Co-Packaged Optics (CPO) is an optoelectronic transmission technology that integrates the optical engine directly onto the same package substrate as the chip, enabling both inter-rack and intra-rack connectivity to address bandwidth bottlenecks, signal attenuation, and thermal challenges in traditional data centers. As AI-driven demand for computational power surges, conventional network infrastructure can no longer meet the bandwidth requirements of the Agentic AI era, making CPO a critical breakthrough direction. Major players such as NVIDIA and Broadcom are actively advancing CPO switch solutions, which currently face key challenges in advanced packaging processes, thermal management, maintenance, and standardization. Compared to other technologies such as NPO, OIO, CPC, LPO, and OCS, CPO is the essential next-generation solution to be mastered, with industry value increasingly concentrating among switch chip manufacturers and advanced packaging providers.
Author and source: Dolphin Research
Since ChatGPT's emergence at the end of 2022, AI has driven one after another semiconductor super-industry opportunity and one after another trillion-dollar market cap company, from computing power (GPUs) and storage capacity to orchestration and control (CPUs).
If there is still one segment in AI infrastructure waiting for a trillion-dollar market cap “next big thing,” Dolphin believes the most promising is the super-connectivity of the AI era. If compute power solves AI’s “intelligence” problem and storage power solves its “memory” problem, then transport power is what will enable long- and short-term memory to enter and exit the cognitive center at rocket-like speeds.
Or, in the words of AI pope Jensen Huang, as compute and memory bottlenecks gradually ease, energy remains a persistent challenge at the high school level—the next critical bottleneck is high-speed networking in the AI era, because traditional cloud-era network infrastructure is entirely inadequate to meet the bandwidth demands of trillions of model parameters, Mixture of Experts (MoE), and local activation in the Agentic AI era.
In this article, we continue exploring network transmission in the AI era, focusing on the optical-electrical transmission technology—CPO—that is gradually being adopted under AI network speed requirements. Dolphin’s research on CPO is divided into:
What is CPO, and can it truly replace traditional copper connections?
Two, can it completely replace the current mainstream pluggable optical modules?
III. Under this trend, how will the competitive landscape among upstream and downstream companies in the industry change?
In this article, we first provide an overview of the fundamental issues in the industry chain.
Here is the detailed analysis
01 What is CPO?
In traditional data center architectures, a key component is the "optical module," which converts optical signals transmitted via fiber into electrical signals for the data center, or converts electrical signals generated within the data center into optical signals for transmission over fiber, serving as a "bridge" and "translator" in data transmission.
In terms of function, the CPO (Co-Packaged Optics) architecture incorporates the capabilities of traditional optical modules but has two notable differences:
1. Different structure
Traditional optical modules are pluggable and resemble the RJ45 connectors on home network cables, but CPO is entirely different—it directly integrates the optical engine responsible for electro-optical conversion and the chip (primarily the switch’s ASIC chip) onto the same package substrate or interposer.
2. Different application scenarios
Optical modules are typically used between racks (scale-out); whereas CPO can be used both between racks (scale-out) and within racks (scale-up). When used between racks, CPO replaces traditional optical modules; when used within racks, it replaces the currently dominant copper connections.
Figure: Schematic comparison of traditional pluggable mode and CPO solution
Source: GTC 2025, Dolphin Research
We can see that recently, both NVIDIA and Broadcom have been actively promoting their CPO switch solutions.
So why is CPO technology receiving such significant attention? Because the demand for computing power in data centers continues to rise, bandwidth requirements for data transmission are growing explosively, and data centers are evolving toward hyperscale computing clusters. During this transition, traditional data transmission technologies are creating numerous bottlenecks:
1. Bandwidth bottleneck
In cabinet-to-cabinet scenarios, the limited panel space of traditional switches, combined with the difficulty of reducing the size of conventional pluggable optical modules, restricts the number of ports available per switch, making it unable to meet increasingly high bandwidth demands.
Currently, the highest-capacity pluggable module supports up to 1.6 Tbps per module, with a single switch panel capable of supporting up to 51.2 Tbps total bandwidth. Future modules of up to 3.2 Tbps may be introduced, enabling a maximum switch capacity of 102.4 Tbps—nearly reaching the practical limit of pluggable optical modules.
2. Signal Integrity Bottleneck
In a rack environment, as transmission rates increase, using traditional copper cables leads to significant signal attenuation and distortion over long distances, further limiting the achievable transmission distance.
