【Introduction】
As switch chips advance to 102.4T and interfaces evolve to 200G/lane, the pressures of power loss, consumption, and bandwidth density in electrical connections are pushing optics closer to the chip. However, the notion that "closer optics is always better" is easily misinterpreted: LPO/LRO refers to signal processing, while NPO/OBO/CPO refers to optical placement; they do not lie on the same substitution chain.
This article (Part 1) uses a two-dimensional map of "Optical Placement × Signal Processing" to dissect the true boundaries of four deployment forms: DSP pluggables, LPO/LRO, NPO/OBO, and CPO. It answers two questions: How far still need high-speed electrical signals travel? And how much power and cost is the system willing to pay to maintain link margin?
Part 2 will shift to manufacturing and operations: What eight types of evidence are missing for CPO to move from Production to scale replication, how the boundary between optics and copper will be redrawn, and a recalculable 3/5-year TCO framework.
Core Conclusion First: Pluggables will not disappear, and CPO will not dominate everything. The deployment boundary is determined by electrical connection length, manufacturing capability, and system TCO. Note: Terms such as "Demo/Release/Sampling/System Validation/Production/Mass Production" in this text refer to specific product-level statuses and do not represent the maturity of the entire technology roadmap. Corporate public disclosures do not equate to independent customer acceptance. For readability, some leading Chinese optical module companies are referred to as "Leading Optical Module Companies A, B, and C." Specific names and original disclosure texts can be requested from the author.
Part 01: Optical Interconnect Is No Longer a Unidirectional Substitution Path
For a long time, DSP pluggables, LPO (Linear Pluggable Optics), NPO (Near-Package Optics), and CPO (Co-Packaged Optics) have often been arranged as a path where "optics continuously moves closer to the chip." However, this ordering mixes two different dimensions: LPO and LRO (Linear Receiver Optics) primarily involve signal processing; NPO, OBO (On-Board Optics), and CPO primarily involve the location of electro-optical conversion. These four typical deployment forms are not four mutually exclusive standards.
Optical Placement answers "Where is the optical engine?" Front-panel pluggables connect to the host via standardized interfaces; OBO places the optical engine on the system board; NPO moves closer to the ASIC but retains structural boundaries distinct from co-packaging; CPO integrates optics and ASIC into a common package or substrate system. Optical I/O located even closer to compute or accelerator chips (collectively XPU) targets chiplets and packaging interfaces, with protocols, thermal management, testing, and maintenance conditions differing somewhat from switch-side CPO.
Signal Processing answers "How much digital recovery capability does the link require?" Fully Retimed pluggables use DSPs; LPO mainly reduces the full DSP retiming path within the module, relying more on host-side SerDes and linear opto-electronic link coordination; LRO typically implements transmitter-side retiming and receiver-side linearity, representing a partial retiming topology. As concepts of optical placement, NPO and CPO themselves do not mandate the adoption or removal of any specific type of retimer.
While optical placement and signal processing can be defined separately, they are not arbitrarily combinable in engineering practice. Electrical path length, host SerDes equalization capability, FEC requirements, packaging interfaces, and test architectures jointly constrain actual solutions. Moving closer to the chip generally helps compress high-speed electrical connections, but may increase the coupling degree of manufacturing and repair between optics and ASICs.
Consequently, next-generation optical interconnects must simultaneously answer two questions:
How far still need high-speed electrical signals travel?
And how much power, cost, and latency is the system willing to bear for signal recovery to maintain sufficient link margin?
Part 02: After 200G/lane, What Bottlenecks Do Old Architectures Face?
CPO was not proposed only in the last two years. The deep force driving architectural differentiation is the evolution of switch ASIC capacity from 25.6T and 51.2T to 102.4T, and high-speed serial links from 100G/lane to 200G/lane. In nominal port configurations, 51.2T can correspond to 64 800G ports, and 102.4T can correspond to 64 1.6T ports; however, increased port bandwidth does not automatically improve the signal integrity of PCBs, connectors, and module interfaces.
