As artificial intelligence (AI) data centers grow in scale, optical communication interconnect architectures are changing rapidly. Alongside the existing pluggable optical module-centric approach, multiple methods are advancing simultaneously — from NPO (near-package optics), which places the optical engine close to the switch chip, to CPO (co-packaged optics), which combines the switch package with optical technology.

Lumentum recently assessed that as AI data center scale-out networks expand, optical interconnect technology is not transitioning to a single method but has entered a stage where multiple architectures coexist. A scale-out network is a structure that connects numerous accelerators and xPUs deployed inside a data center at rack, inter-rack, and cluster levels so they can perform a single task simultaneously.

As the number of semiconductors dedicated to AI training and inference grows, an environment is forming in which thousands to more than 100,000 accelerators exchange data within a single network. Accordingly, the scale of so-called east-west traffic within data centers is expanding to levels different from existing cloud environments.

■ 102.4T Switch Can Connect 128,000 Accelerators

Lumentum explained that using a 102.4T-class switch ASIC, approximately 128,000 accelerators can be connected with only a two-tier switch architecture. Reducing network tiers is advantageous for lowering latency that occurs during data movement. On the other hand, as the volume of data to process increases, the burdens of optical port density, power consumption, and heat management are growing as well.

In particular, during AI training, collective communication in which many accelerators exchange data simultaneously is repeated. For this reason, network congestion, communication latency, or link errors can directly affect overall computing performance.

Currently, 1.6T-class pluggable optical modules are used in AI scale-out networks, and the technology is at a stage where mass production of 200Gbps-class lane technology is expanding. Next-generation 400Gbps-class lane technology is also under development.

Lumentum explained that the spread of AI has not changed the fundamental principles of optical communication. Key requirements of existing optical communication technology — low bit error rate, mass production capability, stable performance under temperature changes, and cost reduction per bit — remain intact. What has changed is the total bandwidth and optical port density that data centers demand, because more data must be processed within the same rack while power consumption and cooling capacity are limited.

According to forecasts by market research firm LightCounting, AI scale-out and scale-up networks are expected to account for about 80% of high-speed optical interconnects shipped from 2026 to 2031.

■ Pluggable → NPO → CPO: Coexistence, Not Sequential Generational Replacement

The evolution of optical interconnect technology is also not proceeding all at once. While existing technologies below 100G and 100G-class lane structures are still widely used, 200G-class lane-based pluggable products have established themselves as the center of the current generation. This technology is regarded as the key step toward 1.6T-class connectivity.

The next step being discussed is NPO. NPO is a design method that places the optical engine close to the switch ASIC while still allowing module replacement and maintenance. CPO is a technology that integrates the optical engine inside or very close to the switch package, focusing on reducing the distance electrical signals must travel while addressing front panel space issues and power consumption limits.

Lumentum believes that rather than moving directly from pluggables to CPO, each technology is highly likely to be used together for a considerable period depending on the application. LightCounting forecasts that CPO is likely to be adopted first in scale-out networks rather than scale-up. However, for the remainder of this decade, pluggable optical modules are expected to continue handling a significant portion of scale-out traffic.

■ EML Shipments Up 8-Fold, Plans to Expand Production Capacity by an Additional 50%

In line with these changes, Lumentum is expanding its product portfolio to address multiple optical interconnect architectures. One of the key light sources used in pluggable optical modules is the EML. According to Lumentum, the company's EML shipments have increased more than eightfold since fiscal year 2020. The company has also announced plans to expand EML production capacity by an additional 50% or more by the end of fiscal year 2026.

In the silicon photonics structures used for NPO and CPO, the light source and modulator are separated. In this case, continuous-wave (CW) lasers supply steady light while a separate modulator applies the data signal. Lumentum supplies related CW lasers using its own indium phosphide production facilities.

When the optical engine moves closer to the switch ASIC, a structure that separates the laser from the optical engine is also used. The external laser source approach has the advantage of isolating the heat generated by the laser from the ASIC package while allowing the laser itself to be replaced in the event of failure.

To this end, Lumentum is developing an ultra-high-power (UHP) laser platform. According to performance data released by the company, the platform recorded optical output of more than 350mW in a 50-degree Celsius environment, with power conversion efficiency above 20%. The linewidth was specified at 500kHz or less, and relative intensity noise at -147dB/Hz or less under a maximum 400mW condition.

■ The Winner Is Not a 'Specific Method' but 'Adaptability and Production Capacity'

The key variable in the competition for the AI optical communication market is likely to be not only technological performance but also production capacity. This is because as shipments of accelerators and switches increase rapidly, optical component companies must also be able to supply in large volume while maintaining high yields and consistent performance. Rather than focusing on a single architecture among pluggables, NPO, and CPO, Lumentum is pursuing a strategy of expanding its supply capability for lasers and optical devices commonly used across multiple technology stages.

Ultimately, the optical interconnect market for AI data centers is highly likely to develop in a direction where different technologies are used in parallel according to network scale, power, heat, and maintenance conditions, rather than a single technology instantly replacing existing methods. Even as the optical engine moves from the front panel to the vicinity of the switch chip and then closer to inside the package, the fundamental conditions of optical communication — generating light stably, modulating it at high speed, and delivering it with minimal loss — do not change. Accordingly, in the future AI optical communication market, rather than predicting the winner of a specific interconnect method, companies that can respond to multiple architectures while securing large-scale production capacity are increasingly likely to hold competitive advantage.

[The market outlook for the AI data center and optical communication industries may vary depending on the pace of technology development, customer investment plans, and changes in semiconductor and network architectures. Companies' production capacity expansion plans and technological performance figures are subject to change based on future actual mass production and demand conditions. Before making investment decisions on specific companies or stocks, it is necessary to comprehensively review official disclosures, earnings, and business progress. This article is not intended to recommend the purchase or sale of any specific stock. This article was written with AI assistance.]