Currently, copper cables can support up to 1.8 TB/s bandwidth (such as NVIDIA’s NVLink copper cables), but are strictly limited to distances under 2 meters, while the bandwidth demand per GPU is advancing toward 3.6 TB/s.
3. Thermal and Power Consumption Bottlenecks
As transmission rates increase, power consumption in traditional communication links rises significantly, and heat dissipation becomes increasingly challenging. We know that U.S. data center construction currently faces major energy constraints, so power consumption issues will create substantial cost pressures.
CPO can theoretically address the above issues effectively; according to NVIDIA, power efficiency can improve by 3.5 times with CPO.
02 Specifically, what are the data transmission scenarios in data centers?
Here, we break down the data transmission technology pathways for the data center across different scenarios and stages:
Figure: Examples of Scale-out and Scale-up
Source: NADDOD, Dolphin Research
1. Scale-up, primarily involves intra-rack connectivity
Primarily involves hardware interconnects within the rack, especially within servers, including but not limited to connections between CPUs, GPUs, network interface cards, DDR memory, and hard drives.
Currently, this section primarily uses copper as the main connection medium, including PCIe slots for connecting the CPU, GPU, and network cards, memory slots (PCB copper traces), SATA cables, and various other copper cables. CPO has the potential to disrupt the current mainstream solution.
2. Scale-out, primarily involving interconnectivity between racks
Primarily involves interconnections between cabinets or servers and switches.
This connection requires optical media as the transmission medium, with fiber optics and pluggable optical modules being the primary solutions today. Similarly, CPO is a key development trend and is advancing faster than in-rack scenarios.
3. Furthermore, interconnections between data centers and between data centers and external networks are not the focus of this article.
From the perspective of major players' strategies, CPO is currently primarily targeted at inter-rack scenarios, but may expand to intra-rack scenarios in the future.
03 CPO is currently in the early stages of promotion; what are the main bottlenecks it faces?
1. Maturity of advanced packaging technology
From a foundational technology perspective, CPO is fundamentally different from traditional solutions such as pluggable optical modules. While traditional optoelectronic components are technologically similar to broader optoelectronic devices and modules, CPO requires packaging the optical engine onto a substrate or interposer, primarily relying on advanced packaging technologies such as CoWoS.
At the same time, CPO differs from what we typically understand as advanced packaging, as it requires not only the integration of electronic integrated circuits but also photonic integrated circuits, necessitating heterogeneous integration through hybrid bonding technologies such as TSMC’s COUPE.
The issue is that, on one hand, the aforementioned advanced packaging technologies are extremely complex to manufacture, and both NVIDIA and Broadcom rely on TSMC’s production capacity, which is limited. In addition, supply constraints may also arise for required materials such as optical couplers, equipment for hybrid bonding and testing, and ABF substrates.
Moreover, the production yield of these advanced packaging technologies, particularly heterogeneous integration, still has significant room for improvement, resulting in costs far higher than those of pluggable solutions. TSMC is currently working to improve the yield of advanced packaging, but this will still require some time.
2. Maintenance and Repair Issues
For traditional pluggable solutions, maintenance and repair are convenient because they are “pluggable.” However, CPO is entirely different—its photonic and electronic modules are directly packaged together with the substrate, interposer, and even the chip, making maintenance and repair significantly more difficult than with traditional solutions.
However, these issues can be addressed—for example, by designing in a higher tolerance for errors or implementing operational redundancies.
3. Thermal Management Issues
The high-density packaging of the optical engine and chip causes significant localized heating during operation, potentially exceeding the laser's tolerance limits, making thermal management a major challenge. To address these issues, more efficient cooling solutions must be introduced, but this also involves increased costs.
4. Standardization Issues
Currently, NVIDIA, Broadcom, and others are actively launching their own complete, independent CPO switch solutions to gain a first-mover advantage in the market. However, industry standards (such as interface and packaging standards) have not yet been established, making it difficult for upstream and downstream players to develop, produce, and configure products based on a unified standard—this remains a key challenge for commercialization.
In summary, solutions to the above issues exist, but they rely on technological maturity and the establishment of standards, both of which take time.
On the other hand, fundamentally, CPO technology must achieve a cost advantage.
This raises the question: regardless of the approach, cost is always a core consideration, but alongside CPO, other more advanced or more conservative pathways are also being pursued—how do they relate to each other? Let’s first clarify the differences between these various technical approaches.