As symbol rates increase, insertion loss, reflections, crosstalk, and equalization pressure tighten the budget for board-level electrical channels. The industry has thus explored along two engineering paths:
Path 1: Enhance SerDes and DSP capabilities to continue supporting front-panel pluggables over longer electrical paths;
Path 2: Shorten high-speed electrical connections via OBO/NPO/CPO to reduce losses through physical architecture changes.
Both paths have device and system progress, but single product statuses cannot summarize entire technical directions.
Along Path 1, Broadcom’s Sian2, released in September 2024, adopts 200G/lane electrical and optical interfaces. Official statements claim support for 1.6T transceiver designs below 28W—this is a design capability supported by a specific device, not the measured average power consumption of all 1.6T modules.
Optical side bandwidth density continues to rise. In March 2026, Broadcom released the Taurus BCM83640, adopting a 3nm, 1.6T (8 electrical lanes to 4 optical lanes) architecture and 400G/lane optical interface. The announced Availability status is explicitly Sampling for early customers. This reflects improved optical side bandwidth density and cannot prove that 400G/lane electrical SerDes is already in mass production.
Parallel to device progress is the exploration of standards for the next speed grade. IEEE established a 400Gb/s/lane Signaling Study Group in March 2026; IEEE 802 LMSC approved related PAR and CSD documents in July of the same year. This constitutes project research and authorization progress and cannot be viewed as formal approval or widespread commercialization of the 400G/lane electrical standard.
Meanwhile, Coherent showcased 3.2T OSFP pluggable technology and 6.4T NPO optical engines at ECOC 2026. Publicly stated status remains technology demonstrations, not mass production deliveries. The simultaneous display of pluggable and NPO solutions reflects that the industry is still exploring high-speed pluggables and near-package solutions that shorten electrical channels in parallel.
On 102.4T and higher speed platforms, pluggables may continue product evolution, and the system value of NPO/CPO may become more prominent. Whether faster modules can be manufactured technically and whether they offer the lowest full lifecycle system cost are two judgments that must be kept separate.
Part 03: Pluggables Remain Competitive: The Real Divergence Lies in Link Margin
Data centers do not purchase a single optical module but rather a network system capable of continuous operation, expansion, replacement, and multi-vendor assurance. The value of front-panel pluggables stems partly from standardized interfaces and mature module testing processes, and partly from field replaceability, gradual deployment per port, and relative decoupling of modules from switch platforms.
In Scale-out networks, these system capabilities are particularly important. Scale-out connects numerous compute nodes and racks, with complex link distances, port configurations, and vendor mixes. Marvell’s public industry judgment suggests that Scale-out will remain primarily carried by pluggables for the next 8–10 years—this is a corporate forecast, not an industry-confirmed unified timeline.
In September 2026, the Ethernet Alliance organized multi-vendor interoperability demonstrations at ECOC, including 1.6T OSFP, 224G SerDes, and link training. Such events prove that the ecosystem is advancing cross-device integration, but do not directly equate to completed large-scale deployment across the entire 1.6T market.
Standard work also requires accurate stage labeling: IEEE P802.3dj D3.3 initiated its third round of standards body review voting on September 25, 2026, with a deadline of October 10 AOE. Voting deadlines do not mean formal approval and publication of the standard have been completed.
3.1 LPO Is Not Simply "Removing the DSP"
The concept behind LPO is to reduce full DSP retiming while maintaining the pluggable form factor, shifting more responsibility for link compensation and signal quality to the host side. This may lower module power consumption and digital processing latency, but interactions between host ASICs, PCBs, connectors, optoelectronic device linearity, temperature, and optical link losses will reflect more directly on system margins.
