NVIDIA Makes Major Bet on Optical Technology, CPO Gains Momentum in AI Infrastructure

Written by Xiao Bing, Chaoxiang Research

On June 1, 2026, at the Taipei Pop Music Center, Jensen Huang, in his signature leather jacket, unveiled the Vera Rubin architecture and the blueprint for the next-generation AI factory. Beneath this highly anticipated keynote, a dominant theme running through the first half of 2026 has become unmistakably clear:

NVIDIA is making a huge bet on light.

In March, NVIDIA invested $2 billion each in Lumentum and Coherent to secure capacity and technology pathways for next-generation silicon photonics lasers. In May, NVIDIA further invested $500 million alongside century-old fiber optic leader Corning to boost U.S.-based optical connectivity manufacturing capacity tenfold and increase fiber production by over 50%. On June 2, Jensen Huang directly declared at an event, “Marvell is poised to become the next trillion-dollar company.”

Stand in the light, believe in the light. This once-popular A-share meme has now been embodied in real money by Jensen Huang, becoming an industry consensus.

Imagine you’ve built ten thousand skyscrapers in a vast city, each housing tens of thousands of brilliant mathematicians (GPUs) solving massive numbers of problems every second. The question is: once they’ve solved them, how do the answers get transmitted? And how do the buildings coordinate with one another?

If you only give them country roads (traditional copper cables), no matter how brilliant they are, they can only wait idly—no matter how fast they calculate, the data gets stuck on the road, and the entire city grinds to a halt.

This is the real dilemma facing AI data centers today.

Since the emergence of ChatGPT, AI has fueled demand for GPU (computing power), HBM (memory capacity), and CPU (orchestration), spawning one after another company with a trillion-dollar market capitalization. However, in AI infrastructure, there is another critical component: data transmission.

The core medium for data transmission is the optical module.

As traditional optical modules begin to struggle to keep up with AI's demands, a next-generation technology called CPO (Co-Packaged Optics) is emerging strongly.

This article will explain, in the simplest terms possible, what optical modules are, why CPO is the future, and which companies in the upstream and downstream supply chain are worth paying attention to, breaking down this trillion-dollar industry.

1.1 Why is light needed?

Inside the data center, chips communicate using "electrical signals," similar to electrical impulses in the human nervous system. But electrical signals have a critical weakness: they don’t travel far, and they easily distort when moving at high speeds.

Copper cables transmit electrical signals like pushing water through a pipe—the longer the distance, the more the pressure drops; the narrower the pipe, the less the flow. Currently, the maximum transmission distance for copper cables is only about 2 meters, with a bandwidth ceiling of around 1.8 TB/s.

Light signals, on the other hand, are completely different. Light transmitted through optical fibers travels like a bullet in a vacuum tube, with almost no attenuation, extremely high speed, and immunity to electromagnetic interference. A single optical fiber, no thicker than a human hair, can theoretically transmit tens of Tbps of data simultaneously.

But the problem is: chips only "understand" electrical signals, and optical fibers only "carry" optical signals.

So, we need a "simultaneous interpretation" to convert electrical signals into optical signals for transmission, and optical signals back into electrical signals for reception.

This translator is the optical module.

1.2 What’s inside a light module?

If you take a light module apart, it’s essentially a precision translation box containing the following key components:

Transmitter (electrical to optical):

• Driver: Amplifies the weak electrical signals from the chip to a level sufficient to control the intensity of the laser’s light emission—like an amplifier in front of a microphone; without it, the signal is too weak for the laser to "hear."
• Modulator: Takes the amplified electrical signal and controls the brightness and timing of light to "write" the digital 0s and 1s into the light. It does not emit light itself, but only directs it.
• Laser: The true "light source," emitting a steady laser beam continuously. The modulator controls its light to "write."

Receiver (optical to electrical):

• Detector / Photodiode (PD): Receives the optical signal transmitted through the fiber and converts it back into an extremely weak current, much like the human retina converts light into neural signals.
• TIA (Transimpedance Amplifier): The current signal generated by the PD is too weak; the TIA amplifies it into a voltage signal that subsequent circuits can process, similar to turning a whisper into a normal speaking volume.

