A fierce debate surrounding optical modules and Co-Packaged Optics (CPO) has intensified since early 2026, sparked by a series of incremental updates on CPO from the prominent North American AI industry analyst "AYZ" in early February, alongside the recent release of earnings reports and conference calls from companies such as Fabrinet, Lumentum, and Coherent. The intensity of this debate rivals the fervor seen after the initial launch of the Blackwell B200 rack, which ignited discussions over "optical replacing copper" versus "copper replacing optical." Taking the popular stock trio "Yi Zhong Tian" as an example, significant divergence in their stock performances has been observed from the start of the year through February 8, 2026. Despite Suzhou Tfc Optical Communication Co.,Ltd. (300394.SZ) issuing a Q4 2025 earnings preview that slightly missed expectations, it has delivered the strongest stock performance year-to-date. In contrast, industry leaders Zhongji Innolight Co.,Ltd. (300308.SZ) and Eoptolink Technology Inc.,Ltd. (300502.SZ), despite reporting impressive earnings, have seen their stock performances fall slightly short of market expectations.
What is the market debating? The origin of the dispute can be traced back to an article published by "AYZ" on Substack on February 2, titled "Co-Packaged Optics (CPO) – The Next Big Thing for NVIDIA (2)?". The article stated that NVIDIA's ultimate vision is to create an OIO (Optical I/O) all-optical interconnected AI server rack, adopting a gradual strategy: first, launching CPO switches for Scale-Out scenarios in 2025; second, incorporating CPO switches in the NVSwitch trays for Scale-Up scenarios; and third, implementing CPO packaging on GPGPU chips, ultimately realizing the OIO rack. What truly captured market attention was AYZ's direct provision of a schematic diagram for the Vera Rubin Ultra rack under the "Scale-up CPO"方案 and the assertion that the Vera Rubin Ultra would utilize an NVL72 x 2 configuration. Each rack would retain the existing NVL72 rack design, with compute trays and switch trays connected by copper cables. Then, the NVSwitch trays of two adjacent cabinets would be interconnected via CPO and optical fibers, forming a Vera Rubin Ultra rack housing 144 GPGPU cards (or 576 GPU chips), with each Vera Rubin Ultra GPGPU containing four GPU chips.
Furthermore, AYZ indicated in the article that CPO would also be deployed in the Scale-Out scenario of the Vera Rubin Ultra rack. Specifically, each CX10 network card in the compute tray would be co-packaged with a 3.2T optical engine, then externally connected via fiber to NVIDIA's Quantum-X or Spectrum-X series CPO switches. Each compute tray contains 4 CX10 network cards (paired with 4 Rubin Ultra GPGPUs) for Scale-out, meaning the entire Vera Rubin Ultra rack incorporates 144 CPO CX10 network cards and 3.2T optical engines. Overall, under the CPO方案 for the Vera Rubin Ultra rack, the Scale-Up scenario utilizes 648 optical engines, and the Scale-Out scenario uses 144 optical engines, corresponding to 144 GPGPUs per rack, resulting in an optical-engine-to-GPGPU ratio of 5.5 to 1.
Subsequently, AYZ further stated on Twitter that for the Vera Rubin rack, NVIDIA would mandate the inclusion of 2 Quantum-X or Spectrum-X series CPO switches per rack. For the Vera Rubin Ultra rack, NVIDIA would mandate an additional 12 CPO NVSwitch trays per rack. Based on this, AYZ projected Vera Rubin rack shipments of approximately 10,000 units in 2026, corresponding to about 20,000 CPO switches (with actual shipments potentially between 15,000 and 20,000 if CPO yields are insufficient). For 2027, combined shipments of Vera Rubin and Vera Rubin Ultra racks are forecasted to be around 100,000 units, corresponding to at least 200,000 CPO switch shipments.
What is the market debating? The rebuttals and support from optical communication companies. AYZ's incremental information about NVIDIA's potential large-scale deployment of CPO in both Scale-Up and Scale-Out scenarios for the Vera Rubin racks naturally sent ripples through the already tense capital markets. First, on the evening of February 2 (Beijing Time), Fabrinet, often considered closely affiliated with NVIDIA's Mellanox, released its FY2026 Q2 earnings report, seemingly providing a boost for CPO. During its earnings call, Fabrinet stated that while CPO had long been perceived as a technology for the "future," the company now feels "the future is here." Fabrinet has already recognized some revenue related to CPO, albeit relatively small amounts currently. The company believes it holds a significant lead in technology and market positioning for CPO commercialization compared to most competitors and is jointly developing CPO projects with three different customers. The timing for meaningful revenue acceleration depends on customer roadmaps and schedules. Overall, due to non-disclosure agreements, the company was reluctant to disclose specific customer timelines for CPO ramp-up but expressed great anticipation and excitement for its scaling.
