Near-packaged optics (NPO) is picking up design wins from co-packaged optics (CPO). SemiAnalysis calculated a hard constraint: at 95% attach yield per engine in a 32-engine package, compound assembly yield lands at 19% (0.95^32 ≈ 19.4%). In June 2026 it published this analysis and cut CPO volume expectations to 2027 for scale-out and 2028–2029 for production. The stock reaction was sharp: Applied Optoelectronics fell 17%, Lumentum 8%, in a single trading session.
NPO sits between two established architectures. Front-panel pluggable transceivers sit 15–30 cm from the switch ASIC; the DSP consumes 6–8W of a typical 800G module's 14–17W budget. CPO integrates the optical engine on the same package substrate, removes the DSP, and cuts optics power by 70%, per Broadcom. Nvidia's 1.6T link telemetry shows per-link power dropping from 30W to 9W. NPO places the optical engine on a separate socketed substrate beside the ASIC — close enough to eliminate the DSP but removable in the field. Linear pluggable optics (LPO) takes the opposite approach: no DSP, no power loss, 7–8.5W per 800G port.
| Architecture | Distance from Switch ASIC | DSP Present | Power per 800G Port | Field-Replaceable | Key Trade-off |
|---|---|---|---|---|---|
| Front-panel Pluggable | 15–30 cm | Yes (6–8 W) | 14–17 W total | Yes | High power; DSP overhead dominates |
| LPO (Linear Pluggable) | Front-panel | No | 7–8.5 W | Yes | No DSP savings path; signal integrity risk |
| NPO (Near-Packaged) | Adjacent, socketed | No | ~DSP-free budget | Yes (socketed) | Density vs. pluggable; socket interface spec in progress |
| CPO (Co-Packaged) | On-package (soldered) | No | ~9 W (from 30 W) | No | 70% power saving; zero rework path; yield risk |
Serviceability is the deciding factor. A soldered CPO substrate has no rework path. Removing a reflowed, underfilled engine requires applying 220°C–260°C within millimeters of the switch ASIC and every adjacent engine; sub-micron fiber alignment cannot survive a second thermal cycle. One failed optical engine condemns the ASIC, substrate, and every attached engine. NPO's socketed design confines failure to a single replaceable unit — the operational model that made pluggables the standard for a decade.
Broadcom and Nvidia are shipping both. Nvidia's Quantum-X Photonics InfiniBand switch carries 144 ports of 800G across 18 silicon photonics engines on detachable optical sub-assemblies with 18 removable external laser modules; Nvidia labels this as CPO, but the field-serviceability aligns with NPO. The Ethernet companion, Spectrum-X Photonics, targets 512 ports of 800G in H2 2026. Broadcom runs 51.2T Bailly CPO in volume at Micas Networks and began Tomahawk 6 Davisson (102.4T) customer shipments in October 2024. In March, at OFC 2026, it debuted a 3.2T VCSEL-based NPO line — a hedge for customers seeking density without soldered risk.
Standards work reflects both options. Ciena, Coherent, Marvell, Molex, Samtec, and TeraHop launched the Open CPX MSA at OFC 2026 to define a common socket interface for NPO and CPO. The OIF's ELSFP standard (August 2023) specifies a hot-swappable external laser module for both; lasers — the highest-failure optical component — sit outside the package in both cases.
CPO suppliers dispute the economic logic. GlobalSemiResearch rebutted that SemiAnalysis's yield math is pessimistic and ignores screening, binning, and spare-engine redundancy in Spectrum-X. Meta presented OFC 2026 failure-rate data showing CPO outperforming pluggables. Broadcom reported over one million cumulative 400G-equivalent port-hours in Meta deployments without a single link flap.
Market sizing shows momentum. Trendforce projects the combined CPO/NPO market at $100 million in 2025, expanding to $39 billion by 2030, with acceleration in 2028–2029. Pluggable transceivers reach $26 billion by 2030. For architects designing cluster fabric, the socket is the hedge. NPO's field-replaceability is not a consolation prize but an exit from a CPO supply-chain risk that has not been solved at production yield.