Signal repair:

• DSP (Digital Signal Processor): Electrical signals can become distorted after long-distance transmission; the DSP acts like Photoshop, restoring blurred signals to clarity. It consumes significant power and is one of the most expensive and power-intensive components in an optical module.
• CDR (Clock Data Recovery): Realigns the timing in degraded signals to ensure precise intervals between 0s and 1s. Typically integrated into DSPs.

Optical path:

• Waveguide: microscopic optical fibers "printed" inside the chip, through which light signals travel.
• Fiber optic interface: The physical interface that connects the optical module to external fiber optic cables.

In one sentence: An optical module = light source + modulator + detector + driver/amplifier circuit + signal conditioning chip.

1.3 The Evolution of Optical Module Speeds

The development of optical module speeds can be compared to the evolution of mobile communications:

Each doubling of speed signifies technological advancement and revaluation across the entire industrial chain. We are currently at a critical juncture transitioning from 800G to 1.6T, which is why the optical module sector has become the hottest trend on China’s A-share market over the past year, with the Wind Optical Module Index rising more than 500% since its 2025 low.

2.1 Bottlenecks of Traditional Optical Modules

Traditional pluggable optical modules are like USB devices—plug them in and they work; if they break, just replace them. This design is flexible and convenient, but in the AI era, it faces three major bottlenecks:

Bottleneck one: Bandwidth ceiling

Traditional switch panels have limited space, and the size of pluggable optical modules is difficult to reduce further. Currently, a single module supports up to 1.6 Tbps, with a single switch reaching a maximum of 51.2 Tbps. Future modules may offer up to 3.2 Tbps, pushing switch limits to 102.4 Tbps, but this is nearly the physical limit of pluggable solutions.

Bottleneck two: Power consumption explosion

Each GPU requires six hot-pluggable optical modules, each consuming approximately 30 watts. To build a supercluster of one million GPUs, the power consumption of the optical modules alone would reach 180 MW—equivalent to the electricity usage of a medium-sized city. This is completely unsustainable.

Bottleneck three: Signal attenuation

Pluggable modules are mounted at the edge of the switch panel, with long PCB traces separating them from the core ASIC chip. The higher the transmission rate, the more severe the signal attenuation over this "last mile," requiring additional signal conditioning chips (DSPs), which further increase power consumption and latency.

2.2 What is CPO?

The core idea of CPO (Co-Packaged Optics) is simple: place the interpreter right next to the brain.

Specifically, this means directly packaging the "optical engine" responsible for optoelectronic conversion on the same substrate or interposer as the switching chip (ASIC), rather than using it as a plug-in peripheral—it becomes a native, chip-level integration.

For example:

• Traditional optical modules are like making a phone call with Bluetooth headphones—the signal must be sent from the phone, encoded via Bluetooth, transmitted through the air, and then decoded by the headphones, with each step introducing loss and delay.
• CPO is like speaking directly into your ear, eliminating all intermediaries—faster and more energy-efficient.

According to NVIDIA's data, power efficiency can improve by 3.5 times after implementing CPO. IDTechEx predicts the CPO market will grow at a compound annual growth rate of 37% starting in 2026, exceeding $20 billion by 2036.

2.3 Key Timeline for CPO

2.4 Challenges Facing CPO

Although CPO represents the future direction, there are still several hurdles to overcome at this stage:

Advanced packaging capacity: CPO requires the heterogeneous integration of photonic and electronic circuits, which demands TSMC’s cutting-edge packaging technologies such as COUPE/SoIC. Current capacity is limited, yield still has room for improvement, and costs are significantly higher than traditional solutions.

Maintenance: With traditional optical modules, you can simply unplug and replace them if they fail. But with CPO, the module is soldered directly onto the chip, making repairs extremely difficult. Redundancy and fault-tolerant designs are required to compensate.

Thermal management: The optical engine and chips are densely packaged together, which may cause local temperatures to exceed the laser's tolerance limit, requiring a more efficient cooling solution.

Standardization: NVIDIA, Broadcom, and others are each promoting their own solutions; a unified industry standard has not yet been established, making it difficult for upstream and downstream players to develop and produce based on a common interface.

