"A dime-sized piece of glass now sells for over six yuan." The Jiangxi optics factory in CCTV Finance's lens captures how AI compute is rewriting the communications supply chain: orders for high-speed optical-module components are booked through end-2027, with leading players' order books stretching toward 2028.

Golden optical signals traveling through fiber and optical modules
Golden optical signals traveling through fiber and optical modules

A 60x leap for a ten-cent filter

Per Cailian Press and Securities Times (Oct 19–20), the filter inside optical modules has evolved from 2.5G PON fiber-to-the-home parts to 1.6T optical-module components, with its value up more than 60x over 11 years. And it's not just component makers: Qichacha data shows China now has 13,700 optical-module-related enterprises, over half founded 10+ years ago, as the sector enjoys volume-and-price growth.

Three signals of the optical boom

· Component orders booked through end-2027; Zhongji Innolight and Yuanjie Semiconductor capacity races aim at 2028

· 1.6T modules enter their commercial first year with volume shipments

· Samsung plans to enter silicon photonics foundry by 2027 (DIGITIMES)

The industry logic is shifting tracks. First, speed transitions are beating forecasts as 800G gives way to 1.6T faster than any prior projection. Second, interconnect architecture is moving toward photoelectric convergence, with NPO (near-package optics) pulling the optical engine closer to switch silicon to crack power and signal-integrity bottlenecks. AI clusters' scale-out and scale-up appetite all becomes optical orders.

Blue laser light passing through an optical glass prism
Blue laser light passing through an optical glass prism

Compute runs, optics leads

Optical modules are AI compute's bloodstream: every GPU's delivered FLOPS depends on optical interconnects moving data between racks and clusters. As inference demand explodes with agent applications, optical prosperity no longer hinges on training capex alone — the biggest difference from the last cycle. Expansion still needs cycle discipline, though: fiber-cable players are already urging "rational capacity growth" to avoid another mismatch.

GPUs set the ceiling of compute; optical modules decide whether compute can reach itself.

NineZenith applies the same "compute-power-heat co-planning, optics over copper" principle in its AI-datacenter and industry-compute offerings: folding interconnect bandwidth, cooling and power into delivery design is what turns compute investment into usable intelligence services.

(Compiled from public reports by Cailian Press, CCTV Finance, Securities Times, Sina Finance and Eastmoney)