Pluggable, NPO, and CPO Optical Modules: Technical Comparison and Deployment Roadmap

Pluggable optical modules, near-package optics (NPO), and co-packaged optics (CPO) are three distinct optical interconnect architectures differentiated by the physical placement of the optical engine relative to the host ASIC. Pluggable modules house the optical engine in a field-replaceable front-panel transceiver; NPO mounts it on the PCB at centimeter-scale proximity to the ASIC; CPO co-integrates the optical engine and switch ASIC within the same package, eliminating the copper trace between them.

The Optical Internetworking Forum (OIF) has published formal interoperability specifications and framework documents for all three architectures, establishing CPO co-packaging and NPO near-package configurations as recognized optical interconnect design paradigms alongside conventional pluggable form factors. Understanding where each architecture sits in the deployment landscape — and why it sits there — is the central question for data center engineers planning infrastructure beyond 800G.

                                                             Pluggable vs NPO vs CPO

The clearest lens for understanding these three architectures is spatial. In a conventional pluggable deployment, the optical engine sits at the front panel of the chassis, connected to the switch ASIC via copper PCB traces that can span the full width of the board. NPO shortens that path to centimeters: the optical engine mounts directly on the PCB in immediate proximity to the ASIC, without crossing the full board. CPO eliminates the inter-component copper path almost entirely, co-packaging the optical engine and ASIC within the same module assembly.

This spatial distinction carries direct engineering consequences. At the lane speeds required by modern AI fabric interconnects, every centimeter of copper trace contributes to insertion loss, signal distortion, and the power overhead needed to compensate through digital processing. The architecture a deployment chooses determines not only its current operational profile but the engineering constraints it will carry as throughput requirements escalate.

Despite the growing industry attention given to NPO and CPO, conventional pluggable optical modules continue to hold more than 95% of the global data center interconnect market. That concentration does not reflect inertia. It reflects a set of compounding technical and operational advantages that alternative architectures cannot replicate on a short timeline.

                                                   Hot-Swap Field Replacement Scene

The most consequential factor sustaining pluggable dominance is manufacturing maturity. Yield rate — the percentage of manufactured units that pass all specification thresholds at first pass without rework — has reached approximately 95% for 800G pluggable products among leading suppliers. In optical module manufacturing, where component tolerances involve sub-micron alignment and assembly processes are tightly coupled across optical, mechanical, and electrical domains, a yield rate at this level represents years of accumulated process refinement. It translates directly into unit cost, supply chain reliability, and the ability to scale output in response to demand.

The hot-swap capability inherent to pluggable form factors adds a second operational dimension that is frequently underweighted in architectural comparisons. Field replacement of a pluggable module requires no scheduled maintenance window and can be completed in minutes by operations staff without specialized tooling. This characteristic makes pluggable infrastructure fully compatible with 7×24 data center operations — a practical constraint that matters considerably in production environments. Supply chain depth compounds this advantage further: multiple qualified manufacturers, globally distributed inventory, and a well-established component ecosystem ensure that availability and cost structures are well understood by procurement teams.

The financial performance of leading suppliers provides concrete evidence of the current commercial weight of the pluggable tier. Zhongji Innolight reported first-quarter 2026 revenue of approximately RMB 194.96 billion, representing 192% year-on-year growth, with net profit of approximately RMB 57.35 billion — up 262% over the same period. The company holds more than 40% of the global 800G optical module market and has completed certification of its 1.6T products on NVIDIA's GB200 platform, marking the transition of the 1.6T product cycle from qualification into active volume deployment.

Product iteration velocity has itself become a defining competitive dimension. The development cycle for successive speed generations has compressed from the earlier norm of three to four years per generation to approximately one to two years — meaning the pluggable ecosystem is not standing still while next-generation architectures mature. The 800G-to-1.6T transition is the current expression of this dynamic, and the demand trajectory in the 1.6T segment supports continued infrastructure investment in the pluggable tier through at least 2026.

Near-package optics occupies the architectural middle ground: measurably superior to pluggable on power consumption and signal performance, achievable without rebuilding the existing supply chain ecosystem, and commercially deployable within the current planning horizon. For data center operators evaluating what comes after 800G pluggable, NPO is the most credible near-term choice.

                                              PCB-Level Optical Engine Placement and Signal Path

The NPO architecture's defining characteristic is the centimeter-scale placement of the optical engine on the PCB adjacent to the switch ASIC. At the 224G per lane speeds specified under IEEE 802.3dj and the OIF CEI-224G electrical interface standard, this placement difference translates directly into measurable signal and power improvements.

