In 2026, discussion about CPO is no longer停留在 “whether it will come” stage. Coherent’s CPO and NPO routes have each secured more than ten long-term agreement customers, and its integrated optics platform is planned to enter mass production in the fourth quarter; NVIDIA is betting on the CPO ecosystem with multi-billion-dollar-level investment; although the penetration forecasts given by industry institutions start from a very low base, they point to a curve rising from less than 1% to one-third within a few years. These signals, taken together, can easily give buyers the impression that pluggable modules are becoming a “transitional technology” and that it seems not worth making long-term plans around them. But this impression ignores a key procurement reality: the transition period itself may be longer than imagined, and every capital expenditure and operations decision made during the transition period still needs to be taken seriously.
The first irreplaceability of pluggable modules in the face of CPO is fault-domain isolation. CPO packages the optical engine and the switch chip together, bringing structural gains in bandwidth density and energy efficiency, but it also binds the failure of optical components and the failure of the switch ASIC into the same physical boundary. For a data center carrying production traffic, the failure of a pluggable module means replacing a component on the front panel, while the failure of a CPO optical engine may mean a larger maintenance granularity and a longer downtime window. The “maintainability” of the pluggable form factor is not as impressive on paper as CPO’s integration level, but in actual operations it is an underestimated risk buffer. When evaluating CPO, buyers need to examine “fault-domain size” and “the definition of the field-replaceable unit” as dimensions just as important as power consumption and bandwidth.
The second practical constraint comes from the maturity cadence of the supply chain. CPO capacity ramp-up is not simply a matter of copying optical module production lines; it involves coordination among advanced packaging, silicon photonics wafers, external light sources, and system-level testing, and a bottleneck in any one link will be transmitted to final delivery. By contrast, although the supply chains for 800G and 1.6T pluggable modules have also experienced tightness in high-end DSPs and InP lasers, their production lines have already gone through years of volume validation, and yield learning and delivery discipline are relatively mature. For buyers that need to complete cluster expansion within the next twelve to eighteen months, betting entirely on CPO capacity commitments is irrational in terms of timing. A more robust approach is to use pluggable modules as the base of “deterministic supply,” while using a limited number of pilot ports to track CPO’s actual deployment performance.
Another value of pluggable modules during the transition period is that they preserve flexibility in architecture choices. The interconnect layers of AI data centers are diverging: the requirements of the Scale-out layer for scalability and distance, and the pursuit of bandwidth density and latency at the Scale-up layer, do not point entirely to the same optimal solution. CPO’s attractiveness in Scale-up scenarios does not mean that it will replace pluggables at the same speed across all interconnect layers. If buyers break down network architecture by layer, they will see that pluggable modules will remain the most pragmatic mainstay at the Scale-out layer for quite a long time, while CPO is better suited to penetrating first from the scenarios of leading customers that are most hungry for bandwidth density and have the most concentrated operations and maintenance systems.
Therefore, before CPO arrives, the “transition-period dividend” of pluggable optical modules does not lie in prices falling indefinitely, but in the fact that buyers still have the power of choice: they can choose gradient configurations across different speeds, choose to validate LPO or NPO first in closed scenarios, and choose to give greater weight to suppliers’validation capabilities and delivery records than to “technology labels.” Using this window period to establish a clear set of scenario-solution matching logic has more long-term value than chasing any single technology narrative.
