The Optical Networking Upgrade Cycle and What It Means for Samsung

The AI compute buildout has a bandwidth problem that doesn't get as much attention as the GPU shortage but is just as real. When you cluster tens of thousands of GPUs in a data center to train a large model, the bottleneck shifts from compute to communication: the interconnects between GPUs, between servers, and between data center pods need to move enormous amounts of data with low latency and high bandwidth. Copper interconnects, which dominate data center networking today, are approaching their physical limits at the distances and speeds required by AI infrastructure.

Optical interconnects — using light rather than electrons — have been the solution in long-haul telecommunications for decades. The push inside the data center is newer, driven by the bandwidth demands of AI workloads. Silicon photonics, which integrates optical components with silicon manufacturing processes, is the technology making this practical at scale: it allows optical transceivers to be manufactured at the cost points that large-scale data center deployment requires.

Samsung has meaningful exposure to this transition through multiple channels. Their semiconductor foundry business produces the ASIC chips that go into optical networking equipment. Their memory business supplies the high-bandwidth memory that optical transceivers increasingly incorporate. And their display and optics capabilities in related divisions create potential for vertical integration that competitors would struggle to replicate.

The investment thesis around Samsung in the context of the optical transition is essentially a second-order AI trade: rather than owning the AI model companies or the GPU manufacturers (both of which trade at aggressive multiples), owning a company that supplies critical components to the infrastructure buildout at a more reasonable valuation. This is the logic behind describing Samsung as a "beneficiary" of the data center optical upgrade rather than a direct AI play.

The risk in this thesis is that Samsung is not the only supplier of any of these components. Coherent, II-VI (now part of Coherent), and a set of Taiwanese ODMs are competitive in optical transceiver manufacturing. Intel has a silicon photonics product line. The competitive dynamics in components markets tend to be brutal when demand is high: customers diversify supply chains deliberately to prevent any single supplier from having pricing power. Samsung's advantage is scale and integration, but those advantages need to translate to actual contract wins to drive earnings.

The timing uncertainty is also real. Data center buildout plans have historically been volatile — demand forecasting for infrastructure is hard even with known AI scaling trends. Samsung's optical networking revenue is meaningful but not yet large enough to move the needle against the backdrop of their memory and logic semiconductor businesses, which are subject to their own commodity cycle dynamics.