At 224G channel rates, a conventional linear-drive pluggable (LPO) optical module incurs an insertion loss exceeding 20 dB along the electrical path from the ASIC to the front-panel module. An NPO implementation reduces that figure to approximately 10 dB — roughly half the pluggable LPO loss — by shortening the copper trace to centimeter scale and eliminating the signal compensation overhead that longer electrical paths accumulate. The result is lower overall system power consumption compared to a pluggable configuration at equivalent throughput, along with improved signal fidelity that reduces the error correction burden at higher lane speeds.

The operational model also differs from CPO in a way that matters for day-to-day data center management. NPO optical engines remain independently field-replaceable. A module failure can be addressed by swapping the NPO component without taking down the switch ASIC or initiating a whole-unit return — a meaningful operational advantage relative to the CPO maintenance model. The ecosystem transition path is equally important: moving from pluggable to NPO does not require dismantling existing supply chain, qualification, or deployment infrastructure. NPO builds on the pluggable ecosystem rather than replacing it, which substantially reduces integration risk for early adopters.

Alibaba's published technology roadmap places 3.2T NPO products in active testing, with pilot deployment targeted for the third quarter of 2026. The same roadmap targets completion of 6.4T NPO development by September 2027 — a schedule that compresses the NPO product cycle into the same one-to-two-year iteration rhythm already established by the pluggable tier.

On the manufacturing side, Huagong Technology has released both a 6.4T NPO product and a 12.8T XPO configuration, representing an early commercial commitment to the NPO product family across multiple throughput tiers. The company's 800G and above optical module overseas sales recorded year-on-year growth of approximately 13,974% — a figure that reflects the sharp acceleration characteristic of product categories entering volume production from a low prior-period base, and confirms a rapid, sustained demand inflection in high-speed optical module deployment globally.

Co-packaged optics represents the most fully integrated optical interconnect architecture among the three: the optical engine and switch ASIC share the same package assembly, eliminating nearly all of the copper electrical path between them. Its documented performance advantages are the most significant of the three approaches. Its path to commercial scale at volume is the most constrained.

                                          Co-Packaged Optical Engine and ASIC Integration

NVIDIA's internal testing of CPO switch configurations documents a 3.5× improvement in power efficiency relative to conventional pluggable equivalents, along with a 63× improvement in signal integrity — the fidelity of optical signal transmission across the channel, encompassing eye diagram quality, jitter margin, and waveform distortion characteristics. These figures represent NVIDIA's internal evaluation results; specific test conditions, baseline configurations, and comparison methodologies are not published in detail, and independent third-party verification has not been released. Even with that qualification, the directional magnitude of the performance differential is consistent with the underlying physics: shorter electrical paths require less compensation, and co-packaging enables electrical interface specifications that conventional front-panel geometries cannot support at equivalent scale.

CPO is widely recognized across the optical interconnect engineering community as the performance ceiling of current architectures — the design endpoint toward which successive generations of both NPO and pluggable improvements are incrementally converging.

The gap between CPO's documented performance potential and its present commercial availability reflects three distinct categories of constraint that compound each other rather than cancel.

Manufacturing yield is the most immediate barrier. A CPO volume production program by Tiannv Communication, originally scheduled for Q3 2026, was subsequently pushed back to Q4, illustrating the process engineering challenges involved in co-integrating precision optical components with switch silicon at the package level. NVIDIA's original internal target for 2026 CPO unit output was approximately 10,000 units; the industry's consensus estimate places actual 2026 shipments at several thousand units. In early-stage technology deployment, a plan-to-actuals delta of this magnitude is typically diagnostic of manufacturing readiness constraints rather than demand shortfalls — a signal that should inform near-term CPO planning timelines.

The CPO maintenance model introduces a second structural challenge. Conventional pluggable modules support a mean time to repair (MTTR) measured in minutes — field technicians can swap a module during live operations without affecting adjacent systems. CPO optical engines, co-integrated within the switch package alongside the ASIC, typically require whole-unit replacement or return to the manufacturer in failure scenarios. Repair cycles measured in days to weeks represent a materially different operational risk profile, with direct implications for redundancy architecture, spare unit inventory requirements, and availability guarantees.

Supply chain dynamics introduce a third layer of complexity that is largely independent of the technology itself. Cloud service providers have expressed concern about deep architectural dependency on a single CPO supplier, given that tighter co-packaging reduces the flexibility to source components from alternative vendors. Traditional optical module manufacturers, for their part, have shown a strong preference for NPO as the near-term transition path — one that preserves substantially more of their existing design and manufacturing infrastructure than CPO would require. The combination of buyer hesitancy and supplier positioning has extended CPO's commercial ramp well beyond what its technical performance metrics alone would suggest.

CPO's 2026 commercial entry does not follow a uniform market path. Two distinct network topology contexts carry very different deployment readiness timelines.

