The headline names of the AI boom are the GPUs. The companies actually shipping the boxes that move data between them rarely get a mention.
One of them, Eoptolink, a Chengdu-based maker of optical transceivers, grew its revenue by 179% in a single year to roughly $1.2 billion — almost entirely by plugging Chinese-made optics into American AI data centers.
What an optical transceiver actually does
An optical transceiver is the small pluggable device that converts electrical signals from a server or switch into light, sends them down a fiber, and converts them back. Inside an AI cluster, thousands of GPUs must talk to each other at enormous speed; the transceivers are the "signal translators" that make that conversation possible. The faster the cluster, the more transceivers each GPU needs.
When Nvidia scaled its training systems from thousands to tens of thousands of accelerators, the bottleneck quietly shifted from compute to interconnect. Suddenly, a humble module became a long-lead-time constraint — and Chinese suppliers were ready.
The 800G ramp and Eoptolink's numbers
Eoptolink (新易盛, stock code 300502.SZ) rode the shift from 400G to 800G links harder than almost anyone. Its 2024 results, disclosed in the company's annual report and carried by Chinese financial outlets, show:
- Revenue of 8.65 billion yuan (about $1.2 billion), up 179.15% year on year.
- Net profit attributable to shareholders of 2.84 billion yuan (about $400 million), up 312.26%.
- A gross margin of 44.7%, up sharply as higher-value 800G products displaced older lines.
- Overseas revenue of roughly 79% of the total — a business built on serving foreign cloud customers.
Those revenue and profit figures appear consistently across the company's audited report and secondary financial coverage, so the growth curve is solid even if the absolute totals still trail the global leader, Innolight, which CMB International named the 2024 market-number-one.
Why China leads the module, not the chip
The awkward truth for Beijing is that Eoptolink is brilliant at assembly, not at the two most valuable pieces inside the module. The digital signal processor (DSP) that cleans up the signal and the laser chip (typically an EML) that generates the light still come largely from US firms such as Broadcom and Lumentum, with Coherent also a key laser supplier. Industry analysis puts Broadcom's overall gross margin near 68%, a reflection of how much value sits in the chip rather than the box around it.
China's optical-module makers — Innolight, Eoptolink, and Suzhou TFC — are estimated to hold roughly seven of the global top-ten seats, a remarkable manufacturing position. But they are often described as the "plumbers" of the AI data center: essential, high-volume, yet dependent on whoever controls the engine and the brain.
The math behind the boom is simple and brutal. As GPU clusters grow, transceiver demand grows faster than accelerator count, because every link between chips needs its own pair of modules; a single high-density rack can swallow hundreds of them. That is why a modest change in cluster architecture — 400G to 800G to 1.6T — translates into outsized, sudden component orders that suppliers struggle to forecast. It is also why the next frontier, co-packaged and near-package optics that move the light source onto the chip package, is being pursued so aggressively: electrified interconnects simply run out of room at these densities.
The 1.6T frontier and compute power (算力)
The next transition is already underway. Nvidia's Blackwell-class systems push network fabrics toward 1.6-terabit links, and Eoptolink was among the earliest vendors to ship 1.6T modules, demonstrating a 1.6T transceiver built on 3-nanometer DSP silicon at the OFC 2025 show. Analyst forecasts cited by CMB International project roughly 49 million 800G and 22 million 1.6T units shipping globally in 2026 — still leaving a multi-million-unit supply gap, most acute at 1.6T.
Here is the strategic knot: the large models (大模型) that define today's AI need ever more compute power (算力), and compute power at cluster scale is gated by how many of these modules can be built — and by whether the DSP and laser chips behind them remain available.
Tariffs and the Thailand hedge
Because so much of Eoptolink's revenue is overseas, US tariffs on China-made goods are a live risk. The company has answered by building in Thailand: its Thai factory's second phase entered operation in early 2025, joining peers like Innolight and TFC that have also expanded Southeast Asian capacity to serve American customers without the tariff burden. Reporting from Reuters and analysts at CLSA notes this geographic diversification as the sector's primary hedge against trade friction.
Honest limitations
The 2024 financials are audited and well-corroborated, but they describe a single boom year in a cyclical, order-driven hardware business; such margins rarely survive a demand air-pocket. The claim that Chinese firms hold "seven of the top ten" seats is an industry estimate, not a formal market-share audit. Most importantly, Eoptolink's dependence on US-sourced DSP and laser chips means a targeted export restriction could hit the company harder than the Entity List hits a purely domestic player. The 1.6T shipment forecasts are analyst projections, not booked orders.
What readers can do now
- Track Eoptolink's gross margin and Thailand capacity as the cleanest read on whether the 800G/1.6T boom is structural or a one-time spike.
- Watch US export policy on DSP and laser chips: those, not the modules, are the real leverage point over China's AI interconnect supply.
- For "China chips" watchers, separate module assembly strength (real) from core optoelectronic-chip independence (still absent) — the gap defines the vulnerability.