In addition to CPO, several related technical pathways are being pursued in parallel. Clarifying these will help you understand each company’s competitive position.

3.1 NPO (Near-Package Optics)

NPO is a "simplified version" of CPO, where the optical engine is placed on the same PCB substrate rather than being packaged onto the ASIC's substrate or interposer. The distance is reduced, but not to the extent of CPO's "face-to-face" integration.

This is a pragmatic compromise, especially in the Chinese market, where advanced packaging capacity at the level of TSMC is lacking; Alibaba, Huawei, and others are actively promoting NPO. Huagong Technology has already launched the world’s first 3.2T NPO product, deployed with leading customers.

NPO can be regarded as a "transitional state" of CPO; it will dominate the Chinese market in the short term but will eventually evolve toward CPO in the long term.

3.2 OIO (Optical I/O)

If CPO involves packaging the optical engine together with the switching chip, then OIO is a more aggressive version that directly packages the optical engine with the compute chip (GPU/XPU), even integrating them at the chip level.

OIO is entirely focused on in-rack scenarios (scale-up), replacing copper cables. Ayar Labs is a pioneer in this field and has already demonstrated a full CPO scale-up rack prototype with Wiwynn at OFC 2026.

OIO is expected to see large-scale adoption in GPU interconnection scenarios between 2028 and 2030.

3.3 LPO (Linear Drive Pluggable Optical)

LPO is a streamlined upgrade to traditional optical modules, directly removing the most power-hungry DSP chip and using analog amplification instead. The benefits are lower power consumption and reduced cost; the drawbacks are higher signal quality requirements, limited long-distance transmission, and performance bottlenecks above 1.6T speeds.

LPO can be seen as a "lifesaving solution" for traditional optical modules, but it does not alter the overall trend toward CPO.

3.4 OCS (Optical Circuit Switch)

OCS is a special type of switch that does not perform optical-to-electrical conversion; instead, it directly reflects optical signals within the optical domain using a "micro-mirror array," like a series of adjustable-angle tiny mirrors that "bounce" light in different directions.

Google is the biggest advocate of OCS, using OCS to replace traditional spine switches. The advantage of OCS is its extremely low power consumption (no need for optical-electrical conversion), but it can only "forward" optical signals without "decision-making" capability (it cannot unpack packets to inspect addresses and determine routing). Therefore, OCS is suitable only for replacing the spine layer and cannot fully replace leaf switches.

CPO and OCS have a more complementary relationship: OCS manages all-optical forwarding at the Spine layer, while CPO handles electro-optical conversion at the Leaf and server layers. Both can coexist without conflict.

3.5 Summary of Technical Approach

CPO is not a standalone product, but a complex systems engineering project involving numerous upstream and downstream processes. Understanding these processes is key to recognizing investment opportunities.

4.1 Top-level architecture designer, "the client of clients"

One of the most profound changes in the CPO era is the shift in bargaining power along the supply chain.

In the traditional pluggable era, optical module manufacturers could independently define products and ship them independently. With CPO, the optical engine is soldered into the chip package; whoever defines the chip architecture defines CPO. Control has shifted from optical module manufacturers to platform providers and switch chip manufacturers.

NVIDIA (NVDA): The most aggressive player advancing CPO today, having sequentially launched the Quantum-X and Spectrum-X series of CPO switches at GTC 2025/2026, and securing upstream light source and fiber capacity in early 2026 through a $4 billion investment in Lumentum and Coherent, and a $500 million commitment to Corning.

Broadcom (AVGO): The actual pioneer in CPO mass production. Its Tomahawk series of CPO switches began with the first-generation Humboldt in 2021, and by 2025, the Tomahawk 5-Bailly became the industry’s first mass-produced CPO solution, with over 50,000 units shipped annually. The third-generation 200G/lane platform is now on the horizon. Broadcom’s strategy leans more toward “selling shovels”—it does not manufacture complete systems but instead sells CPO switching chips to major cloud providers for their own assembly.

Marvell (MRVL): Following a customized approach by acquiring companies such as Celestial AI to integrate 3D SiPho optical engines into its proprietary XPU architecture, delivering highly integrated CPO computing platforms for specific clients.

