Most of China's semiconductor industry is running away from the West — substituting, localizing, designing out American content wherever the Entity List bites. The optical transceiver makers are running in the opposite direction. Zhongji Innolight (中际旭创, 300308.SZ) books over 90% of its revenue overseas, supplies more than half of Nvidia's 800G optical modules, and counts Alphabet, Amazon, and Meta among its largest disclosed customers. It is now the single largest weight in the CSI 300 — bigger than CATL, bigger than Moutai — a Chinese national champion whose commercial lifeline runs entirely through North American data center budgets. Of the 1,279 Chinese semiconductor companies in the Semisino database, the optics complex is the only cluster with major direct revenue exposure to the Western AI buildout. That makes these companies the purest China-listed play on hyperscaler capex, and simultaneously the most exposed Chinese semi names to Washington's next move. In June 2026 the Pentagon put Innolight on its Section 1260H list of Chinese military companies. Both things are true at once, and that tension is the whole story.

The segment at a glance

Optical transceivers convert electrical signals to light and back, moving data between servers, switches, and GPU racks over fiber. The module itself is a dense little assembly: a laser (EML, VCSEL, or a CW source feeding a silicon photonics chip), a photodetector, a DSP to clean up the signal, and a great deal of precision packaging — lenses, fiber arrays, ceramic substrates, alignment and burn-in. China dominates the assembly layer. Per LightCounting, seven of the global top ten optical module suppliers in 2024 were Chinese, up from two (Accelink and Hisense Broadband) in 2016. Innolight is the global number one; Eoptolink (新易盛, 300502.SZ) has climbed to the number two or three slot depending on the quarter; industry estimates put the pair at roughly 60% of Nvidia's 800G module volume, with Coherent and Lumentum splitting most of the rest.

The Semisino database tracks roughly 75 companies in the optoelectronics segment across four categories — Optical Transceivers, Laser Chips, Silicon Photonics, and Optical Components — and the segment splits cleanly along one axis: who sells into the global AI supply chain, and who sells into China's carriers and domestic cloud. The exporters (Innolight, Eoptolink, TFC, CIG, Linktel) have been the best-performing stocks in the entire Chinese semiconductor universe. The domestic-facing names (Accelink, Hisense Broadband, and most of the components tier) have done fine, but they are riding a different, smaller wave.

The money involved is no longer niche. LightCounting puts the AI-driven optical transceiver market at $16.5 billion in 2025, heading for roughly $26 billion in 2026 — about 60% growth in each year. TrendForce projects shipments of 800G-and-above transceivers jumping from roughly 24 million units in 2025 to on the order of 63 million in 2026 as 1.6T ramps. The A-share market has noticed: Innolight's market capitalization crossed RMB 1 trillion (roughly $190 billion) this spring, Eoptolink sits near $85 billion, and TFC near $40 billion. The trio — nicknamed 易中天 by A-share traders, a pun stitching their names into a celebrity historian's — carried much of the CSI 300's 2025 performance.

Why AI rewired this industry

The Chinese module industry was built in three stages. The domestic 4G/5G buildout of the 2010s was the training ground — a $2–3 billion annual home market, per LightCounting, that let Chinese assemblers grind out yield and cost learning on 100G and 200G modules. The 400G cloud cycle of the early 2020s was the export breakout: US hyperscalers wanted price, volume, and speed, and Suzhou delivered all three. Innolight became a core supplier to Google; Eoptolink embedded into the AWS chain.

Then AI clusters changed the math. A traditional server needed two to four transceivers. An AI training cluster needs multiple 800G links per GPU — optics demand now scales with GPU shipments, and each speed generation (800G in 2023–25, 1.6T from late 2025) resets pricing and margins upward before cost-down kicks in. When Nvidia needed millions of 800G modules delivered on schedule for the GB200 ramp, there were only a handful of companies on earth with the capacity, and most of them were Chinese. That scarcity handed the assemblers pricing power they had never enjoyed: Innolight's gross margin hit a record 46% in Q1 2026; Eoptolink's 2025 net margin, at roughly 38%, was the highest of any publicly traded transceiver vendor anywhere.

The result is a genuinely strange geopolitical artifact: at the exact moment Washington is trying to wall Chinese firms out of AI compute, the physical layer of American AI data centers is being assembled, tested, and shipped — increasingly from Thailand — by Chinese companies whose share prices trade on Nvidia's earnings calls.