Scale-out networks handle inter-rack and inter-cluster connectivity across the data center fabric — the links that connect distinct compute nodes and racks within a large AI training environment. CPO is expected to begin small-volume deployment in this context in the second half of 2026, where its insertion loss and power advantages relative to pluggable are particularly compelling at the cable lengths and port densities involved.

Scale-up networks refer to the high-bandwidth intra-rack or intra-node interconnect between GPU and accelerator chips within tightly coupled compute clusters — the fabric that directly links GPUs during training operations, where bandwidth density and signal integrity requirements are the most demanding of any data center interconnect scenario. CPO deployment in Scale-up applications is not expected before the second half of 2028 at the earliest. Manufacturing maturity, ecosystem standardization, and the operational tooling needed for this environment require additional development beyond what the 2026 Scale-out deployment wave will establish.

The table below summarizes the primary engineering tradeoffs across all three architectures:

The three-year window from 2025 to 2028 is best understood as a structured technology handoff rather than a simultaneous multi-option decision for most operators.

Through 2025, the pluggable tier remains the undisputed primary deployment vehicle. 800G and early 1.6T products are in volume production, supply chains are mature, and no alternative architecture offers comparable manufacturing scale or cost structure. Investment in pluggable infrastructure during this period is not a failure to anticipate what comes next — it is appropriate allocation to the only architecture capable of delivering required capacity at required volume.

2026 marks the practical entry point for NPO at scale. The Alibaba roadmap commitment to 3.2T NPO pilot deployment in Q3 2026 confirms that a major hyperscale cloud operator has moved NPO from evaluation into production planning within the current calendar year. For operators evaluating 3.2T and higher throughput requirements, NPO becomes the technically and commercially credible option for new deployments.

CPO's commercial trajectory through 2026 and 2027 remains bounded by the manufacturing and ecosystem constraints described above. Small-volume Scale-out deployment from H2 2026 will generate the first systematic field data on CPO yield, reliability, and operational performance — data that will be essential inputs for any serious Scale-up CPO evaluation. Research from Bernstein projects LPO shipment volumes to exceed CPO even by 2030, with all three form factors — pluggable, LPO, and CPO — continuing to coexist across different deployment contexts for the foreseeable future. This framing from an equity research perspective reinforces a structural market reality: CPO does not eliminate pluggable. The three architectures serve distinct performance tiers, deployment horizons, and operational requirements, and will complement each other across most of the decade.

From 2028 onward, CPO is expected to become commercially viable for Scale-up network applications if current manufacturing development trajectories hold. The optical interconnect technology evolution is not an elimination race; it is a relay in which each architecture carries the capacity load appropriate to its maturity stage.

                                         Optical Interconnect Technology Timeline 2025–2028

The three-path structure of the optical interconnect market does not present a forced binary choice. It presents a sequencing problem, and the sequencing logic follows directly from the technical and commercial data reviewed above.

For infrastructure currently deploying or evaluating 800G, continued investment in the pluggable tier is technically justified through the near term. Yield rates, supply chain depth, hot-swap operational reliability, and the 1.6T product maturity cycle collectively ensure that pluggable infrastructure purchased today will deliver full performance value through the medium term. The opportunity cost of deferring 800G-scale pluggable investment while waiting for NPO does not exist in any practical sense.

For operators planning next-generation 3.2T and above deployments, the NPO evaluation window is open as of 2026. The Alibaba pilot deployment provides a reference architecture; the insertion loss advantage at 224G channel speeds provides a clear technical motivation; and the ability to leverage existing supply chain relationships reduces integration risk. Key signals to monitor include pilot deployment outcomes from Q3 2026 onward, multi-vendor NPO ecosystem development timelines, and qualification schedules from switch ASIC vendors incorporating NPO optical engine interfaces.

CPO planning in 2026 and 2027 is most appropriately framed as a forward-looking qualification activity rather than an immediate procurement decision. The monitoring signals that should gate any serious CPO deployment commitment are: manufacturing yield recovery data from the initial Scale-out programs (the H2 2026 deployments will produce the first field evidence), emergence of multi-vendor CPO supply chain options that reduce the lock-in risk documented by cloud operators, and standardized maintenance and sparing protocols that allow operations teams to manage CPO within conventional data center availability frameworks. Scale-up CPO is not a 2026 or 2027 deployment decision for most operators — it is a 2027-or-later evaluation, appropriately paced by production data from Scale-out deployments.

The most durable strategic framing for this landscape is portfolio management rather than technology selection. Each of the three architectures carries the capacity responsibility appropriate to its current maturity stage, and each will hand off a portion of that responsibility to the next as development timelines advance. Engineers who calibrate investment in each tier to that tier's actual commercial readiness — rather than front-loading commitment to unproven architectures or deferring necessary pluggable investment while waiting for CPO to mature — will be best positioned to capture the performance gains of the transition without absorbing the execution risk of premature adoption.