04 Comparison of Technical Approaches
1. CPO
The CPO we are discussing, co-packaged optics, refers to integrating the optical engine and the chip on the same substrate, where the chip can be a switching ASIC or other computing chips such as a GPU, but typically refers to a switching ASIC.
2. NPO
NPO stands for Near-Packaged Optics, which is less advanced than CPO and does not yet achieve integration at the level of the same substrate or interposer, but rather integrates components on the same PCB motherboard.
In China, companies such as Alibaba and Huawei are promoting NPO solutions, which can largely be seen as a compromise due to a lack of advanced packaging capacity. However, these solutions may become the dominant approach in the Chinese market for a period of time, thereby somewhat impacting NVIDIA’s market penetration in China.
Figure: Different integration methods shown: (from top to bottom: plug-in, NPO, CPO (integrated on the package substrate), CPO (integrated on the interposer), and OIO, which will be discussed below)
Source: ASE, Dolphin Research
3. OIO
OIO (Optical I/O) can be seen as an advancement over CPO, where there is no involvement of switching chips; instead, it primarily relates to compute chips, referring to the integration of optical engines with compute chips within the same package, or even directly at the chip level, targeting entirely in-rack scenarios.
Figure: Demonstration of different integration methods: pluggable, CPO, OIO
Sources: TSMC, Openlight, Dolphin Research
With that in mind, let’s further clarify the data center architecture:
A data center can be viewed as several interconnected components:
The server is dedicated to computational tasks and is equipped with computing chips such as GPUs and CPUs, along with memory and storage drives.
The switch handles network communication between servers and between servers and external networks, using ASIC chips to facilitate data exchange;
In addition, there is the storage system, which in current mainstream data center architectures is primarily distributed across server nodes and housed internally within the servers, integrated with them.
Based on the above architecture, we can envision the applications of CPO. Building on this, let’s discuss why CPO initially began with switching chips.
Here, we can draw an analogy for the role of switches—switches can be thought of as overpasses within a data center. It’s easy to imagine that the data transmission bandwidth demands, port density, and associated power consumption bottlenecks they face are the greatest, making the need for CPO all the more urgent.
4. CPC
CPC, or Co-Packaged Copper, refers to the direct integration of high-speed copper connectors onto the package substrate.
The cost advantage of this technology path is very clear, but it still cannot resolve the bandwidth bottleneck and attenuation issues associated with copper media, resulting in limited application scenarios—primarily suitable for connections between GPU/CPU nodes, switches, and storage chips within a rack. Currently, NVIDIA’s intra-rack solutions still use copper connections, but they may transition to optical interconnects in the future.
5. LPO
LPO, or Linear-Drive Pluggable Optics, is a streamlined version of pluggable optics that eliminates the internal DSP/CDR chip while retaining and enhancing the analog chips—Driver and TIA (whose functions we will explain later)—to enable direct signal driving.
In simple terms, this means removing the power-hungry DSP chip from the optical module and abandoning signal error correction, while enhancing the analog chip to directly amplify the signal—regardless of its accuracy—so that the electrical signal from the switch ASIC can directly drive the laser.
Figure: Comparison diagram of traditional model versus LPO structure
Source: Bryon Moyer, Semiconductor Engineering, Dolphin Research
However, the same issue exists here: since the PCB traces are not omitted (causing signal attenuation) and the requirements for signal quality are even higher, long-distance transmission remains limited. Moreover, as speeds advance to higher levels (above 1.6T), signal integrity issues become particularly pronounced. In other words, while the structure is simplified, performance is also compromised.
In summary, although compromise solutions such as NPO, CPC, and LPO exist, as data centers move toward higher speeds and larger clusters, these approaches will ultimately hit bottlenecks—CPO is the next-generation solution that must be breakthrough.
6. What is an Optical Circuit Switch (OCS), and does it pose a threat to CPO?
Here, we inevitably come to the Optical Circuit Switch (OCS). The core feature of OCS is that it performs no optoelectronic conversion throughout the process; instead, it establishes physical optical paths directly within the optical domain using an optical switch matrix.
Figure: OCS schematic
Source: Orbray, Dolphin Research
It can be intuitively imagined as consisting of rows of mirrors (a micromirror array) that can adjust the angle of each mirror according to instructions to reflect light in different directions.
On the surface, OCS appears to directly forward optical signals, replacing the traditional switch’s optical-to-electrical and electrical-to-optical conversion processes, suggesting that this technology path might eliminate the need for CPO (at least in the switch layer). However, this is not actually the case.