The exposure ranking

Ranked by direct revenue exposure to the Western AI buildout — overseas revenue mix, hyperscaler/Nvidia customer concentration, and 800G/1.6T product mix:

| Rank | Company | Ticker | 2025 revenue (RMB) | Growth | Overseas mix | Exposure | |---|---|---|---|---|---|---| | 1 | Zhongji Innolight | 300308.SZ | 38.2bn | +60% | ~91% | Nvidia, Google, Amazon, Meta | | 2 | Eoptolink | 300502.SZ | 24.8bn | +187% | ~80% | Nvidia, AWS-centric | | 3 | TFC Communication | 300394.SZ | 5.2bn | +59% | high, via top-5 clients | Optical engines/FAU into global chain, Nvidia CPO | | 4 | Cambridge Industries (CIG) | 603083.SH | ~4bn-scale | — | high | Hyperscaler ODM, Malaysia/Mexico | | 5 | Linktel | 301205.SZ | 1.3bn | +41% | meaningful | Datacom modules, Malaysia ramp | | 6 | HGTECH / HG Genuine | 000988.SZ | 11.9bn (parent) | +44% | mixed | 800G SiPh share ~18% | | 7 | Accelink | 002281.SZ | — | — | low | Domestic-weighted IDM |

Zhongji Innolight is the franchise. 2025 revenue of RMB 38.2 billion (+60%), net profit of RMB 10.8 billion (+109%), with optical modules 98% of the business and overseas customers 91% of sales. Then it accelerated: Q1 2026 revenue of RMB 19.5 billion was up 192% year on year, and that single quarter's RMB 5.7 billion net profit exceeded the company's entire 2024 earnings. Market estimates put Innolight at roughly 35–40% of the global 800G module market and 50–70% of the emerging 1.6T segment; J.P. Morgan ranks it co-leader of the broader optical components market alongside Coherent at roughly 20% apiece. Customer concentration is the mirror image of the opportunity: Alphabet alone is roughly 22% of revenue, Amazon 11%, Meta 6%, with Nvidia flowing through both direct and ODM channels. The company runs a 70,000-plus square meter Thailand plant — capacity on the order of 200,000 high-end modules a month — as its primary base for North American shipments, has added a Mexico facility, and has reportedly filed for a Hong Kong secondary listing that could raise up to $7 billion. It is also, as of June 2026, on the Pentagon's 1260H list, of which more below.

Eoptolink is the momentum trade. Revenue nearly tripled in 2025 to RMB 24.8 billion (+187%), net profit rose 236% to RMB 9.5 billion, and gross margin reached 47.8% — the best economics in the industry, helped by a leaner product line concentrated almost entirely in datacom and an early, aggressive bet on LPO (linear pluggable optics, which deletes the DSP from the module). Eoptolink is understood to be deepest in the AWS supply chain while also taking a large slice of Nvidia volume, and its Sriracha, Thailand campus — built out in two phases through 2025 — is where its 1.6T ramp is happening. At OFC 2026 it was showing 1.6T variants in every flavor plus 6.4T near-package demos, a signal that it intends to fight on the CPO transition rather than concede it.

TFC Communication (天孚通信, 300394.SZ) is the picks-and-shovels play one layer down: fiber array units (FAUs), lenses, ceramic ferrules, and increasingly complete optical engines sold to the module makers and — critically — into co-packaged optics programs. 2025 revenue grew 59% to RMB 5.2 billion with net margins near 40%, and the company claims to have been first in the world to deliver 800G and 1.6T optical engines, with 1.6T now in full-scale delivery. TFC's HKEX filing shows its top five clients contributed over 90% of revenue before 2025 — extreme concentration, but concentration in exactly the right customers. Its strategic kicker is CPO: as the interface between optical engines and the switch ASIC, FAU content per system runs 3–5x that of traditional pluggables, which is why TFC is the one Chinese name consistently linked to Nvidia's CPO supply chain. Thailand Phase II came online in 2025.

The second tier is smaller but genuinely exposed. Cambridge Industries Group (剑桥科技, 603083.SH) began mass delivery of 800G modules in May 2025, has its overseas R&D team sampling 1.6T designs on 3nm 200G-per-lane DSPs, and is spending $56 million on Mexico capacity to serve North American hyperscalers from a Malaysia-plus-Mexico footprint. Linktel (联特科技, 301205.SZ) grew 2025 revenue 41% to RMB 1.3 billion as its Malaysia base ramps. Source Photonics — unlisted, Chinese-owned since its acquisition by Chinese capital, with West Coast roots — is the sleeper: it ranked roughly fourth and fifth globally in 400G and 800G shipments in 2025 respectively, and it is one of the few Chinese-owned module makers shipping in-house EMLs, including 200G-per-lane parts for 1.6T. HG Genuine, the optical module arm of HGTECH (华工科技, 000988.SZ), holds roughly 18% of the 800G silicon photonics module market on the back of in-house SiPh chips it claims cost 40% less than overseas equivalents. Accelink (光迅科技, 002281.SZ), the state-linked IDM under the CICT umbrella, is the most vertically integrated of all — the only domestic player producing 25G EMLs in-house — but its revenue skews to carriers and domestic cloud, making it the least exposed of the majors to Western AI capex and the most protected if that capex ever gets fenced off.