Google (GOOG): A unique player that is both the biggest advocate of the OCS approach and a major customer of CPO. Google replaces spine-layer switches with OCS, but still requires CPO for electro-optical conversion at the leaf and server layers, making Google both a "competitor" to and a "buyer" of CPO.

4.2 Advanced packaging and manufacturing, welding light and electricity together

The core technical challenge of CPO lies in heterogeneous integration packaging, which involves packaging photonic chips (silicon photonics or InP) and electronic chips (CMOS ASIC) made from different material systems and processes onto the same substrate or interposer. This is not traditional packaging of "soldering components onto a board," but rather requires hybrid bonding technology with sub-micron precision, presenting a difficulty comparable to chip manufacturing itself.

TSMC (TSM): The absolute core of this segment. Both NVIDIA and Broadcom’s CPO solutions rely on TSMC’s COUPE platform and SoIC 3D packaging technology. In February 2026, TSMC advanced COUPE to risk production, with its 6.4T/package solution developed in collaboration with AMD expected to enter high-volume production in the second half of 2026. In essence, TSMC’s advanced packaging capacity and yield directly determine the pace of CPO mass production.

Advanced Semiconductor Engineering (ASE) (ASX): As the world's largest packaging and testing manufacturer, ASE is also a key participant in CPO advanced packaging.

Amkor (AMKR): Amkor in the United States is also vying for CPO manufacturing orders.

In the A-share market, Hua Tian Technology (002185) and JCET Group (600584) are the primary beneficiaries in the packaging segment.
Huatian Technology's packaging business directly benefits from the adoption of CPO technology; JCET, a subsidiary of JCET Group, participates in advanced packaging and possesses technical expertise in heterogeneous integration. However, it should be noted that at this stage, the core aspects of CPO packaging remain highly concentrated in TSMC's hands, with domestic packaging firms primarily benefiting from peripheral support and mid-to-low-end packaging and testing.

Worth highlighting separately is Fabrinet (FN), the leading EMS provider in optical precision manufacturing, which produces high-end optical modules for companies such as Coherent and Lumentum, playing a role similar to TSMC in the semiconductor industry.

4.3 Lasers, the "Heart" of CPO

If the chip is the "brain" of CPO, then the laser is its "heart"—without a light source, all photonic-electronic conversion would be impossible.

There is competition between two technological approaches in the laser industry.

EML lasers (electro-absorption modulated lasers) are a traditional approach that integrates laser emission and signal modulation onto a single chip, making them ideal for high-bandwidth, long-distance transmission. This technology has extremely high barriers to entry, with only a handful of global suppliers. Lumentum (LITE) was the first to mass-produce 200G EML in 2023 and demonstrated the world’s first 400G EML in 2025; Coherent (COHR, formerly II-VI) followed closely behind, with the two together holding over 80% market share. Sumitomo Electric (5802.T) and Mitsubishi in Japan are also strong players in traditional EML, but their capacity expansion has far lagged behind demand growth.

CW lasers (continuous wave lasers) are an emerging approach that completely separates "light emission" from "modulation": the laser only generates a steady, continuous beam of light, while modulation of the signal is handled by modulators on a silicon photonic chip.

The CW route offers lower power consumption and better cost efficiency, naturally aligning with CPO and silicon photonics architectures. More importantly, Chinese manufacturers have already achieved breakthrough progress on the CW route.

Source Crystal Technology (688498) holds over 30% global market share in 10G laser chips, has shipped millions of CW laser chips, and is currently developing and testing 100G EML products. In Q1 2026, revenue grew by 321%, and net profit increased more than 11-fold, making it one of the most dynamic players among upstream optical chip companies.

Shijia Photonics (688313)'s CW light sources have been validated and adopted by multiple leading manufacturers; the newly developed CWDFB laser achieves power exceeding 1000mW at 50℃.

Changguang Huaxin (688048) covers high-power semiconductor laser chips, VCSEL laser chips, and silicon photonics chips.

