[Relevant Companies Involved in This Article] Nvidia, Broadcom, Marvell, Google, AMD, TSMC, Foxconn, Quanta, Flex, Zhongji InnoLight, Eoptolink, WUS Printed Circuit.
Nvidia CPO Progress
From the perspective of physical layer and system interconnect evolution, traditional pluggable optical modules face severe signal attenuation and power consumption bottlenecks as single-channel rates advance toward 200G/400G and single-switch bandwidth exceeds 100Tbit/s. As the hardware leader in computing units, Nvidia has launched two CPO products, Quantum-X and Spectrum-X, which are technical solutions specifically designed for building scale-out networks in next-generation ultra-large-scale AI clusters.
Looking first at the hardware physical architecture analysis, these two CPO switches are highly homologous at the optoelectronic physical layer. The core SerDes single-channel rates are both based on the 200G specification, and both adopt a 200G single-wavelength micro-ring silicon photonics chip architecture to achieve optical channel modulation. However, they differ in network protocol stacks and bandwidth configurations. The core switching rate of Quantum-X reaches 100T/s, supporting 144 800G ports or 72 1.6T ports (which corresponds to the 72-way 1.6T port interconnect of NV72). Spectrum-X leans more towards the Ethernet ecosystem, offering two tiers: 100T (128×800G) and 400T (4×128×800G), capable of supporting scale-out interconnects for up to 576 GPUs.
For the core silicon photonics of CPO, the single-channel 200G/s has already reached the standard for commercial mass production in the industry. However, when sprinting toward 400G per single wavelength, it is limited by the internal modulators of silicon photonics chips and remains to be developed. Therefore, Nvidia's choice to expand system capacity horizontally rather than blindly pursuing single-channel rates has indeed yielded significant results. The production capacity for its NV72, NV576, and 10,000-card interconnect solutions has long been fully booked, and its system capacity continues to evolve from 100T to 200T, 400T, and even 800T. Similarly, regarding the reliability of mass production, Nvidia adopts a solution combining an External Laser Source (ELS) with liquid cooling. Because lasers are extremely sensitive to temperature, prone to aging, and have high failure rates, placing the ELS on the panel side in a pluggable form factor and equipping it with a liquid cooling module can control the temperature, thereby stabilizing the output wavelength of the laser. Consequently, the temperature drift issue of the micro-ring resonator is improved, and the long-term power consumption of the system is significantly reduced.
Additionally, the author has learned from multiple sources that regarding the "mass production progress" currently concerned by the market, information from the packaging and testing sides indicates that the "mass production" referred to by officials and the industry is closer to small-batch prototype delivery and early customer validation stages, rather than mature, large-scale, and continuous mass shipments. The initial customers are also mainly concentrated among service providers such as CoreWeave, Lambda, and Oracle, with an overall expectation of around 1,000 units. The targets are also primarily focused on the construction of 10,000-card clusters and preliminary feasibility testing. Furthermore, achieving large-scale mass production of CPO solutions still faces two major challenges that need to be overcome:
First is the issue of operation and maintenance (O&M) services. Traditional pluggable optical modules support hot-swapping of individual ports, whereas in the CPO architecture, the optical engine and ASIC are highly integrated. Damage to a single component could directly scrap a CPO package valued at 300,000 USD. Investors need to weigh and consider the O&M costs and break-even points for data centers.
Second is ecological competition. Large CSPs (such as Meta, Microsoft, and AWS) currently rely heavily on Broadcom (Broadcom TH5/TH6) solutions for switching networks, and 1.6T pluggable optical modules are about to enter a period of large-scale centralized procurement. Meanwhile, the simultaneous efforts of LPO and LRO are further reducing power consumption based on glass DSPs, largely retaining the modular O&M advantages. Furthermore, as a comprehensive solution, NPO (Near-Package Optics) is also highly sought after by a group of domestic manufacturers. For example, Alibaba, Tencent, and ByteDance are all actively deploying NPO solutions.
AI Computing Unit Delivery Progress
Similarly, the author has also learned from the front lines that the deployment of AI computing infrastructure is currently undergoing a leap from "chip-level" to "rack-level" complex system integration. For example, the Blackwell (GB200/GB300) and Rubin (Vera Rubin NVL72/CPX/LPU) series products have seen continuous downward revisions in market expectations. At the beginning of 2026, the market optimistically expected the demand for NVL72 in the second half of the year to reach 15,000 racks, but the latest industry chain data indicates that the expectation has been significantly revised down to around 7,000. The core reasons for the lowered expectations mainly stem from two factors:
1. Mass production progress and yield rate issues: According to the yield rates of ODM manufacturers such as Foxconn, Quanta, and Flex, their overall system yield rate is around 70%, which is significantly lower than the 95% production line yield rate of standard general-purpose servers, and also lower than the 85% yield rate of the previous generation A100/H100 traditional GPU servers. For complex rack products that include multiple sets of Compute Trays, NVLink Switch Trays, and complex liquid cooling pipelines, various issues such as signal interference and crosstalk caused by the doubling of component quantities, as well as micro-leakage in liquid cooling, urgently need to be resolved.
1. Soaring costs of memory and components: Not only the AI server industry, but the entire electronics industry is affected by rising component prices. The prices of HBM3/3e, HBM4, large-capacity DDR5, and NAND continue to surge. The price of a single 64G DDR5 module in the market has more than quintupled, and the price trend will continue to rise. According to industry calculations, the total price of memory in a single server has even surpassed the GPU chip itself, leading to high BOM costs for the entire machine.
Meanwhile, when tallying orders from large North American CSPs, the author also sees that for the 2027 outlook, giants such as AWS, Google, Microsoft, Meta, Oracle, and OpenAI are all targeting their orders at the Rubin NVL72 series. The overall volume of the Rubin platform is estimated to be about 70,000 racks, and the vast majority of these orders will be handed over to Taiwanese ODM manufacturers with system engineering validation capabilities. On the other hand, driven by policy orientation regarding AI sovereignty and regional policies, orders from some domestic computing power manufacturers are beginning to target standard server unit manufacturers such as Huawei, HPE, Dell, and Zhongke Shuguang. Especially in the GB series, the proportion at the beginning of 2026 rapidly climbed from 5% to 20%–25% in the GB300 stage, and will reach 30%–35% in the Rubin series.
At the same time, according to media reports, the computing units of current telecom operators and the newly built machine rooms for national-level projects are backed by massive industrial funds. Unlike commercial entities such as CSPs, they are not subject to strict capital expenditure and ROI constraints. (CSP capital return requirements, with AI computing assets depreciated over 5 years and amortized by about 20% annually, the realization cycle for commercial returns will reach about 8 years, and may further slow down amid competition.) In contrast, sovereign AI funds are not constrained by short-term financial returns, and their computing power usage is also relatively flexible. They can be later converted to scientific research or educational purposes, so they lean more towards the high-computing-power Rubin platform. Therefore, a large part of the current shipment demand for this series is driven by such sovereign AI funds.
Conclusion
The competition for AI computing power has entered the second half. CPO technology is indeed leading, but mass production has always been an engineering issue. Cost and capital return are the primary issues that CSP manufacturers must face first. Therefore, as pioneers, pluggable optical modules and NPO are destined to have significant room for development. Regarding delivery volume and cycles, the current yield rate is indeed a major bottleneck, and sovereign AI and government-guided funds will take the baton to continue driving the rapid development of the industry.
Reference: Exclusive Insight "Analysis of the Technical Architecture, Mass Production Progress, and Market Promotion Prospects of Nvidia CPO Switches Quantum-X and Spectrum-X"