The upstream chip question

Here is the irony at the center of the trade: China's biggest AI-export winners are themselves import-dependent at the chip level. Crack open an 800G module from Suzhou and the DSP is from Broadcom or Marvell — an effective duopoly with no mature Chinese alternative — and the EML laser is overwhelmingly from Lumentum, Coherent, Mitsubishi, or Sumitomo. Fewer than five companies worldwide make EMLs at commercial scale, and per TrendForce, Nvidia's pre-booking of EML capacity has pushed industry lead times beyond 2027; in March 2026 Nvidia went further, committing roughly $2 billion each to Lumentum and Coherent to lock up laser supply. Domestic localization of 25G-and-above optical chips is still estimated below 15%. The Chinese module industry, in other words, is an assembly powerhouse sitting on a chip desert — it captures the revenue line of the AI optics boom while Broadcom, Marvell, and the laser duopoly capture the structural margin.

The domestic response runs on three tracks. First, laser chip localization: Yuanjie Semiconductor (源杰科技, 688498.SH) grew 2025 revenue 139% to RMB 601 million and swung to profit as its CW laser sources shipped in volume into data center silicon photonics modules; by external sales it now ranks sixth among laser chip makers globally and second in lasers for SiPh interconnects, with 100G and 200G EML programs in qualification. Startups like Zetta Semiconductor demoed 100G and then 200G PAM4 EMLs in 2025, and Source Photonics and HG Genuine ship in-house lasers — but high-volume, high-yield 200G-per-lane EML remains a Western and Japanese franchise for now. Second, architecture workarounds: silicon photonics substitutes a simpler CW laser (where Yuanjie and Shijia Photons, 688313.SH, play) for the scarce EML, and Innolight is already the global leader in 800G SiPh modules with SiPh estimated above 40% of its mix; LPO deletes the DSP line item outright. Third, the components tier — Shijia's AWGs and MPO connectors, Broadex (博创科技, 300548.SZ) in planar lightwave circuits, Advanced Fiber Resources (300620.SZ) in passives, Everbright Photonics (688048.SH) in laser chips — is scaling beneath the module makers. None of this eliminates the dependence; SiPh wafers still come largely from GlobalFoundries, Tower, and TSMC. It changes the shape of the leash rather than removing it.

Risks: tariffs, Washington, and CPO

The risk stack is unusually legible because it all runs through one customer geography. Chinese-origin transceivers already carry 25% Section 301 tariffs, which is precisely why the industry decamped to Thailand and Malaysia — Innolight, Eoptolink, and TFC all now serve North America primarily from Thai lines. But the tariff hedge is not free of risk: US Customs is escalating scrutiny of Southeast Asian transshipment and the "substantial transformation" question, and a DOJ trade fraud task force stood up in 2025. The December 2025 USTR Section 301 action on semiconductors added a new sword of Damocles: an initial tariff of 0% on a broad list of Chinese chips, scheduled to rise in June 2027 to a rate not yet announced. And the Pentagon's June 2026 addition of Innolight to the 1260H list carries no direct sanctions but is historically a waiting room — hyperscaler procurement teams have now been handed a compliance reason to accelerate qualification of Coherent, Lumentum, and Fabrinet-assembled alternatives, and Fabrinet remains the only large-scale non-Chinese merchant assembler if diversification gets serious.

The technology risk is co-packaged optics. Nvidia's Quantum-X InfiniBand CPO switches arrived in late 2025 and Spectrum-X Photonics targets 2H26, built on TSMC's COUPE process with a partner list — Coherent, Corning, Lumentum, Foxconn, SENKO — conspicuously free of Chinese module makers, TFC's component role aside. CPO integrates the optical engine into the switch package and, in the limit, deletes the pluggable module and its maker. The consensus view, which we share, is that pluggables and CPO coexist well into the late 2020s — general-purpose data centers prize the serviceability of pluggables, and 1.6T pluggable volumes are still exploding — but the direction of travel favors whoever owns the silicon, and that is Nvidia, Broadcom, and TSMC, not Suzhou. A-share valuations north of RMB 1 trillion for Innolight leave little room for the scenario where CPO penetration steepens after 2027.

What to watch

Four markers over the next 12 months. First, the 1.6T ramp split: whether Innolight holds its estimated 50–70% share as Eoptolink, Coherent, and HG Genuine scale — 1.6T pricing will reveal whether current 45%-plus gross margins were a cycle or a franchise. Second, Washington's sequencing: the June 2027 Section 301 semiconductor rate announcement, any move to extend controls or tariffs to Thai-origin modules, and whether the 1260H listing metastasizes into procurement restrictions. Third, EML localization: Yuanjie's 100G/200G EML qualification and Zetta's ramp are the difference between an assembly industry and a vertically sovereign one. Fourth, CPO order flow — specifically whether TFC's FAU content in Nvidia programs grows fast enough to prove Chinese suppliers can migrate down the stack as the module migrates into the package. The optics exporters are the rare Chinese semi story that gets paid by the Western AI buildout rather than in spite of it. That is a wonderful place to be, right up until it isn't.