Second, on the evening of February 3 (Beijing Time), Lumentum released its FY2026 Q2 earnings report, further fueling the CPO discussion after Fabrinet's comments. Lumentum stated during its earnings call that it expects approximately $50 million (possibly more) in CPO-related orders (primarily for CW laser sources, etc.) to convert into revenue in the fourth quarter of this year, followed by several hundred million dollars (specific scale undisclosed) in CPO-related orders converting to revenue (i.e., shipping) in the first half of 2027.
Concurrently, Lumentum indicated that it expects significant volume ramp-up for CPO in Scale-Up scenarios by the fourth quarter of 2027. Looking ahead, it is reasonable to speculate that CPO penetration in Scale-Up scenarios will increase rapidly, eventually completely replacing copper cables.
Third, around noon on February 4 (Beijing Time), some A-share listed companies provided very clear and specific assessments on the matter, effectively pouring cold water on the CPO excitement. These companies argued that technology transitions require long lead times (typically over a year) for planning, sampling, and testing, and they have not observed such signs currently. The CPO方案 is primarily driven by chip manufacturers (like NVIDIA and Broadcom) aiming for tight coupling with GPUs within the rack, potentially being prioritized for customers within their ecosystem, which differs from the mainstream choices of CSP (Cloud Service Provider) clients. Communications with major CSP customers confirmed that their network plans for 2026-2027 are centered on pluggable solutions, with no indication of any CSP planning large-scale CPO deployment in the current year, next year, or even the year after. Additionally, these A-share companies believe that pluggable optical module technology will continue to evolve. They anticipate that even in the 3.2T era post-2028, pluggable optical modules will remain the dominant solution for Scale-Out scenarios. (The 3.2T era will see coexistence of EML laser solutions and silicon photonics solutions; single-wave 400G EML is currently in the testing phase, but silicon photonics holds advantages in cost and production scalability and is expected to be a significant technological choice.)
Furthermore, leading brokerages like CITIC Securities and GF Securities have expressed firm conviction that CPO will not undermine the fundamental position of optical modules in Scale-Out scenarios, arguing that the optical module sector has been severely mispriced by the capital markets.
Finally, on the evening of February 4 (Beijing Time), Coherent released its FY2026 Q2 earnings report, offering clear support for the assessments made by some A-share companies regarding optical modules. Coherent stated that while CPO was initially intended for Scale-Out scenarios, the company believes its future significant growth will actually be driven by Scale-Up applications. Based on the company's communication and collaboration with CSPs, Coherent believes the market size for CPO in Scale-Up scenarios will far exceed that of Scale-Out, potentially by several orders of magnitude. Coherent explicitly stated that pluggable optical modules will remain the dominant form factor in Scale-Out and Scale-Across scenarios for at least the next decade and are expected to see very strong demand and growth in the coming years.
Has the debate reached a conclusion? The essence of the dispute stems from divergence. Setting aside the differences and synthesizing the statements from major optical communication companies reveals a clear commonality: the large-scale rollout of CPO solutions in Scale-Up scenarios is imminent. This potential incremental demand could be substantial, and the trend of "optical replacing copper" appears to be a definitive shift. However, significant disagreement remains regarding whether CPO solutions will be adopted on a large scale in Scale-Out scenarios. At the very least, neither Lumentum nor Coherent has explicitly stated that CPO will definitely not be used in Scale-Out scenarios in the future. There is currently insufficient evidence to completely disprove AYZ's conclusion regarding CPO adoption in the Scale-Out scenario of the Vera Rubin Ultra rack.
As for whether the established pluggable optical module companies can capture a significant portion of the potential incremental demand from CPO adoption? This is akin to asking if someone proficient in "classical mechanics" can easily master "quantum mechanics." The answer might be singular: without an omniscient perspective, quantum mechanics remains, well, quantum mechanics!
Finally, Valuation Hub, in a report dated December 31, 2025, titled "From 800G, 1.6T Optical Modules to CPO: MPO Fiber Patch Cords Evolving from Quantity to Quality," clearly expressed three viewpoints that remain relevant to this discussion: First, while professional investors will closely track the timing of CPO volume ramp-up and marginal changes in the North American AI supply chain through various means, astute investors might ponder whether there exists a "picks and shovels" industry relatively immune to specific technological disruptions – one that can benefit from the 1.6T optical module ramp and potentially benefit even more from the subsequent CPO adoption wave. Second, recall the debates following the Blackwell rack launch regarding "optical retreating vs. advancing" and the stock performances of optical module versus copper cable companies. In other words, simplistic expectations based solely on "earning growth or valuation expansion" may overlook the complexities of capital markets. Third, regardless of the specific scenario – Scale-Up, Scale-Out, or Scale-Across – the technological evolution of AI data centers is unequivocally moving towards a future dominated by "light." The market's performance in the coming period is expected to be even more dynamic.
Comments