What Is NPO? Architecture Differences from CPO, Adoption Examples, and Taiwan’s Supply Chain

Co-packaged optics (CPO) is one approach. Near-packaged optics (NPO) provides another: placing optical engines close to the main-chip package while preserving separate package boundaries.[1]

The difference extends beyond product appearance to who performs assembly, when testing occurs, what gets replaced after a failure, and how value is distributed across the supply chain. Understanding NPO requires examining architecture, commercialization, and specific supply relationships together.

1. NPO and CPO differ in where optical and electrical functions are integrated

An optical engine converts between electrical and optical signals. It may contain a photonic integrated circuit, drivers, receiver circuits, and control functions. How these are divided or integrated varies by technology platform.

A typical NPO design places packaged optical engines near the main-chip package, connected through short electrical links. They may sit on a PCB, daughterboard, or additional high-performance substrate. CPO integrates the engines and main chip within a shared packaging system, moving electrical interconnection further into the package.[1]

*This compares typical architectures. Actual products must be assessed from package cross-sections and interface designs.[1][2][5][14]*

Market terminology is not entirely consistent. Some NPO products are also classified as on-board optics (OBO). Assessment should examine package boundaries, actual electrical paths, and interfaces—not merely whether the engine and main chip share a substrate.

NPO also does not necessarily mean a socketed optical module. Lightmatter’s Passage L20, for example, uses a reflow-solderable BGA package, while replaceable NPO sockets represent another design category.[5][14] “Near-packaged” therefore does not guarantee direct field replacement, much less hot swapping.

2. Manufacturing can be separated, but CPO engines can also be tested before integration

One attraction of NPO is that the main-chip package and optical engine can be manufactured, packaged, and tested independently before board-level integration. This offers more flexibility in supplier responsibilities and assembly scheduling and helps screen out defective components before integration.[1]

Pretesting is not exclusive to NPO, however. NVIDIA’s public CPO technical materials describe screening optical engines before integration and using engines that have passed testing in subsequent assembly.[2]

*This compares integration levels, not a universal process sequence. Fiber-installation timing also varies by design.[1][2]*

The real difference is where integration risks concentrate. NPO leaves more challenges at board-level interconnection and assembly. CPO requires earlier coordination of optical, electrical, mechanical, and thermal design within the shared package. Even individually tested components can develop contact, coupling, or thermal-deformation problems after assembly.

3. Which system vendors have publicly adopted NPO?

“Using NPO” can mean either incorporating it into a switch or compute-system design, or a data-center operator actually deploying that equipment. Clear public examples exist for the former; the latter requires further evidence of customer identity, deployment scale, and delivery status.

The following are representative systems confirmed in official materials, not a complete market list.

These examples establish concrete NPO designs for switches and AI compute systems. A commercial launch, optical-engine mass production, and large-scale complete-system deployment are distinct milestones. The materials reviewed do not establish widespread, large-scale NPO adoption by major cloud operators.

4. Who is supplying or developing NPO optical engines?

Optical-engine suppliers are not necessarily the final system adopters. Separating these roles clarifies the ecosystem’s maturity.

NPO is not synonymous with one optical technology. It can use silicon photonics or other optoelectronic approaches, as Broadcom’s VCSEL demonstration illustrates. Lasers, fiber, reach, and manufacturing methods may differ, so one fixed bill of materials cannot describe all NPO products.[6][8]

Broadcom also develops CPO and stated in a 2025 announcement that Tomahawk 5–Bailly CPO systems had entered mass production.[7] The industry is therefore not neatly divided into NPO and CPO camps: one company may pursue multiple approaches for different system needs.

5. Breaking down the NPO supply chain: what has public support?

The supply chain includes photonic chips, electronics, light sources, fiber connections, electrical connections, packaging, assembly, and testing. The table includes products, platforms, and services explicitly mentioning NPO, rather than treating general optical capabilities as NPO orders.

*Public products or capabilities differ from confirmed supply relationships. Clearer named partnerships or product links include OpenLight–Tower, Ranovus–Jabil, and the Corning interconnect listed in Passage L20 specifications.*

These materials indicate that when similar optical technologies are used, NPO and CPO share substantial upstream capabilities; the main differences arise in packaging integration, electrical connections, and system assembly. Public information is insufficient to convert that overlap into a reliable supply-chain percentage.

Likewise, silicon-photonics or high-speed interconnect capabilities at a foundry, packaging company, or laminate supplier do not independently establish participation in a specific NPO product. Analysis must return to named partnerships, specified adoption, qualification, and shipment announcements.

6. Taiwanese participation: optical engines, sockets, and external light sources

For Taiwan, NPO’s significance extends beyond international chip vendors’ architecture choices to the products Taiwanese companies can supply. Public disclosures span optical engines, optoelectronic platforms, electrical connections, external light sources, and packaging development, with progress ranging from research and demonstrations to sampling and trial production.

