--- title: "The 800G boom made two Chinese module makers the world's top suppliers — and exposed a chip gap" date: 2026-11-01 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-optical-module-ai-datacenter language: en --- # The 800G boom made two Chinese module makers the world's top suppliers — and exposed a chip gap > Innolight and Eoptolink now lead the global 800G optical transceiver market, but the lasers and DSPs inside those modules still come from US suppliers. A look at who owns the value chain in AI data-center interconnect. A server rack in a Virginia data center hums, and somewhere inside it a stream of ones and zeros leaves the copper traces and becomes light. That conversion happens inside a small metal box screwed into the front of a switch. Most of those boxes shipped this year were built by companies you have probably never heard of — and most of them are Chinese. ## Key takeaways - Two Chinese firms, Innolight (中际旭创) and Eoptolink (新易盛), rank first and third in the global optical transceiver market, with Chinese companies holding seven of the top ten slots in 2024. - The boom is driven by 800G modules for AI clusters; analyst forecasts see shipments of 800G-and-above parts roughly tripling from 2025 to 2026. - Module assembly is where Chinese firms are strong; the highest-margin components — EML lasers and DSP chips — are still controlled by US suppliers such as Broadcom, Lumentum, and Coherent. - The next architectural shift, co-packaged optics (共封装光学), moves the optics onto the chip package and could reshape who captures the profit. ## The box that turns electricity into light An optical transceiver is the device that converts electrical signals from a switch or accelerator into light pulses, sends them down a fiber, and converts them back. In an AI training cluster, thousands of these modules sit between the GPUs and the network fabric. As clusters grew from thousands to hundreds of thousands of accelerators, the volume of "plumbing" required to keep them talking exploded. This is why a quiet component category became one of the hottest segments in the semiconductor supply chain. The rate that matters most right now is 800G — eight times the capacity of the 100G modules that dominated a few years ago. ## The 2024 ranking that set the tone Market research firm LightCounting's 2024 global top-ten ranking showed Chinese vendors taking seven of the ten positions. Innolight (中际旭创) held the top spot with about US$3.3 billion in revenue, up roughly 114 percent year over year, while Eoptolink (新易盛) jumped from seventh to third with about US$1.2 billion in revenue, up roughly 175 percent. Both companies focused almost entirely on high-speed Ethernet modules for North American cloud providers — the fastest-growing slice of the market. The numbers behind those rankings are public. Innolight's 2024 audited results show revenue of 23.86 billion yuan (≈ US$3.3 billion) and net profit of 5.17 billion yuan (≈ US$720 million). Eoptolight's 2024 results show revenue of 8.65 billion yuan (≈ US$1.2 billion). These are exchange-filed figures, not estimates. The demand curve is steep. TrendForce projects shipments of 800G-and-above transceivers climbing from about 24 million units in 2025 to roughly 63 million in 2026 — a 2.6x jump. Very few manufacturers can deliver at that scale on schedule, which is why the leaders' positions look durable in the near term. ## Where the profit actually sits Open one of these modules and you find four main pieces: - A laser source (EML, VCSEL, or a continuous-wave laser) — the "engine" that makes light. - A photodetector that turns light back into electricity. - A digital signal processor (DSP) that cleans the signal and corrects errors. - The optics, printed circuit board, and packaging that hold it together. Assembly and packaging are exactly where Chinese module makers excel. But the two components with the fattest margins sit upstream. EML lasers are dominated by Lumentum and Coherent; optical DSPs by Broadcom and Marvell. Those are concentrated, high-barrier segments — a small number of US suppliers versus dozens of module assemblers worldwide. The assemblers compete on scale, yield, and on-time delivery; the chip and laser makers sit on design IP that is hard to replicate. That is where pricing power lives, regardless of any single quarter's margin. The contrast is structural, not just a matter of percentages: shipping the most boxes is not the same as controlling the components that decide the box's performance. Chinese firms assemble the 800G module; US firms supply the chips that define its capabilities. ## The next turn: 1.6T and co-packaged optics The industry is already moving to 1.6T modules, which double the per-port rate again and strain the power and tolerance budgets of pluggable designs. Beyond that sits co-packaged optics (共封装光学, or CPO), which removes the pluggable module entirely and integrates the optical engine next to the switch chip inside the package. LightCounting has forecast that CPO ports could approach 30 percent of 800G and 1.6T ports by 2027. That shift is a double-edged sword for Chinese assemblers. On one hand, it threatens to displace the pluggable-module volume that built their lead. On the other, transitional schemes such as linear pluggable optics (LPO) remove the DSP and play to their packaging strengths. According to Li An, chief scientist at BrainNet (脑机网), the move toward co-packaged optics will test whether Chinese module makers can convert assembly strength into chip-level control, rather than simply riding the volume wave. ## Honest limitations This article relies on company filings and analyst reports (LightCounting, TrendForce) that estimate market shares and shipment volumes; those estimates vary by a few points depending on the firm. We did not independently verify per-customer concentration such as the share of any single hyperscaler's orders, and we treat any such figure as an industry estimate rather than fact. The margin figures for chip suppliers are company-wide and not isolated to the optical segment, so they indicate direction rather than exact module-chip economics. Rapid price declines in 800G parts could change the financial picture faster than written here. ## What readers can do now - If you follow the AI infrastructure supply chain, track 1.6T qualification and CPO roadmaps from the switch-chip makers (Broadcom, NVIDIA, Marvell) — that is where the next margin shift will be decided. - Watch the gross margins of the laser and DSP suppliers versus the module assemblers; the gap between them tells you who really controls the profit. - For non-specialists, the takeaway is simple: the physical layer of AI compute is more geographically concentrated than the headline "GPU shortage" suggests, and that concentration carries its own supply-risk story. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-optical-module-ai-datacenter Free to quote with attribution and a link to the original.