult to accelerate with capital than chip quotas and test cadences are the physical constraints of US domestic infrastructure. AOI invested $150 million to expand its Sugar Land Phase II factory in Texas, originally planning to push monthly capacity to 650,000 units by the end of this year. However, due to sharing the ERCOT Texas grid with densely launching AI data centers nearby, it got stuck on long lead times for high-voltage cables and transformers, forcing a delay of several months and a downward revision of the year-end capacity target to 500,000 units. This raises the question: If emerging catch-up players are trapped by capacity ramps, why do US photonic giants with stronger capital and technical capabilities also refuse to take on large-scale assembly tasks for pluggable optical modules in the US?
II. Zero Prepayments and Billions in Material Stockpiling: Who Bears Supply Chain Risks for Cloud Giants?
If production line磨合 and power infrastructure can be digested with time, opening the balance sheets in Q3 2026 reports reveals a second, more hidden commercial barrier: In the AI hardware cycle of rapid technology iteration, leading Chinese optical module manufacturers are using their own balance sheets to bear the heaviest working capital risks across the entire chain for North American cloud giants. Outsiders seeing "2027 optical module orders cover the full year" often mistakenly assume module factories hold guaranteed contracts locking in long-term returns. However, Zhongji InnoLight disclosed honestly in its HK IPO prospectus published in July this year that major customers typically do not sign unconditional "take-or-pay" long-term purchase commitments with high liquidated damages, and may adjust purchase volumes, modify product specifications, or delay project progress at any time. In practice, North American cloud giants rarely pay prepayments to pluggable module manufacturers. Instead, they issue annual share guidance and rolling purchase orders. Module factories ship goods to overseas Vendor Managed Inventory (VMI) warehouses, and revenue recognition is only completed upon actual customer pull.
Comparing statutory financial statement items as of September 30, 2026, this asymmetry in power between upstream and downstream is clear. Under the "Contract Liabilities" item representing customer prepayments, Zhongji InnoLight's book value is only CNY 11 million, and Eoptolink's is only CNY 305 million, less than 2.8% of quarterly revenue. AOI management also admitted on the call that cloud giants scrutinize their balance sheets strictly, making the possibility of paying prepayments for pluggable modules negligible. However, looking upstream, to secure schedules locked until 2028–2029 for Indium Phosphide laser wafers, Tower and GlobalFoundries' 12-inch silicon photonics foundry quotas, and 3nm DSP chips, Zhongji InnoLight's advance payments on current assets surged 1,068.43% from the beginning of the year to CNY 1.567 billion, with inventory rising to CNY 23.878 billion. Eoptolink's current advance payments skyrocketed 7,695.01% to CNY 1.322 billion (plus long-term equipment and material prepayments reaching CNY 3.24 billion), with inventory rising to CNY 17.413 billion. This extreme tight balance of "zero downstream prepayments, tenfold upstream advances for materials," combined with mature mass-production yields of 85% to 90% (similar US domestic lines hover around 60% to 70% yield, with manufacturing costs 30% to 50% higher) and a mature Southeast Asian delivery base (Eoptolink's Thailand revenue alone reached CNY 21.9 billion in 2025), constitutes the commercial trump card for China's module dualists to capture approximately 60% of global data center optical module revenue and two-thirds of shipment volume.
Calculating this financial ledger explains why US photonic powerhouses Coherent and Lumentum are unwilling to enter the competition for pluggable optical module assembly. Commercially, nearly 60% of Lumentum's latest fiscal quarter revenue came from direct sales of high-margin laser chips, with Non-GAAP gross margins reaching 53.2%. The recent announced massive long-term agreements of $20 billion and over $5 billion by both companies also lock in Co-Packaged Optics (CPO), Near-Packaged Optics (NPO) integration solutions, and External Light Sources around 2028. Even in their retained proprietary data center optical module business, packaging and testing lines are heavily rooted in Thailand, Malaysia, and Vietnam. Lumentum's net fixed assets in Thailand already account for nearly 40% of its global total. Asking US original manufacturers accustomed to over 50% chip gross margins to violate Wall Street capital efficiency principles and build heavy-capital, thin-margin, high-yield-threshold assembly plants back in the US simply does not close the loop commercially. So, since US giants firmly control upstream high-end chips, why doesn't Washington cut off chip supply to force a migration of the supply chain?
III. Holding High-End Chip Cards, Why Doesn't Washington Dare to Cut Current Supply?
Pushing the view further to upstream materials and chip architectures touches the core underlying constraint of this game: Sino-US relations in the high-speed optical communication chain are not unidirectional dependencies but form a deeply interlocked bidirectional structure of "Chips—Substrates—Silicon Photonics Packaging." On the surface, the highest-value 3nm DSP electrical chips in optical modules (Marvell and Broadcom combined account for 90% share) and 200G Electro-absorption Modulated Laser (EML) chips (Lumentum, Broadcom, and Mitsubishi Electric combined account for 72% share) are highly concentrated in US and Japanese hands. However, at the most source substrate material stage, whether for 200G EMLs or Continuous Wave (CW) lasers required for silicon photonics modules, growth must occur on Indium Phosphide single-crystal substrates. Moreover, the Indium Phosphide substrate area consumed by a single 1.6T EML module is 2.7 to 2.8 times that of the 800G era. Since China included Indium Phosphide and related indium items under export license controls in February 2025, AXT, one of the world's three major substrate leaders with main production lines in Tongmei, Beijing, stated directly in its Q3 2026 report that export licenses are the most severe operational challenge, unable to predict when shipments to the US will resume. Lumentum also explicitly acknowledged in its annual report that China's export restrictions to Japan impacted its global substrate supply chain. Even after paying an $87 million deposit to AXT to lock in long-term agreements, substrate supply may still face shortages.
