en.Wedoany.com Reported - Marvell Chairman and CEO Matt Murphy pointed out at the recently held Computex conference that the true bottleneck of AI infrastructure is no longer computing power or memory, but connectivity. NVIDIA CEO Jensen Huang agreed with this assessment at the event, emphasizing that under the "decomposed and distributed" computing model of AI agents, connectivity has become a necessity.
Over the past three years, global AI infrastructure investment has been heavily concentrated in GPU procurement and HBM memory stacking, leading to a consensus in the industry that "computing power is the moat." However, as large model inference and agent applications drive exponential growth in data center scale, data flow has surged, with bandwidth and latency demands far exceeding the coverage of single-point computing power improvements. The system bottleneck has shifted to the collaborative efficiency of clusters with tens of thousands or even millions of cards, which is essentially a network connectivity issue requiring high-bandwidth, low-latency networks to solve the connectivity challenges in large-scale distributed computing.
Addressing this connectivity challenge requires a full-stack layout covering everything from millimeter-scale packaging inside chips to cross-data center distances spanning thousands of kilometers, involving different physical layers, protocol layers, and packaging technologies. This includes intra-rack or inter-node interconnects (scale-up), intra-data center networks (scale-across), and long-distance coherent optical transmission across data centers (scale-out). The core trend in full-stack evolution is "optical replacing copper," breaking through the physical limitations of copper cables. When bandwidth doubles, the effective transmission distance of copper cables is halved, severely compressing the survival space of copper cables within racks, as optical interconnects penetrate from backbone networks into rack interiors.
Although optical interconnects are the ultimate direction, fully achieving all-optical interconnects still faces constraints such as high costs and incomplete supply chains. Jensen Huang stated at the conference that copper should be used as long as possible, and optics should be used where necessary. Over the next five to ten years, data centers will simultaneously use large amounts of copper and optics. In the short term, racks will still rely on high-performance copper interconnects, but as the industry evolves toward 1.6T and beyond, technologies like co-packaged optics (CPO) become key to breaking through density and power consumption bottlenecks.
The global trend of "optical replacing copper" is accelerating into growth momentum for China's optical communication industry chain. In terms of optical fiber and cable, AI computing power demand is driving growth in ultra-low-loss fiber and hollow-core fiber. China's "Big Four" fiber companies generally achieved high revenue and net profit growth in the first quarter of 2026, with related stock prices and market capitalizations continuously hitting new highs. In optical modules, Zhongji Innolight has become a core beneficiary of AI computing power, with large-scale shipments of 800G high-speed optical modules, leading to a dozens-fold increase in stock price and a market cap exceeding 1.4 trillion yuan. In optical components, TFC Communication successfully achieved mass production of 1.6T optical engines and completed R&D for CPO supporting components in 2025, achieving sustained and stable growth. In optical chips, Yuanjie Technology achieved batch delivery and scale revenue for chips at 25G and above rates, with product technical indicators comparable to international manufacturers, and its stock price also surged dozens of times.
Despite the high enthusiasm in the capital market, China's optical communication industry chain still relies on imports for high-end optical chips and high-end components.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