Seven Taiwanese or Taiwan-affiliated companies with explicit NPO product or development information are listed below. This is a participation list, not a list of suppliers already shipping in volume. Their supply-chain roles also differ.

*Lintes, FIT, RVi, and Centera are supported by official announcements or presentations. Elite Advanced Laser, LOTES, and ShunSin disclosures include media records of executives’ public statements. Listed speeds refer to different interfaces or module capacities and cannot directly rank technical capability.*

Concrete examples of overlap with Taiwan’s CPO supply chain

RVi explicitly says the same optical-engine product line supports NPO and CPO deployments. Elite Advanced Laser says its external laser sources address both architectures. These examples show that some optoelectronic technologies, packaging, and manufacturing capabilities can carry from CPO into NPO, without implying identical component specifications or assembly.[21][23]

Centera and Elite Advanced Laser also provide a named division of responsibilities. Centera’s official presentation describes Centera designing CPO external light sources for switch vendors and Elite Advanced Laser manufacturing them. This public relationship clarifies design and manufacturing roles, but cannot establish the same arrangement or shipment status for all Centera NPO projects.[22]

The opportunities identified here include optical-engine design and packaging, external light sources, and board-level optical / electrical connections. Socketed NPO interfaces in particular must meet high-speed signal, mechanical-contact, and thermal-integration requirements. Lintes’ and FIT’s published solutions show concrete activity in this segment.[13][14][20]

Accton and Alpha Networks: standards participation, with products still to be tracked

LOTES and FIT are also listed as Open CPX participants.[17] For these companies, record standards participation, product announcement, completed qualification, and actual shipments separately. This identifies the problem each supplier addresses and avoids treating one AI-optics theme as an identical realized opportunity for every participant.

7. NPO bottlenecks: electrical channels, thermal-mechanical design, and production testing

Moving the optical engine near the main chip still leaves a high-speed electrical channel. Traces, vias, solder joints, and socket contacts affect loss and reflection. Every detachable interface also adds contact and mechanical conditions requiring validation. TE’s NPO socket design treats high-speed signals, compression force, and overall structure as interrelated issues.[14]

Whether NPO lowers system cost therefore cannot be determined from optical-engine price alone. Shorter electrical paths may save power, but sockets, board materials, assembly precision, and testing add costs. The engineering assessment here is to compare the complete channel and system rather than one component.

PCBs need low loss; thermal conductivity depends on the heat-flow path

“Low electrical properties” is not a precise description of NPO board requirements. More useful measures are low dielectric loss (Df), a stable dielectric constant (Dk) suited to impedance design, and the effects of copper foil and board processing on high-speed channels. Material selection must also account for dimensional stability, moisture absorption, heat resistance, and reliability. High-speed materials documentation lists electrical, thermal, and mechanical properties separately.[19]

Low loss, high-temperature resistance, and high thermal conductivity are different properties. Engines near high-power chips need effective thermal management, but heat may primarily leave through structures above the package or a cold plate. Passage L20 specifications, for example, list cold-plate cooling.[5]

NPO therefore does not universally require high-thermal-conductivity PCBs. Conductivity direction, heat-source location, cold-plate contact, and optical-engine temperature tolerance all affect selection. Some locations need efficient heat removal; others need to limit heat transfer from the main chip into the engine. Establish the heat-flow path before choosing materials.

Replaceability still requires thermal-cycling and assembly validation

For socketed NPO, contact pressure, substrate warpage, thermal-expansion differences, and fiber tension can affect long-term stability. GIGALIGHT also identifies thermal-expansion matching and compression structures as engineering challenges.[9][14]

“Removable” and “easy to maintain” are consequently different achievements. Systems must preserve electrical contact, optical connectivity, and cooling after removal and reinstallation. Soldered NPO likewise requires control of reflow, optical coupling, and final assembly yield. Pretesting reduces integration risk but cannot replace complete-system validation.

8. Shared standards are emerging, bringing competition toward system integration

Supply-chain overlap appears in standards activity as well. In September 2026, Open CPX MSA released specifications covering both architectures, including 6.4T and 7.2T connection formats, internal or external light sources, and liquid-cooling designs. Founding members include Ciena, Coherent, Marvell, Molex, Samtec, and TeraHop.[17]

OIF also lists a 12.8 Tb/s NPO work item, planning implementation agreements for 12.8T and 6.4T optical engines using 200G-per-lane technology.[18] These efforts help establish common interfaces, but publication of a specification and completed interoperability validation across vendors remain different stages.

The public examples suggest that NPO’s value lies in enabling optical engines, main-chip packages, and system cooling to be configured more independently. This creates opportunities for board-level design, connectors, fiber management, and assembly testing to play larger roles in optoelectronic interconnect upgrades.

NPO and CPO share many underlying technologies, but their commercial competitiveness ultimately depends on the complete system’s balance among power, density, manufacturing yield, and maintenance costs.