Here, we outline how switch architectures are constructed in data centers:
(1) Within the mainboard: First, we know that the core computation within a data center is performed by GPUs. After GPU computation is complete, the data must be transferred to the CPU, which processes it before sending it to the network card (containing an ASIC), or the data can be transmitted directly from the GPU to the network card.
The above processes can be implemented on a single motherboard, or at least within a single server.
(2) Inside the rack: Afterward, data must be transmitted from the servers to the rack’s switch. Multiple servers within a single rack can interconnect at high speed, but a switch must be installed at the top of the rack to communicate with external networks, enabling data exchange between the rack and external systems. This switch is called a Top-of-Rack (ToR) switch.
And the above steps are implemented within the same rack.
(3) Between Racks: A data center consists of a cluster of multiple racks; how is communication between racks orchestrated? This is where spine switches come into play. Spine switches manage high-speed connections between all leaf switches and to external networks outside the data center, serving as the central hub of the entire switch network within the data center.
Diagram: Spine and Leaf switches in a data center
Source: Bryon Moyer, Semiconductor Engineering, Dolphin Research
OCS is primarily used to replace spine switches.
First, spine switches are expensive and power-intensive, making the demand for alternatives the most urgent.
Second, the role of OCS is limited—it can only forward signals (reflect light), much like a mirror. In contrast, traditional switches have more comprehensive functionality: they must unpack data packets, examine IP addresses, and then determine where to route them. For example, since OCS can only execute commands without any decision-making capability, it is feasible to use it solely as a Spine switch. However, if you aim to replace Leaf switches as well, additional components—such as SmartNICs—would be required to handle packet processing. This would complicate the architecture and may not necessarily be the optimal solution.
Looking at it this way, the architecture becomes clear:
Although currently both NVIDIA’s Quantum X800-Q3450 and Broadcom’s Tomahawk 6-Davisson, which follow the CPO architecture, are Spine switches, and Google’s OCS switches are also designed to replace traditional Spine switches, there is indeed direct competition between the two.
Ultimately, although OCS has the potential to replace spine switches, further downstream, for higher-volume applications—such as the electrical-to-optical conversion between optical engines and ASICs on leaf switches, connections between server motherboards (via NIC ASICs or NVSwitch), and connections between computing chips on the motherboard, as well as between computing chips and NIC ASICs—CPO will still be required. Therefore, in the future, the two technologies will be complementary rather than competitive.
What are the industrial chain segments involved?
(1) First, let's analyze the principle and architecture of CPO.
CPO can be regarded as an upgraded version of the optical engine, which is responsible for optoelectronic conversion and primarily consists of the following components:
1. Photonic Circuit Section
(1) Modulator: Converts electrical signals (0/1 digital) into optical signals by controlling the intensity and timing of light.
(2) Detector: A PD (Photodiode) that converts optical signals into electrical signals.
(3) Waveguide: Can be understood as microscopic optical fibers printed onto the chip.
2. Electronic Circuit Section
(1) Driver: Amplifies the weak electrical signals received from the switch or server into precise electrical signals that control the laser's emission, so the next stage after the driver is the modulator.
(2) TIA (Transimpedance Amplifier): Amplifies the extremely weak electrical signal generated by the photodiode and converts it into a voltage signal suitable for processing by subsequent circuits, making the TIA the next stage after the photodiode.
3. The light source, i.e., the laser
The modulator itself cannot emit light, but it can control light; therefore, a light-emitting component—the laser—is required to work with it.
Diagram: Light Engine Structure
Source: Zong ZeGuo et al., "Research on 400G FR4 Silicon Photonic Transceiver Module," Dolphin Research
In addition, there are two other components:
4. DSP and CDR are both used to repair electrical signals. DSP compensates for physical impairments in the electrical signal, while CDR extracts an accurate clock from the degraded signal and re-timing the data, with DSP chips typically integrating CDR functionality.
Like LPO, CPO removes the power-hungry, costly, latency-inducing DSP from the optical engine. However, in CPO, some DSP functions are integrated into the switching ASIC, whereas LPO uses an analog chip to amplify the signal. Additionally, CPO integrates the CDR into the high-speed SerDes.
What is high-speed SerDes? High-speed SerDes consists of a serializer (Ser) and a deserializer (Des), located inside ASIC chips, which respectively pack parallel data from within the chip into high-speed serial data streams, or unpack and reconstruct high-speed serial data streams back into multiple low-speed parallel data streams.