Yongding Co., Ltd. (600105) subsidiary Dingxin Photonics has built a rare domestic IDM laser chip factory, with mass production already underway for 100G EML and 100mW CW high-power silicon photonic light sources. Accelink Technologies (002281) is one of the few domestic companies with full-chain capabilities and in-house development of high-end optical chips, including EML.

In March 2026, NVIDIA invested $2 billion each in Lumentum and Coherent, with accompanying procurement commitments spanning from 2027 through 2030. Lumentum will use the funds to build a new wafer fab in the United States, with its laser production capacity expected to grow at an 85% CAGR from 2026 to 2030. Coherent will channel the investment into expanding its indium phosphide (InP) capacity at its Sherman, Texas facility. These investments send a clear signal: lasers represent the most critical and strategically valuable segment in the CPO supply chain, with the largest supply-demand gap.

4.4 Silicon photonic chip, the "brain" of the CPO optical engine

Silicon photonics is the mainstream implementation approach for CPO optical engines. Its core idea is to directly "pattern" optical components such as waveguides, modulators, and detectors onto a chip using standard CMOS silicon processes, treating optical elements like semiconductor devices. This approach naturally enables large-scale integration and allows shared manufacturing platforms with electronic chips, significantly reducing costs through mass production.

Overseas entities have accumulated deep expertise in silicon photonics.

Broadcom (AVGO) is one of the first semiconductor giants to invest in silicon photonics, with its CPO switch optical engine built on its proprietary silicon photonics platform.

Intel Photonics, a team under Intel (INTC), has over a decade of accumulated expertise in silicon photonics research and development. Although its activities in the consumer market have been limited, it has consistently been a key player in optical interconnects for data centers.

Marvell (MRVL) has integrated silicon photonics capabilities through acquisitions such as Celestial AI, with its 3D SiPho optical engine supporting 200Gbps optical interfaces. Cisco (CSCO) acquired Acacia Communications in 2019 for approximately $4.5 billion to gain access to a leading industry silicon photonics coherent technology platform.

Domestic manufacturers are also accelerating their pursuit.

LightSpeed Technology (002281) has achieved mass delivery capability for its 400G and 800G silicon photonic chips and jointly launched a 1.6T silicon photonic optical module with Cisco at OFC 2026.

Yuanjie Technology (688498) offers high-power silicon photonic light sources that complement silicon photonic modules.

Shijia Photonics (688313) is a market leader in PLC splitters and AWG chips, and is expanding into the silicon photonics chip domain.

Silicon photonics technology has high versatility and is compatible with various cutting-edge approaches such as CPO, LPO, and thin-film lithium niobate, making it a strategic focus for major manufacturers. Infinera previously disclosed that the proportion of silicon photonics solutions in its 800G products is rapidly increasing, indicating that silicon photonics is not only exclusive to CPO but is also increasingly being adopted in traditional pluggable optical modules.

4.5 Fiber optic connection components, a new opportunity spawned by CPO

If the previous stages primarily represent upgrades to existing markets, the optical interconnect components are a purely incremental market created by CPO. These components are scarcely used in traditional pluggable optical module solutions but have become essential in CPO architectures, making them one of the most elastic segments in the supply chain.

(1) FAU (Fiber Array Unit)

In traditional optical modules, fibers can simply be plugged into standardized interfaces. But with CPO, the fibers must be aligned with micron-level precision to the waveguides on the surface of the optical chip—any slight misalignment prevents proper light coupling. The FAU performs this task by precisely arranging and securing multiple fibers to ensure each one perfectly connects with its corresponding waveguide on the chip.

In traditional optical modules, a single FAU is worth about $15, but the polarization-maintaining FAUs used in CPO have seen their value surge to dozens or even $100. Based on NVIDIA’s 115.2T switch, each unit requires 72 FAUs, bringing the total FAU value per machine to $6,000–$7,000. From 2025 to 2026, the FAU market size is expected to grow rapidly from RMB 6–7 billion to over RMB 10 billion. Moreover, expanding FAU production is challenging with high yield requirements, resulting in tight supply.

