Networking Industry Pivots to Near-Packaged Optics as CPO Yield Risks Mount

The semiconductor and networking industries are increasingly adopting Near-Packaged Optics (NPO) as a strategic hedge against the manufacturing and reliability risks of Co-Packaged Optics (CPO). This shift comes as the AI boom drives an urgent need for high-bandwidth, low-power interconnects that traditional pluggable transceivers can no longer efficiently provide.

To address these demands, NPO architecture places the optical engine on a separate, socketable substrate located beside the ASIC. This design reduces the electrical path length, which eliminates the need for power-hungry Digital Signal Processors (DSPs) typically required to push signals across longer copper traces. Unlike CPO, where components are permanently integrated, NPO remains field-replaceable, allowing operators to swap failed engines without discarding the entire switch.

The Power-Reliability Trade-off

The industry's interest in CPO is driven primarily by extreme power efficiency. For example, Nvidia's 1.6T link power drops from approximately 30W in a pluggable configuration to just 9W when using CPO. However, this efficiency comes with a significant "blast radius" risk. Because CPO engines are reflowed and soldered directly to the package, a single engine failure can potentially condemn the entire switch ASIC and substrate, as the heat required for rework is extreme.

This reliability gap has led analysts to push back volume expectations for CPO. The primary concern is compound yield; when multiple optical engines are integrated into a single package, the probability of a total package failure increases significantly if any single engine is defective.

Standardizing the Middle Ground

In response to these challenges, major industry players are moving to standardize the "middle ground" offered by NPO. At OFC 2026, Ciena, Coherent, Marvell, Molex, Samtec, and TeraHop launched the Open CPX MSA. This agreement aims to standardize socketed optical-engine interfaces for both NPO and CPO, ensuring interoperability and reducing vendor lock-in for the next generation of AI clusters.

Broadcom also showcased its commitment to this path at OFC 2026, presenting a 3.2T VCSEL NPO product line specifically targeting short-reach multimode (50m) scenarios, alongside a 3.2T silicon photonics single-mode NPO product. These offerings cater to buyers who require high density but are unwilling to accept the permanent commitment of soldered CPO components.

Implications for AI Scaling

The transition to CPO is a critical bottleneck for scaling massive AI clusters. If the industry cannot resolve the yield and serviceability issues inherent in fully integrated optics, the reliance on NPO will extend the lifecycle of current silicon photonics supply chains. This delay in the adoption of the most power-efficient architectures could impact the deployment timelines of next-generation AI switches and shift the valuations of optical component manufacturers.

Industry observers are now watching whether the Open CPX MSA can accelerate the adoption of socketed solutions. While CPO remains the theoretical gold standard for power, NPO provides a pragmatic, serviceable alternative that allows the industry to scale AI infrastructure without risking catastrophic hardware losses.