(2) Now let’s look at all the stages involved in the entire CPO industry chain:
1. First, the overall CPO
The optical engine in CPO includes both the photonic circuit and electronic circuit components mentioned above, and together with the ASIC chip, forms the main body of the CPO switch. First, let’s address a core question: Who will build this CPO?
Traditional optical modules, as standalone units composed of optical components and discrete devices, can be fully provided by specialized manufacturers, such as InnoLight, Eoptolink, and Coherent, which are well-known to us. What about CPO? Clearly, these companies can no longer dominate it.
We tend to believe that the industrial value trend under CPO will be as follows:
(1) Exchange manufacturers and platform providers that control core technologies: Companies that control data center system platform and switching chip providers such as NVIDIA, Google, Broadcom, and Marvell to define architectures and standards, and sell complete solutions;
(2) Contract manufacturers: Companies such as TSMC, UMC, and Amkor for wafer fabrication, photonic integration, and advanced packaging services;
(3) Upstream suppliers: Companies such as Coherent and Lumentum continue to manufacture and supply optoelectronic components.
(4) Traditional optical module manufacturers: During the transition period, companies such as InnoLight and NeoPhotonics continue to provide optical engine modules, offering intermediate solutions like NPO and LPO, as well as compromise CPO designs based on maintainability considerations.
2. In addition to the CPO's core optical engine, several other components require attention.
(1) Laser
CPO can only integrate optoelectronic conversion components; directly integrating lasers remains challenging, so external lasers are still required. At the same time, CPO significantly increases the power requirements for lasers (by at least 3 to 4 times), which also substantially raises the demands on performance and reliability, thereby significantly increasing their value.
However, there is a choice of technical approach:
1) EML laser: A traditional approach that integrates the laser and modulator; its advantage lies in suitability for high-bandwidth, long-distance communications above 200G. This pathway is dominated by major players such as Lumentum, II-VI (Coherent), and Sumitomo.
2) CW Lasers: An emerging approach that fully isolates the laser, offering advantages in cost and power consumption, and better alignment with the future CPO architecture. CW laser suppliers enjoy greater flexibility, with Chinese manufacturers such as Yuanjie Technology, Shijia Photonics, and Changguang Huaxin having achieved mass production of 70mW/100mW products and secured significant orders.
Figure: Schematic illustration of the difference between EML and CW lasers
Source: Sumitomo Electric, Dolphin Research
Next are the four major optical fiber components, which are rarely used in traditional pluggable optical module architectures:
(2) Fiber Array Unit (FAU): Used to precisely mount fibers to achieve high-precision alignment between fibers and waveguides.
Figure: Fiber Array Unit
Source: Corning, Dolphin Research
(3) Polarization-Maintaining Fiber (PMF): A special type of fiber designed to maintain the polarization state of light.
(4) Fiber Distribution Box (Fiber Shuffle): Used to organize fibers, allowing the order of fibers within complex, high-density equipment to be rearranged.
Figure: Fiber Shuffle Illustration
Source: Hyoptic, Dolphin Research
(5) Fiber optic connector (MPO, Multi-Fiber Push On): Used for connecting multiple fiber cores.
Figure: MPO Port Schematic
Sources: Senko, US Conec, Dolphin Research
Why are the above components rarely used in traditional optical modules?
(1) In traditional models, optical fibers are directly inserted into standardized interfaces, but under CPO, the fibers must be precisely coupled to the waveguides on the surface of the optical chip, requiring the use of FAUs;
(2) The traditional approach involves direct modulation, is insensitive to the polarization state of light, and previously suffered from extremely high costs of polarization-maintaining fiber (PMF), making it unsuitable for industrial-scale applications. However, CPO relies on an external laser for light sourcing, and laser polarization states can cause significant energy loss; therefore, PMF is essential.
(3) Traditional setups typically involve only two fibers—one transmit and one receive—so there are no complex fiber connections required to the backplane, making manual handling sufficient without the need for Fiber Shuffle; however, Fiber Shuffle is mandatory under CPO.
(4) Similarly, traditional modules do not require many interfaces, but under CPO, when speeds reach 400G or higher, eight or even 16 fibers must transmit in parallel. With limited panel space, an MPO multi-fiber connector is required.
We will analyze the market opportunities and investment prospects related to the industries involved in CPO in the next part.