(2) PMF (Polarization-Maintaining Fiber)

Traditional optical modules use direct modulation and are insensitive to the polarization state of light. However, CPO uses an external laser; during transmission of the laser through optical fiber to the optical engine, any change in polarization state can cause significant power loss. Polarization-maintaining fiber serves as a "dedicated channel" that ensures the light's polarization direction remains unchanged throughout, and although its cost is considerably higher than that of standard fiber, it is indispensable in CPO architectures.

(3) Fiber Shuffle (Fiber Distribution Box)

Traditional optical modules typically have only two fibers—one for transmission and one for reception—making manual cable management sufficient. However, under CPO, the number of fibers increases dramatically to dozens or even hundreds, requiring these high-density fibers to be rearranged and organized so that each fiber connects precisely from the optical engine to the correct external interface. Fiber Shuffle is essentially the data center’s version of a "cable organizer," and it is indispensable in CPO architectures.

(4) MPO (Multi-Fiber Optic Connector)

If the CPO reaches a rate above 400G, it requires eight or even 16 fibers to transmit in parallel, while panel space is extremely limited. MPO is a "multi-port plug" that can connect multiple fibers at once, and its demand has surged in the CPO era.

In this segment, Corning (GLW), a U.S. stock, is the undisputed global leader in optical fiber and optical materials, serving as a core supplier for FAU and fiber optics, as well as a strategic partner to NVIDIA in a $3.2 billion collaboration. In 2025, Corning’s optical communications business generated $6.3 billion in revenue, a 35% year-over-year increase, making it the company’s largest and fastest-growing business segment. Unlisted companies US Conec and SENKO are also key global players in the MPO/MTP connector market.

In the A-share market, Tianfu Communications (300394) is the absolute leader in this segment, offering a full range of products including FAU fiber arrays, LENS arrays, and MPO connectors, and is a core supplier for NVIDIA and Broadcom’s CPO solutions. In the first half of 2025, the proportion of active optical components increased by 8 percentage points year-over-year to 63.78%, primarily driven by growth in CPO-related packaging orders, with a gross margin of 42%.

TaiChenGuang (300570) is the domestic leader in MPO connectors, and its products have been indirectly certified by NVIDIA.

Guangku Technology (300620), in addition to its core business of lithium niobate modulators, has entered the mainstream supply chain with its 90-degree bent fiber array and has a unique presence in the OCS all-optical switching device sector.

Changxin Bochuang is an integrated photonic device supplier, offering a full range of MPO, AOC (active optical cables), and AEC products, and has entered the supply chains of Google and NVIDIA.

4.6 Fiber Optic Connection Components: The New Opportunity Created by CPO

Compared to traditional optical modules, CPO has significantly increased demand for precision fiber optic components. These components were rarely used in traditional solutions but have become essential in CPO architectures, making them one of the most elastic growth segments in the supply chain.

(1) FAU (Fiber Array Unit)

In CPO, optical fibers must be aligned with the waveguides on the optical chip surface with micrometer-level precision, and that’s what FAUs do. In traditional optical modules, a single FAU costs about $15, but the polarization-maintaining FAUs used in CPO can cost dozens or even up to $100 each. Based on NVIDIA’s 115.2T switch, each system requires 72 FAUs, resulting in a total value of $6,000–$7,000.

From 2025 to 2026, the FAU market size is expected to grow rapidly from RMB 6-7 billion to over RMB 10 billion.

(2) PMF (Polarization-Maintaining Fiber)

Traditional optical modules are insensitive to the polarization state of light, but CPO uses an external laser source, and if the polarization state changes, significant optical power loss can occur. Polarization-maintaining fiber serves as a "dedicated channel" that ensures the polarization state of light remains unchanged throughout transmission.

(3) Fiber Shuffle (Fiber Distribution Box)

The number of optical fibers under the CPO has surged, requiring the complex, high-density fibers to be rearranged and organized—like a cable manager for data centers. Traditional optical modules, with only one transmit and one receive fiber, don't need this at all.

(4) MPO (Multi-Fiber Optic Connector)

If the CPO reaches above 400G, it requires eight or even 16 fibers to transmit in parallel. MPO is a "multi-port plug" that can connect multiple fibers at once, and its demand has surged in the CPO era.

