[Relevant Companies Involved in This Article] TFC Communication, InnoLight Technology, YOFC (Yangtze Optical Fibre and Cable), Samsung, ASE, Lumentum, T&S Communications, Broadcom, Corning, Fujikura.
Mass Production Pacing of NPO and CPO
Although NPO and CPO are often discussed together, their industrial implementation paces have shown significant divergence. Currently, the relatively mature NPO in the industry can achieve a 30-40% lower power consumption compared to pluggable modules, while CPO achieves a 50-60% reduction. However, the difficulty of mass production varies drastically. Comparatively, NPO mounts the optical engine on the PCB adjacent to the ASIC, significantly shortening the transmission distance of board-level interconnect electrical signals. Most solutions no longer require DSPs within the modules. CPO goes a step further by packaging the optical engine and ASIC on the same substrate, further shortening the interconnect distance and thus reducing power consumption even more.
Another major advantage of NPO over CPO is that the optical engine is independently packaged and interconnected via the PCB, allowing for pluggable replacement. In contrast, CPO features substrate-level packaging where a single failure compromises the entire unit, necessitating a full replacement. For clusters with enormous costs, maintainability is critical. This significant distinction directly dictates customer selection logic. Based on recent new product releases and exhibition information in the industry, the current supply chain yield and maturity of CPO are insufficient, with high variance in mass production maturity. It is also deeply tied to specific chip manufacturers. Moreover, the significantly higher initial costs compared to standard optical modules remain a major concern for customers. On the other hand, NPO products have lower complexity, better maintainability, thermal management, and scalable delivery capabilities, making it the preferred volume ramp-up path for the current industry.
In early April this year, Google placed an order for approximately 12 million NPO products for its newly built AI clusters, with a total value of about USD 14 billion. Such a massive expenditure reflects confidence in the first-mover commercialization logic of NPO. Meanwhile, the author has also collected deployment timelines for NPO among leading domestic cloud providers. The fastest deployment will be completed in Q4 this year. 2027 will be a year of significant volume ramp-up, growing from about 7 million to a peak of 20 million to 30 million units (nearly a fourfold increase, noting that different statistical calibers are used and this represents only partial data). Furthermore, InnoLight Technology, an industry leader, explicitly stated in an institutional conference call that NPO is expected to be applied first in scale-up scenarios. CSP customers have already provided clear demand guidance, with mass production anticipated in 2027.
The mass production pace for CPO is also accelerating, albeit with a more complex pathway. NVIDIA announced the full mass production of its Spectrum-X silicon photonics CPO switch in August this year. Based on TSMC's COUPE platform, it debuted the world's first 200Gbps microring modulator. Its PIC monthly capacity is planned to increase from about 500 wafers to at least 25,000 wafers by 2028. Meanwhile, Broadcom's 51.2T Bailly CPO switch, assisted by Micas Networks in its mass production introduction, currently reduces power consumption by up to 70% compared to traditional pluggable optical transceiver modules. It has been delivered to hyperscalers like Meta for deployment validation.
Development of Core Upstream and Downstream Components
The industrialization process of CPO and NPO is simultaneously driving massive changes in the supply of upstream core components. For instance, the ELS (External Laser Source) is a core component of CPO. Typically, an ELS module consists of 4 or 8 CW laser source chips, 8 or 16 lenses, 4 or 8 isolators, and 4 or 8 polarization-maintaining (PM) fibers. The entire module is priced at around USD 500. Individual components are also expensive, with the cost side reaching nearly USD 350. A single CW laser source chip costs close to USD 32, directly accounting for half of the cost expenditure. Research on TFC Communication, a domestic A-share listed company, shows that its FAU (Fiber Array) and ELS products, which are directly supplied for NVIDIA's latest CPO switch applications, make it the only A-share company currently directly entering the switch BOM in NVIDIA's officially announced CPO supply chain.
The first-generation CPOs from Broadcom and NVIDIA are currently in stable mass production, and the R&D for the next one or two generations is also advancing rapidly. Coherent, with its independently developed CW laser chips, is expected to capture the largest market share. Meanwhile, TFC, InnoLight, and Eoptolink still need to procure from Lumentum. On the other hand, Lumentum, as the most renowned optical chip company in the industry, has entered mass production of its independently developed 400mW single-wavelength 1310nm ultra-high-power laser, with remarkably rapid progress.