Facing the supply bottleneck of traditional upstream laser chips—the global shortage of 200G EMLs is projected to reach 70% in 2027, with Coherent's self-produced 6-inch Indium Phosphide lasers all reserved for internal use and never sold externally, and Lumentum's capacity sold out until early 2029—leading Chinese module manufacturers took the lead in engineering breakthroughs via "Self-developed Silicon Photonics Chips + External CW Light Sources." Relying on on-chip "1-to-2" or even "1-to-4" splitting designs, the number of lasers required per 1.6T module was reduced directly from 8 expensive EMLs to 4 or even 2 CW lasers, slashing laser material costs from approximately $160 down to $15–$20. This architectural leap not only rapidly pushed the 1.6T silicon photonics penetration rate to 60%–72% in 2026 but also opened a scaled introduction window for second-source suppliers like Yuanjie Semiconductor in the field of high-power 70/100mW CW light sources for local optical chip manufacturers.
More realistic commercial counter-measures come from the buyer's market itself. Leading Chinese optical module manufacturers, which swallow two-thirds of global shipments, are precisely the largest cash buyers of high-end chips from Broadcom, Marvell, Lumentum, and Coherent. If the administrative side forcibly cuts off access to current commercial optical modules, US optoelectronic chip giants would suffer heavy blows to their quarterly revenue. Simultaneously, Microsoft alone currently has 35 data centers under construction in the US and 134 planned, which would immediately fall into a shutdown predicament of "having compute cards but no high-speed optical networks." It is under this "mutual loss" hard constraint that the recent policy wind direction from Washington quietly rewrote the timeline.
IV. Behind "Retreating to 3.2T": What Window Period Is Left for the Chinese Industrial Chain?
After seeing the triple constraints of capacity, payment terms, and materials, the time considerations behind the recently circulating compromise solution—"abandoning bans on current products, changing to only restricting Chinese optical modules with transmission speeds reaching 3.2T"—become obvious. According to the IEEE 802.3 Single Channel 400G Study Group and the real engineering progress of global leading vendors, the "3.2T/6.4T production launch" mentioned by the industry currently refers entirely to CPO or NPO optical engine sub-units inside switches. True front-panel pluggable 3.2T optical modules (8×400G) remain at the exhibition prototype and DSP sampling stage, with scaled commercial mass production waiting until late 2027 to 2028, and widespread installation deployment not until around 2029.
2028–2029 is not an accidentally selected technology generation point, but a time window where multiple overseas supply chain layouts and institutional deadlines converge: On the capacity side, Coherent's Texas 6-inch Indium Phosphide expansion line backed by Nvidia's $4 billion investment, Lumentum's North Carolina 6-inch new plant, JX Metals' ¥120 billion substrate expansion project in Japan, and TSMC's COUPE 3D Co-Packaged Optics platform are all scheduled for concentrated capacity ramp-up during this period. On the institutional side, the supply chain bill under review in the US Senate (S. 5548, aiming to penetrate overseas subsidiaries and firmware origins, but currently limited to federal critical security systems with a five-year transition period), and the deadline for the FCC to terminate qualifications for non-reciprocal testing labs within China (December 1, 2028), also point to the same distant node. Using an unwritten "3.2T distant restriction" to respond to domestic political pressure, while continuing to default to Chinese leaders supplying North American data centers at full load through full-process manufacturing bases in Thailand and Malaysia (meeting customs substantial transformation rules) during the most urgent 800G and 1.6T rush in 2026–2027, constitutes the practical tacit understanding reached by all parties in the face of physical capacity iron laws.
For the Chinese optical communication industry sitting at the center of the table, these two precious years of "capacity exemption" do not mean enjoying assembly dividends passively, because the anchor of US security reviews is irreversibly sinking from "physical country of origin at customs (checking where packaged)" to "parent company control and hardware/firmware Bill of Materials penetration (checking who holds control, adopting chips and firmware from where)." On the future table, the path to converting current scale advantages into long-term safety cushions is very clear: Looking downstream, it involves accelerating the definition rights of slot-type NPO and 12.8T liquid-cooled XPO optical engines that do not rely on binding to a single wafer fab, based on 1.6T silicon photonics cash flows. Looking upstream, it involves steadfastly pushing the breakthroughs torn open today in 70/100mW CW light sources and passive high-density connectors into the deep waters of single-channel 200G/400G EMLs, large-size Indium Phosphide wafers, and high-speed electrical chips. After all, in the game of computing infrastructure, what truly prevents opponents from casually flipping the table has never been goodwill, but indestructible delivery capabilities and opportunity costs too expensive to switch.