--- title: "The light pipes feeding AI: how Chinese optical modules became indispensable" date: 2026-10-04 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-silicon-photonics-interconnect language: en --- # The light pipes feeding AI: how Chinese optical modules became indispensable > As AI clusters hit an interconnect wall, China's pluggable optical-module makers — Innolight, Eoptolink, Accelink — supply the bulk of the world's 800G transceivers, even as the next leap, co-packaged optics, stays led from the US and Europe. The bottleneck in the AI boom is no longer only the chip that computes. It is the light that carries the answers between tens of thousands of those chips, and the cables are starting to melt under the load. A quiet corner of Chinese manufacturing now sits at the center of that problem. ## Why photons, not just electrons A modern AI training cluster is a mesh of GPUs that must talk to each other constantly. Electrical copper links lose the signal and burn power over distance, so the industry moves data as **light** through optical transceivers — the small modules plugged into switches and cards. As models grew, so did the speed grade: 400G gave way to **800G**, and **1.6T** is now ramping. Industry analysis puts optical modules at roughly 12% of AI infrastructure cost in 2023, climbing toward 18% by 2025. The catch is physics. At 800G and beyond, the pluggable module's own power draw and signal loss become the constraint — which is why the next debate is about moving the optics **onto** the chip package itself. ## China owns the pluggable layer At the pluggable transceiver manufacturing level, Chinese firms are the unambiguous global leaders. Reported 2024 figures tell the story: - **Zhongji Innolight (中际旭创)** — about **¥23.86 billion (≈ US$3.4B / HK$26B)** revenue, up roughly 123% year-on-year, with an estimated ~40% share of the 800G market. - **Eoptolink (新易盛)** — about **¥8.65 billion (≈ US$1.2B / HK$9.5B)** revenue, up roughly 179%, with ~22% share. - **Accelink (光迅科技)** — ~15% share, and a demonstrated low-power **1.6T OSFP224** module using a 3nm DSP plus silicon photonics. Estimates suggest Chinese manufacturers together supply the majority of Nvidia's 800G transceiver volume, and roughly 60% of global optical-module shipments overall. To dodge US tariffs, these firms have built plants in Thailand, Malaysia, and Vietnam — a "China design + Southeast Asia manufacturing" model that keeps them inside global supply chains. ## Silicon photonics is the bridge technology **Silicon photonics (硅光)** integrates lasers, modulators, and detectors onto a silicon chip, cutting cost and power versus discrete parts. At the 1.6T generation it has become the mainstream approach for high-speed modules; several Chinese vendors showed 400G/800G/1.6T silicon-photonics products at CIOE 2025 in Shenzhen this September, alongside next-step **3.2T** optical engines. The appeal is concrete: pairing silicon photonics with a 3nm DSP, as Accelink demonstrated, attacks the heat-dissipation bottleneck that limits density in AI racks. Vendors also pursue **LPO (Linear-drive Pluggable Optics)**, which drops the power-hungry DSP chip to cut module power by roughly half — a pragmatic bridge while the bigger architectural shift matures. ## The frontier still lies elsewhere Here is the uncomfortable split. At the **pluggable** layer China leads. At the **co-packaged optics (CPO)** frontier — bonding the optical engine directly to the switch ASIC to slash power at scale — the leaders are TSMC, Broadcom, Intel, Cisco, and coherent-laser specialists in the US and Europe. CPO switches are expected to reach commercial rollout around 2026, with the clearest value being a large cut in total cost of ownership for hyperscale data centers. In other words, China excels at making the proven thing at scale and at low cost, while the next leap in interconnect architecture is still being defined by US and European foundries and system designers. The standards fight — COBO internationally versus China's CAICT domestically — will decide who shapes that future. ## What readers can do now - If you run infrastructure, treat 800G/1.6T optics as a strategic procurement category, not a commodity; lead times and supplier concentration are real risks. - If you follow tech sovereignty, separate "who makes the modules" (China, today) from "who defines CPO and the lasers" (US/Europe) — the leverage shifts at the package boundary. - If you build products, prototype LPO or silicon-photonics links now; the power savings are large enough to change rack design, not just line items. ## Honest limitations Revenue and market-share figures for Innolight, Eoptolink, and Accelink come from industry analyses and trade reporting (including CIOE 2025 coverage and company/analyst summaries); I have not re-audited the companies' audited financial statements, and the ~40%/22%/15% 800G share splits are estimates that vary by source. The "60% of Nvidia 800G volume" and "majority of global shipments" claims are analyst estimates, presented as directional. The CPO commercialization timeline (~2026) reflects vendor and analyst roadmaps, not shipped product. I have excluded any stock recommendation; this is a technology and supply-chain structure analysis, not investment advice. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-silicon-photonics-interconnect Free to quote with attribution and a link to the original.