LPO MSA released the 100G-DR-LPO specification on March 25, 2025, and the 400G-FR4-LPO specification on September 25. Both belong to the 100G/lane system; 400G-FR4 refers to a total rate of four 100G channels, not a single channel at 400G. Linear solutions for 200G/lane remain a future direction.
Marvell demonstrated a 1.6T silicon photonics optical engine and a 200G/lane LPO reference design at OFC 2025; the sampling target mentioned in their announcement was the optical engine for selected customers, which cannot be directly equated to full LPO modules having formed large-scale procurement.
3.2 The Value of LRO Lies in Retaining Partial Recovery Capability
LRO adopts a typical "transmitter-side retiming, receiver-side linear" architecture. Compared to full DSP pluggables, it reduces some processing overhead; compared to fully linear LPO, it retains certain recovery capabilities. Its suitability depends on host SerDes, bidirectional channel loss distribution, error rate requirements, and system compatibility, and cannot be pre-set as the optimal compromise for all scenarios.
Commercialization stages cannot be broadly labeled solely by technology route. According to the semi-annual report of Leading Optical Module Company A for 2026, the company is one of the first to deliver 1.6T products in mass production and disclosed that it has launched and mass-produced LPO optical modules; however, this disclosure does not prove that LPO has reached the same level of maturity across all speed specifications and customer systems.
For the pluggable ecosystem, the future debate is not whether DSP or LPO is the only choice, but rather how full retiming, partial retiming, and linear implementations will divide applicable boundaries among power consumption, link margin, interoperability, and field operations.
Part 04: NPO/OBO: Intermediate Positions May Be Independent Choices
NPO/OBO attempts to strike a balance between front-panel pluggables and CPO: moving the optical engine to the system board or near the ASIC to reduce high-speed board-level electrical connections, without solidifying all optical functions into the ASIC co-package. They describe optical placement; specific products can still configure different drivers, equalization, or retiming architectures. See Figure 2 for details.
An intermediate position does not imply naturally simple engineering implementation. Once the optical engine leaves the mature front-panel module, the system must re-coordinate PCB layout, liquid cooling thermal design, FAU connections, fiber management, EMI, field replacement, and test boundaries. If key performance metrics can only be verified after integration into the whole machine, device manufacturers and system customers need to redefine quality responsibilities.
From public progress, NPO technology is at various stages including tech demos, releases, sampling, and system validation. According to an announcement by Leading Optical Module Company B on March 17, 2026, its 3.2T NPO solution completed sampling tests and full system validation with top domestic CSPs; the disclosed solution includes optical engines, external light source modules, and fiber management. This conclusion is a corporate public disclosure, not an independent customer acceptance record, nor can it be inferred as stable large-scale procurement.
NPO products from different companies are at different development stages. According to the 2026 semi-annual report of Leading Optical Module Company A, its 6.4T NPO solution was released in March; Coherent’s 6.4T NPO belongs to ECOC technology demonstrations. Product releases and technology demos correspond to different progress states, and a single maturity label should not summarize the entire class of NPO technology.
In the next phase, NPO needs to prove not only the effect of shortened electrical connections but also packaging and coupling yields, the degree of independent testability, repair operations, and cross-platform reuse costs. If optical engines can be modularly replaced and board-level coordination costs are controllable, they may form independent deployment value; if testing and repair costs increase significantly, their economic viability may be weakened.
Part 05: Conclusion of Part 1 and Preview of Part 2
Part 1 answered "Where should optics be placed, and how much signal processing should be retained." The conclusion is: The four forms are not mutually exclusive standards but rather different combinations of 'Optical Placement × Signal Processing'; the modularity and supply chain ecosystem of pluggables retain value in open Scale-out networks, NPO/OBO compete for intermediate positions, and CPO continues validation in high-density short-path scenarios.
But placement and architecture are only necessary conditions. The real watershed lies in manufacturing and operations: After co-packaging, combined yield, repair boundaries, and business interruption costs determine whether a route can move from Production to scale replication.