4.7 Fiber Optic Cables, the Infrastructure Backbone of the CPO Era

Although optical fiber cables are not direct components of CPO modules, they serve as the physical medium for optical interconnection—without fiber, optical signals have nowhere to travel. The explosive construction of AI data centers is driving demand for optical fiber into a supercycle.

The simultaneous surge in volume and price during this cycle is extremely rare. In March 2026, the price of China’s G.652.D single-mode fiber soared to 83.4 yuan per core-kilometer, more than doubling from January levels and setting a new all-time high. The last time a price increase of this magnitude occurred was during the peak of China’s Broadband China initiative in 2018. On the demand side, the four major North American cloud providers have planned a combined capital expenditure of $725 billion for 2026, a 77% year-over-year increase; Meta alone signed a long-term optical cable contract worth $6 billion with Corning.

Corning (GLW), a U.S. stock company and global leader in optical fiber preforms, is increasing its domestic optical connectivity manufacturing capacity in the U.S. tenfold with support from NVIDIA’s $500 million investment.

Changfei Fiber Optic (06869/601869), listed on both the Hong Kong and A-share markets, is the world’s largest manufacturer of optical fiber preforms and optical fibers. In Q1 2026, its net profit surged 226% year-over-year. At OFC 2026, Changfei showcased hollow-core fiber with a single spool length of 91.2 km and an attenuation of only 0.04 dB/km, reaching globally leading levels and representing the next generation of optical fiber technology.

Zhongtian Technology (600522) is one of China's leading optical cable manufacturers, with end-to-end capabilities in both submarine and terrestrial cables.

Hengtong Optical Fiber & Cable (600487) offers a full range of fiber optic cable products and has forward-looking initiatives in F5G solutions.

FiberHome Communications (600498) is a core enterprise in the Wuhan Guanggu optical communication industry chain, backed by China Information and Communication Technology Group.

4.8 PCB/board, the backbone of CPO

Whether it's traditional optical modules or CPO switches, both rely on high-performance PCBs (printed circuit boards) and ABF substrates. However, the CPO era has brought a qualitative shift in PCB requirements: higher signal integrity is essential (since the optical engine is positioned close to the ASIC, demanding stricter signal trace precision), low-loss materials have become mandatory (high-end materials like Megtron 6/7 cost 5–8 times more than standard FR-4), and enhanced multi-layer stacking capabilities are required. Meanwhile, PCBs for optical modules themselves are also evolving toward higher speeds—the PCBs used in 800G/1.6T optical modules carry significantly higher value than those in previous generations.

Shenghong Technology (300476) is undoubtedly the leading AI player in this segment. It is a core supplier of server substrates for NVIDIA’s GB200, with AI server PCB revenue accounting for over 50%. In the optical communication field, Shenghong has achieved mass production of 800G switch PCBs and industrial-scale manufacturing of 1.6T optical module PCBs, serving both CPO and optical module application scenarios. Its global market share in AI computing PCBs leads the industry, making it the most comprehensive player in the intersection of CPO and PCB technologies.

Dongshan Precision (002384) follows a dual-main-business strategy focusing on AI computing PCBs and optoelectronic modules, with net profit growing 119%-152% year-over-year in Q1 2026, driven primarily by accelerated investment in AI infrastructure.

Shanghai PCB Co., Ltd. (002463) is a traditional leader in high-speed PCBs for data centers, steadily supplying products to major global server and switch platforms.

Shennan Circuits (002916) differentiates itself through its advanced IC substrate capabilities, enabling it to capture higher-value segments ranging from PCBs to chip packaging substrates.

4.9 DSP and SerDes chips, redefined by CPO

In traditional pluggable optical modules, the DSP (digital signal processor) is the single component with the highest power consumption and cost, responsible for repairing damaged electrical signals during transmission—indispensable, yet a "power hog."

One of the most significant power savings in the CPO solution comes from eliminating the standalone DSP chip. However, this does not mean that signal processing tasks disappear—they are simply redistributed: the core functions of the DSP are integrated into the switching ASIC, while CDR (Clock Data Recovery) is integrated into the high-speed SerDes. SerDes (Serializer/Deserializer) resides within the ASIC chip and is responsible for packaging parallel data from inside the chip into high-speed serial data streams, or converting received serial streams back into parallel data. CPO demands that SerDes speeds leap from the current 112 Gbps to 200 Gbps or higher, placing extremely high requirements on ASIC design capabilities.

