Author: Feng Ning
Recently, the semiconductor industry has been stirred up by a single term—CPO (Co-Packaged Optics). NVIDIA and SK Hynix, one holding the lifeline of computing power and the other gripping the throat of memory storage, have rarely found themselves unable to sit still at the same time.
01. SK Hynix and NVIDIA Double Down on CPO
Recently, SK Hynix and its global research team published a paper in the journal Nature Electronics. The paper explores the development of Co-Packaged Optics (CPO) for high-performance computing and AI (Artificial Intelligence), identifies key technical challenges, and outlines the development trajectory of next-generation optical interconnect technologies.
Seunghoon Hong, Head of the AI Infrastructure Team at SK Hynix, and Professor Kyusang Lee from the Department of Electrical and Computer Engineering at the University of Virginia (UVA), served as corresponding authors, leading a collaborative research effort with researchers from the University of Illinois Urbana-Champaign (UIUC), Nanyang Technological University (NTU), Massachusetts Institute of Technology (MIT), and Yonsei University.
On a broader level, the paper proposes a comprehensive technology roadmap, elaborating on how memory, Advanced Packaging, and Optical Compute Interconnect (OCI) should co-evolve to support next-generation AI systems. Crucially, it also demonstrates how SK Hynix is moving beyond HBM innovation to help define the architecture of next-generation AI infrastructure at the system level.
Meanwhile, the researchers have set clear technical targets for next-generation AI infrastructure, including a bandwidth of over 100 Tb/s per node, energy consumption below 1 pJ/bit, and chip-to-chip latency of less than 10 nanoseconds. The paper also presents a comprehensive technology roadmap, outlining the evolution from configurations based on 2D and 2.5D interposers to 3D heterogeneous stacking, as well as the key technical challenges that must be addressed for commercial deployment.
Coincidentally, on August 14, NVIDIA officially announced that its Spectrum‑X Ethernet Photonics Co-Packaged Optics switch has entered full-scale mass production. Aimed at AI factories, it builds next-generation large-scale scalable network infrastructure. This is the world's first mass-production-grade 200G/lane Co-Packaged Optics switch system. According to official data from NVIDIA, compared to traditional pluggable optical module solutions, it achieves a 5x improvement in network power efficiency, a 5x increase in uninterrupted AI application runtime, and extends the Mean Time Between Failures (MTBF) by 10 times, significantly addressing the power consumption and reliability shortcomings of hyperscale AI cluster networks.
CPO has suddenly become the common focus of both SK Hynix and NVIDIA, which is no coincidence. There is only one driving force behind this: data centers can no longer keep up.
02. Data Centers Need CPO
The earliest generation of data centers extensively used Direct Attach Copper (DAC) cables for internal interconnects. Switch chips relied on OSFP and QSFPDD connectors to interface with DAC cables, completing short-distance transmission using purely electrical signals. The copper cable solution had prominent advantages: low cost, simple deployment, no need for optoelectronic conversion devices, and controllable power consumption in short-distance scenarios. However, its shortcomings were equally fatal. Under high-speed signals, copper wires suffer from severe attenuation, with transmission distances mostly limited to a few meters. As bandwidth increases, power consumption rises rapidly, making it completely incompatible with high-density, ultra-high-bandwidth AI computing clusters. Today, it is gradually being replaced by optical interconnects.
Subsequently, the industry fully entered the era of pluggable optical modules, which is the mainstream state of the industry today.
Switch chip boards connect to external independent pluggable optical transceiver modules via OSFP and QSFPDD connectors, and then complete long-distance signal transmission through optical fibers. The biggest advantage of pluggable optical modules is flexibility: they support hot-swapping, allowing business recovery by directly replacing the module in case of failure. Upgrades and iterations are also very convenient, and the standardized form factor is compatible with the vast majority of cloud providers' data centers. However, the shortcomings of this architecture are infinitely magnified in ultra-high-bandwidth AI clusters: there is still a centimeter-level high-speed electrical trace between the switch chip and the panel optical module. Signal loss and crosstalk occur during high-speed signal transmission. To compensate for signal loss, the Retimer and equalizer inside the optical module continuously consume power, keeping the overall network power consumption persistently high.
For ordinary internet data centers, 800G and 1.6T pluggable optical modules are still sufficient. However, when facing AI training clusters with hundreds of thousands of GPUs, the number of switch ports reaches tens of thousands. After stacking a large number of ports, the additional power consumption and signal loss become a heavy burden on the entire computing cluster, and the traditional pluggable architecture begins to hit the physical ceiling.
CPO was born precisely to resolve this contradiction. The CPO architecture moves the electro-optical conversion process inside the chip package, compressing the high-speed electrical signal transmission distance to a few millimeters. This avoids the signal loss caused by long-distance high-speed copper traces and eliminates part of the DSP signal processing overhead, achieving comprehensive optimization of power consumption, latency, and reliability.
03. Before 200G/lane, What Speed Did CPO Use?
However, with the release of NVIDIA's 200G/lane CPO Ethernet switch, the immediate reaction of many readers is confusion: hasn't the industry been extensively discussing 800G and 1.6T optical modules? Is NVIDIA falling behind by only releasing 200G now? What is the difference between CPO and traditional pluggable optical modules? What does this mean for the booming domestic optical module companies?
To answer these questions, it is first necessary to clarify a set of easily confused concepts: 200G/lane describes the speed of a single optical lane, while 800G and 1.6T are the total speeds of a single port; the two do not belong to the same dimension.
