Samil PwC says AI bottleneck is shifting to optical connectivity

AI infrastructure constraints are shifting from GPUs and HBM toward the networking layer needed to connect thousands to tens of thousands of chips, according to a Samil PwC report released on the 13th and a separate co-packaged optics market study added by ResearchAndMarkets.com. Both reports say copper-based electrical links and conventional pluggable optics face growing power, distance, density and cost limits in AI data centres, increasing the strategic importance of optical communications. Samil PwC describes a transition from rack-to-rack optical fibre links to Co-Packaged Optics, or CPO, and eventually optical GPU chip-to-chip interfaces, while the ResearchAndMarkets study frames CPO as a key interconnect architecture for both scale-out network switches and scale-up GPU and AI-accelerator optical I/O. Samil PwC said CPO has entered early commercialization and cited a market forecast of 38% compound annual growth to $580 million by 2030, while the newer market study says broader adoption will depend on solving manufacturing-test yield, field reliability, serviceability and qualification challenges, with large-scale NVIDIA CPO production potentially shifting to 2028-2029. Both reports cite industry moves by NVIDIA, TSMC and others, and Samil PwC says the shift could create openings for South Korean companies in optical cables, transceivers and longer-term silicon photonics and high-speed interconnect technologies.

The information on this website is generated using AI and we cannot guarantee its accuracy. Please use it as reference information only.