When it comes to AI computing power investment in US stocks, most investors' first thoughts turn to NVIDIA's GPUs, the HBM memory arms race between SK Hynix and Micron, or TSMC's advanced process capacity. But inside the data center, there is another unassuming yet indispensable link — optical modules and optical interconnects. If GPUs are the "heart" of an AI data center, then optical modules are the arteries running throughout the body: they are responsible for transmitting massive amounts of data among thousands upon thousands of GPUs. Without them, even the most powerful computing power would be an isolated island. This article breaks down the logic of this "invisible artery" sector in plain language, and outlines the investment perspectives and risk points in the US stock market.

1. Why Do AI Data Centers Use "More Optical Modules Than GPUs"?

To understand the demand for optical modules, we must start with a basic physical fact: high-frequency electrical signals can only travel a very limited distance, with attenuation deteriorating sharply beyond a few meters, while GPUs in AI training clusters are often distributed across different racks, different rows, or even different server rooms. In other words, copper works within a server, but light is a must between servers — and the transceiver devices at both ends of each optical fiber are the optical modules.

Now look at the network architecture: large-scale GPU clusters generally adopt a Leaf-Spine fat-tree topology, and each time data passes through a switch "hop," a pair of optical modules is needed for transmission and reception. This leads to a counterintuitive conclusion: the number of optical modules in a cluster is often several times the number of GPUs. The larger the model parameters and the higher the parallelism, the more frequent the communication between GPUs — and the interconnect network, ironically, becomes the real bottleneck of training efficiency. The industry often says "computing power can be piled up, but interconnects are hard to scale," and optical modules are precisely the key variable in this "hard-to-scale" equation.

2. From 400G to 800G, and Then to 1.6T: Generational Iteration Is Accelerating

The core metric of optical modules is speed. In H100-generation clusters, 400G optical modules were the absolute mainstay; as new-generation platforms like Blackwell ramp up and per-GPU interconnect bandwidth doubles, 800G modules have entered the mass-production ramp-up stage. Notably, the iteration cycle of optical modules has been compressed from the historical four years per generation to about two years — driven by the white-hot AI competition: if interconnect bandwidth can't keep up, the utilization of GPU clusters built at a cost of hundreds of millions of dollars will be dragged down.

For next-generation platforms, 1.6T products have entered the industry chain validation and pre-research stage, and upstream segments

  • Challenges: Once the optical engine is co-packaged with the switch chip, field maintenance becomes more difficult, and the reliability of lasers, thermal management, and ecosystem maturity still need time to be proven;
  • Industry trends: Major switch chip and system vendors have already incorporated CPO into their roadmaps, and the penetration rate of silicon photonics solutions is expected to keep rising.

    • It is important to stress that CPO will not replace pluggable solutions overnight. Over the next few years, multiple routes—including pluggable optics (including the low-power LPO direction) and CPO—will evolve in parallel. For investors, the key is to track each company's technology bets and mass-production pace, rather than simply latching onto buzzwords.

    4. The US-Listed Optical Interconnect Sector: Which Companies Are Involved?

    On the global optical module shipment rankings, Chinese manufacturers have long held top positions; in the US stock market, however, the companies investors can directly access fall roughly into several categories:

    • Component and module makers: Coherent (COHR) is vertically integrated in lasers and optical modules; Lumentum (LITE) is strong in optical chips and lasers; Applied Optoelectronics (AAOI) focuses on data center optical modules;
    • Contract manufacturing: Fabrinet (FN) is the "picks and shovels" player in optical module assembly, and its order structure is often seen as a bellwether for industry conditions;
    • Switching and networking: Arista (ANET) is the leader in data center Ethernet switches; NVIDIA's networking business (InfiniBand and Spectrum-X Ethernet) bundles optical interconnects into its own compute moat;The essence of the AI race is a systems battle of "compute + storage + networking + power." With GPUs and HBM now fully priced in, the value of the networking layer is being recognized by an increasing amount of capital. Generational upgrades in optical modules, the commercial rollout of co-packaged optics (CPO), and the continued expansion of cloud giants' capital expenditures together form the medium-term logic of this track. For investors in US stocks, understanding optical modules means understanding the "other half of the story" of AI data centers — wherever data flows, value follows.

      Risk Disclaimer: This article is an analytical summary based on publicly available industry information and does not constitute any investment advice. US stocks are highly volatile; invest with caution.

      Detail Page Advertisement