The formula for the total port speed of an optical module is: Total Speed = Single Lane Speed × Number of Lanes, where the number of lanes is generally a power of 2, such as 4, 8, or 16. Here are a few common industry examples: 8×50G=400G, 8×100G=800G. Therefore, 200G/lane is not a backward new technology; rather, it is precisely the underlying physical foundation for next-generation high-speed interconnects.
Prior to the mass production of 200G/lane this time, technological exploration in the CPO industry had long been stalled at 50G/lane and 100G/lane. So why must the industry push the single-lane speed to 200G/lane? The root cause lies in the underlying network pressure brought by the explosion in AI computing power.
As GPU architectures iterate from Hopper to Blackwell and evolve toward the future Rubin architecture, the external communication bandwidth of a single GPU has experienced explosive growth. Today, large-scale AI training clusters often deploy hundreds of thousands or even millions of GPUs, and the total switching capacity that switches need to handle has reached unprecedented levels. If it continues to stay at 100G/lane, achieving a total switch capacity at the 102.4T level would require doubling the number of physical lanes. The direct consequence is the rapid expansion of the switch chip area, the routing density of the PCB (Printed Circuit Board) hitting physical limits, and the overall power consumption getting completely out of control. Cabinet power supply and cooling cannot bear the load, and the old path of simply stacking the number of lanes is no longer viable.
The essence of CPO is to break the traditional architecture of "separation between switch chips and optical modules" that has persisted for many years. It co-packages the optical engine, which includes lasers, modulators, and detectors, with the ASIC switch chip on the same substrate. This compresses the electrical signal transmission path from the centimeter level in traditional solutions to the millimeter level, fundamentally resolving the power consumption, signal attenuation, and bandwidth bottlenecks of traditional interconnect architectures.
Of course, NVIDIA is not the only player in the CPO track. Broadcom released its third-generation CPO switch chip, Tomahawk 6 Davisson (TH6 Davisson), in October 2025, achieving a switching capacity of 102.4 Tbps, directly doubling the bandwidth of the previous generation. Even earlier, Broadcom had launched the Bailly CPO switch, integrating eight 6.4 Tbps silicon photonic optical engines. Compared to traditional pluggable solutions, it reduces interconnect operating power consumption by 70% and improves silicon area usage efficiency by 8 times.
04. The Embarrassment of Optical Module Companies
However, when dissecting the complete supply chain list for NVIDIA's SpectrumX, it can be found that the core collaboration and division of labor for this mass-produced CPO switch are very clear: TSMC (Taiwan Semiconductor Manufacturing Company) is responsible for the process manufacturing of silicon photonic chips, SPIL (Siliconware Precision Industries) undertakes wafer-level and chip-level packaging and testing, Lumentum provides laser chips, TFC Communication is responsible for the assembly of laser module sub-assemblies, and Foxconn (Hon Hai) completes the assembly and delivery of the complete system.
One point worth pondering: this list of core suppliers does not include traditional pluggable optical module assembly giants such as Innolight and Eoptolink.
This directly exposes the industry reality: under the CPO architecture, there is no need for traditional form-factor, panel-pluggable optical modules. The optical engine is no longer an independent external component but is directly integrated into the switch package. This leads to the most core proposition in the entire industry: does the arrival of CPO mean the end of the era of pluggable optical modules? What kind of industry changes will domestic leading enterprises, whose core advantages lie in the assembly and integration of pluggable optical modules, face?
First, the core conclusion must be clarified: the mass production of CPO does not mean that pluggable optical modules will quickly exit the market. The two will coexist for a long time, and the industry will enter a dual-track driven stage.
Network planning information from multiple A-share listed optical communication companies and cloud service providers shows that from 2026 to 2027, the network construction plans of global cloud providers will still rely on pluggable optical modules as the main force, and no leading cloud provider plans to deploy CPO switches on a large scale in the next two to three years. TrendForce data shows that at this stage, the penetration rate of CPO in the AI data center optical interconnect market is only about 0.5%, still in the early stage of commercialization; IDC predicts that the window for large-scale commercialization of CPO will fall after 2027-2028.
At the same time, the inherent shortcomings of CPO at the operation and maintenance (O&M) level must be viewed objectively, which is also the key reason why cloud providers dare not rashly replace them entirely. When a traditional pluggable optical module fails, O&M personnel can directly replace a single module, completing the fault repair in a few minutes with almost no impact on business. However, in the CPO architecture, the optical engine and the switch chip are deeply bound. Once the photonic chip or optical engine fails, it may cause 832 ports to fail simultaneously. Maintenance often involves returning the entire machine or the entire switching unit for repair, significantly increasing O&M complexity and downtime costs.
Only when the yield rate of future CPO products continues to climb and the unit cost drops to a critical point will the value distribution of the entire industry chain undergo a substantial shift: the focus of value in the industry chain will shift from the traditional optical module assembly segment to the upstream segments of optical chips, silicon photonic chips, and advanced optoelectronic packaging.
Returning to the realistic situation of domestic manufacturers, under the CPO wave, the domestic optical communication industry is not simply facing a "crisis," but rather a restructuring of the industry chain, where opportunities and challenges coexist.
Before the leading domestic optical module companies lie three realistic paths. First, hold the basic market, continuously iterate 800G/1.6T/3.2T pluggable optical modules, and capture the mainstream demand for global AI computing power construction in the next 2-3 years. Second, extend upstream, deploy optical engines, silicon photonics, high-speed optical chips, and co-packaged optoelectronics, stepping out of pure module assembly and cutting into the high-value segments of CPO. Third, deploy transitional technology routes such as Near-Package Optics (NPO). The technical difficulty of NPO is relatively lower than that of CPO, making it more suitable for the realistic conditions of domestic cloud providers. It is an important track for domestic manufacturers to achieve technological implementation.