---
title: "The 1.6T Link AI Clusters Are Moving To: What Innolight's Numbers Actually Show"
date: 2026-11-22
category: Chips & Compute
site: NeuroAI
canonical: https://neuroai.site/a/na-chips-optical-transceiver-16t
language: en
---

# The 1.6T Link AI Clusters Are Moving To: What Innolight's Numbers Actually Show

> As AI clusters grow, the cables between servers become a constraint; a Chinese optical-module maker shows how central that unglamorous part has become.

Walk down the aisle of a data center and the most visible thing is light: tiny points blinking on the front of servers. Those points are the ends of fiber links, and they are doing a job that has quietly become as important as the processors themselves. If the chips are the brains of an AI cluster, the optical links are the nerves.

Innolight (中际旭创), a Chinese supplier of high-speed optical modules (光模块), has become a useful lens on this shift. Its published results show a company riding the move from 800G to 1.6T links, the generation now being deployed in AI data centers.

## Key takeaways

- Innolight says its products span 10G through 1.6T optical modules, with 400G, 800G and 1.6T as its focus range.

- The company reported 2025 revenue of ¥38.24 billion (≈ US$5.35B / HK$41.7B), up 60.25 percent, and net profit of ¥10.797 billion (≈ US$1.51B / HK$11.8B), up 108.78 percent.

- It says 1.6T modules began volume delivery in the second half of 2025, with demand continuing into 2026.

- The industry is also developing co-packaged optics (CPO, 共封装光学), which moves the optical engine closer to the switching chip.

- For readers, the point is that AI speed is a system property: a faster accelerator is worth less if data cannot move between machines.

## What an optical module does

Servers process electricity. Long-distance, high-bandwidth communication travels as light through fiber. An optical module is the translator at each end: it turns an electrical signal into a light signal for transmission and back into electricity on arrival.

That sounds like a small component, and physically it is. But multiply it by the number of links in a cluster and it becomes a dominant cost and power line. A modern AI training job is distributed across thousands of accelerators, which constantly exchange fragments of work. Every exchange passes through the network. If each link is slow or power-hungry, the expensive accelerators sit idle waiting for data.

This is why the industry measures progress in generations: 400G, then 800G, now 1.6T. Each step doubles the bits a single port can carry, which means fewer ports, fewer cables and less power for the same total bandwidth. Innolight says it demonstrated 1.6T products in 2023 and showed a 3nm 1.6T OSFP module at the 2025 Optical Fiber Communication conference. It also says it completed a single-channel 400G demonstration in 2025 — one lane of a multi-lane 1.6T design.

## The numbers, and what they mean

The company's own site reports 2025 revenue of ¥38.24 billion (≈ US$5.35B / HK$41.7B) and net profit of ¥10.797 billion (≈ US$1.51B / HK$11.8B). Its annual report, summarized in Chinese financial media, says 1.6T modules began shipping to key customers in the third quarter of 2025 and ramped faster in the fourth quarter.

A company growing that quickly is not selling a niche component anymore. It is selling a piece of infrastructure that large computing customers buy in volume. Trade publication reporting, citing research firm LightCounting, indicates Chinese suppliers took a substantial share of the global optical-transceiver market in 2025, with several of the largest vendors based in China. That is a statement about manufacturing and supply chains, not about the superiority of any one design.

The company has also disclosed plans for a Hong Kong listing. Financial figures cited here are from its published results and do not depend on share prices, which move daily.

## The next argument: pluggable or co-packaged

Today's modules are pluggable: small units that can be swapped from the front of a switch. That is convenient for maintenance and lets customers mix suppliers. The cost is power, because the signal has to travel from the chip, through the board and into the module before becoming light.

Co-packaged optics (CPO, 共封装光学) puts the optical engine much closer to the switching chip. Shorter electrical paths mean less energy per bit, which matters enormously at 1.6T and beyond. The trade-off is flexibility: a co-packaged engine is not something a customer can casually replace.

This is a genuine engineering disagreement, not a marketing slogan. Pluggable modules have a mature ecosystem and a huge installed base. Co-packaged designs offer a power advantage but require closer coordination between chip and optical vendors. A supplier like Innolight has to invest in both paths, because a sudden shift in customer preference would otherwise leave it holding the wrong inventory.

Chinese energy-sector reporting has framed this as a power race: as computing clusters expand, data centers consume more electricity, and the optical layer is one of the places where efficiency can still be won. That framing is about engineering constraints, not about any government target.

## Why a reader should care

AI services feel immaterial: a question goes in, an answer comes back. The physical reality is a building full of heat, electricity and fiber. The price and availability of optical links influence how fast a service can be expanded, how much it costs to run and how quickly a new model can be offered to the public.

There is a human dimension too. Data centers compete with other users for electricity and water. Every percentage point of network efficiency reduces the amount of power needed for the same computational output. That does not make a data center harmless, but it makes the difference between a tolerable and an unreasonable footprint.

Seen this way, the optical module is a good corrective to the habit of discussing AI only in terms of model size. Models are software. Software runs on a physical network, and that network has suppliers, factories and power bills.

## Honest limitations

This piece is an industry deep read based on the company's published figures and product statements, plus trade-press reporting of research-firm rankings. I have not audited Innolight's financials beyond what it publishes, tested its modules, or verified market-share figures independently. Performance claims about 3nm processes, single-channel 400G demonstrations and shipment timing are company statements. The article deliberately avoids any discussion of export rules, trade restrictions or regulatory proceedings, and it does not assume that one optical architecture will prevail. Share prices and stock-market reactions are excluded.

## What readers can do now

- When you read that an AI model is "faster," ask whether the claim is about computation or about data movement; both are required.

- If you work in procurement or IT, treat the network between accelerators as a first-class cost line, not an afterthought.

- Follow the power consumption of AI services as closely as their capabilities; efficiency improvements in components like optical modules are what make large-scale service economically possible.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-optical-transceiver-16t
Free to quote with attribution and a link to the original.
