Optical connections emerge as the new battleground in AI infrastructure

As AI data centers grow in scale, the bottleneck shaping AI competitiveness is shifting away from securing semiconductors such as GPUs and HBM chips toward the "connectivity" that links them together. Analysts say the network and power efficiency needed to bind thousands — or even tens of thousands — of GPUs have emerged as the new battleground of the AI era.

Samil PwC announced Thursday that it has published a report titled "The Bottleneck Is Now 'Connectivity': Optical Communications in the AI Era and Opportunities for Korean Companies." The report analyzes the transition in AI infrastructure data transmission from conventional copper-based electrical connections to optical communications, and outlines the implications for Korean companies.

According to the report, conventional copper cables become less power-efficient as network bandwidth increases. In the high-bandwidth environments used by major AI data centers today, the effective transmission range of copper can be limited to just a few meters. As AI data centers grow larger, the distance between server racks also increases, causing electrical signals to weaken in transit.

The report introduces optical communications technology — gaining traction as an alternative to electrical connections — in three stages.

The first stage involves deploying optical communications for long-distance connections between racks to achieve low-latency, low-power environments. Data is processed as electrical signals but converted into light and transmitted along optical fiber during the transfer phase. Optical fiber cables have very low signal attenuation, enabling high-speed transmission over medium and long distances. Optical transceivers handle the conversion between electrical and optical signals.

The second stage is co-packaged optics (CPO) technology, which integrates optical engines into the same package as switch chips, drastically reducing the electrical signal path inside equipment to improve power efficiency. The report said CPO has entered the early commercialization phase and projected the global market to grow at an average annual rate of 38 percent, reaching $580 million by 2030. The third stage involves converting the interface between GPU chips to optical signals — a next-generation technology seen as capable of overcoming the physical distance constraints of computing chips and transforming AI system architecture.

Major AI infrastructure companies are adopting optical communications technology to maximize data connectivity, processing speed and efficiency, the report said. Nvidia formalized its optical communications strategy in March by investing $2 billion each in optical technology firms Lumentum and Coherent, while TSMC, the world's leading foundry, is building a manufacturing base for CPO components including silicon photonics chips.

Korean companies are primarily active in the optical cable and optical transceiver segments. Taihan Fiberoptics has secured cost competitiveness through vertical integration and signed supply contracts for optical cables destined for AI infrastructure in the United States, while OE Solutions and Lightron Fiber-optic Devices are developing optical transceiver products to capitalize on market expansion.

Meanwhile, the report raised the possibility that advances in optical communications could shift AI systems from a "density-focused" model to a "flexible, distributed architecture." In a high-speed optical connection environment, the constraint of physically co-locating GPUs and HBM chips would gradually ease, potentially changing how memory is deployed and how systems are designed. The report advised Korean memory chip companies to proactively explore memory utilization structures and architectures optimized for optical communications environments, and to develop forward-looking research and strategy.

"Just as Korea created an overwhelming lead in memory semiconductors, there is every reason to believe that specialized areas where Korea's strengths can shine will emerge in optical communications as well," said Lee Seung-hwan, a partner and AX Node leader at Samil PwC. "We need to leverage our accumulated semiconductor capabilities and precision manufacturing competitiveness to expand our technological reach into adjacent areas such as silicon photonics and high-speed interconnect chips."

Samil PwC, which closes its fiscal year in June, posted sales of 1.11 trillion won ($784 million) for the year ended June 2025, up about 8.4 percent from the prior year. Operating profit for the same period surged 55.6 percent year-on-year to about 25.4 billion won, reflecting both top-line growth and improved profitability.