Turning to polarization-maintaining (PM) fibers, according to industry demand guidance for 2027, 10 million ELS units can support 20 million 3.2T optical engines, corresponding to 20 million FAU sets. Each FAU set requires 4 PM fibers of about 1 meter each, bringing the total demand to nearly 80 million meters, commonly referred to in the industry as 80,000 fiber-kilometers. However, the current full-capacity production of industry leader Corning is less than 15,000 fiber-kilometers. The overall industry supply-demand gap is more than twofold, leading to continuous price increases for PM fibers. Meanwhile, an executive director of YOFC also stated at the World Artificial Intelligence Conference 2026, "Driven by the new generation of high-performance CPUs and silicon photonic components, the demand for PM fibers will grow by 10 to even 20 times in the next 1 to 2 years. PM fibers have become a rigid demand for AI optical interconnects."
Moreover, for the supply chain, stable quality and consistent lead times are crucial in customer considerations. According to minutes from a closed-door meeting of the North American supply chain, the biggest bottleneck in the mass production of the Spectrum-X CPO switch is the FAU fiber array. The yield of the "fiber threading process" for FAU arrays among well-known enterprises such as Senko Advanced Components, Sumitomo, Corning, and FOCI Fiber Optic Communications has always been a massive issue. For high-density 24/36-fiber products under a manual-equipment collaborative mode, the current stable mass production yield is entirely stuck in the 70% to 75% range. However, research on TFC Communication shows that its production line currently maintains a stable mass production yield in the 90% to 95% range. This highly precise assembly process means that even slight mating errors can damage the product. Yield control on mass production lines highly depends on the long-term experience and proficiency of operators. Currently, the quality of skilled workers from mainland domestic manufacturers can be said to be generationally ahead of those in Vietnam, Thailand, and other regions.
Potential Future Evolution of the Industry
As "star products" in the CPO mass production supply chain, CW laser sources and FAUs naturally attract significant attention. Similarly, another rising star is also receiving more focus. Broadcom released its latest VCSEL-based NPO solution at OFC 2026, showcasing the mass production of 100G-VCSEL 3.2T NPO and the R&D progress of the next-generation 200G VCSEL. Under 100G PAM4 transmission, the NPO solution allows for 6-13 dB insertion loss, with power compressed to 1pJ/bit. Most astonishing is its cost; mass production can achieve a cost of USD 0.1/GB, and it achieved "Zero failures" in continuous experiments of 53 million operating hours.
Meanwhile, VCSEL-NPO operates at a 1060nm wavelength, is less affected by temperature changes, and does not require additional temperature control and compensation mechanisms, inherently saving costs. At the same time, the industry's supply of indium phosphide (InP) remains continuously tight. Silicon photonic chips require a large amount of InP wafers, and short-term capacity crowding will cause costs to surge. In contrast, VCSEL can utilize the currently mature gallium arsenide (GaAs) process. Production lines can directly manufacture them after simple modifications, offering a significant cost advantage. Broadcom positions its VCSEL-NPO as the core transitional path from pluggable optical modules to the ultimate CPO solution, and it will not replace the mainstream status of silicon photonic CPO in the short term.
Conclusion
The pacing divergence between NPO and CPO, the supply bottlenecks in PM fiber and FAU yields, and the parallel exploration of VCSEL and silicon photonics routes all indicate that the supply chain is still in its early stages, with undefined standards, unstable yields, and unstarted capacity. The true watershed lies not in conceptual hype, but in who can best coordinate optics, electronics, thermal management, and power within the packaging, while stabilizing mass production yields in a deliverable range.
References:
"In-Depth Analysis of the CPO and NPO Supply Chain: Technological Landscape and Supply Chain Competition of External Laser Sources, FAUs, PM Fibers, and Passive Components", "OFC 2026 - Ultimate Power Consumption Challenges and Technical Route Decryption for Photonic Scale-Up Interconnects - OFC Panel", "LC North America Research Minutes - 2026 Q2", "Broadcom OFC Speech", "Coherent OFC Speech"