Broadcom (AVGO) is the undisputed leader in integrated ASIC and SerDes switching designs; its Tomahawk series chips feature built-in high-speed SerDes that directly drive CPO optical engines without requiring additional signal conditioning chips.

Marvell (MRVL) has a unique advantage in customized switching ASICs, enabling it to build integrated CPO computing platforms tailored for specific clients.

Specializing in SerDes and connectivity chips, Astera Labs (ALAB) positions itself as a smart connectivity chip provider, offering PCIe/CXL retimers and SerDes IP. Credo (CRDO) focuses on high-speed SerDes IP cores and holds a significant share in the data center connectivity market. Alphawave Semi (AWE), listed in London, is also a major player in high-speed connectivity IP.

4.10 Optical Module Manufacturers: From Main Actors to Transforming Players

In the traditional pluggable era, optical module manufacturers were the absolute key players in the supply chain, independently procuring optical chips, electrical chips, and mechanical components to assemble complete optical modules sold directly to data center customers. But with CPO, the optical engine is integrated into the ASIC package, weakening the role of standalone optical modules and leaving optical module manufacturers with a fundamental question: Will my slice of the pie be eaten?

The answer is: Not in the short term, but a transition is necessary in the long term.

In the short term, pluggable optical modules remain in a super cycle of prosperity. In Q1 2026, InnoLight (300308) achieved revenue of nearly RMB 19.5 billion, a 192% year-over-year increase, and net profit of RMB 5.7 billion, up 262% year-over-year. Before CPO fully replaces pluggable modules, demand for 800G/1.6T optical modules continues to grow at a doubling rate. NeoPhotonics (300502) is also accelerating volume production of its 1.6T products. Among the top 10 global optical module manufacturers, seven are Chinese companies, with InnoLight firmly holding the number one position.

In the medium term, optical module manufacturers are pursuing multiple strategies in preparation for the CPO era. First, they continue to supply plug-in optical modules at 800G/1.6T/3.2T speeds to fully capitalize on current market profits. Second, they offer transitional solutions such as NPO and LPO; Huagong Technology (000988) has already launched the world’s first 3.2T NPO product, which is now being adopted by leading customers. Third, they are transitioning into suppliers of CPO optical engines, shifting from selling entire vehicles to selling engines—a natural evolution, as the core processes of optical engines (optical chip packaging, fiber coupling, and testing/validation) closely overlap with those of optical modules. Fourth, they are entering the OCS all-optical switch market; InnoLight has leveraged digital liquid crystal technology, backed by Google and Amazon, to enter this space.

LightSpeed Technology (002281), a seasoned optical communications leader with state-owned enterprise backing, has integrated the entire value chain from chips to components, modules, and subsystems, and is now capable of mass delivery of 1.6T silicon photonic modules.

U.S. stocks Coherent (COHR) and Fabrinet (FN) are also key players in the optical module market; the former is a leading player in both optical modules and optical chips, while the latter, known as the "king of contract manufacturing," handles nearly all high-end optical modules. Management recently stated that CPO is "more real than ever" and has already begun generating related revenue.

Short-term (2026–2027)

This is the "final feast" for pluggable optical modules and the "from 0 to 1" phase of CPO.

800G/1.6T pluggable optical modules remain in short supply, with leading companies such as InnoLight and Eoptolink continuing to experience strong performance growth. Meanwhile, CPO is beginning its first large-scale shipments (primarily at the Spine switch level), driven by NVIDIA and Broadcom.

Key beneficiary segments: optical modules (InnoLight, NeoPhotonics), laser devices (Lumentum, Coherent, Source Photonics), and fiber optic connectivity components (TFC, Tech-Fiber).

Medium-term (2027–2029)

CPO is expanding from Spine to Leaf, and the share of pluggable optical modules in scale-out scenarios is being eroded by CPO. NPO reached its peak in the Chinese market as a transitional solution. 3.2T modules are now commercially available.

Key beneficiary segments: Advanced packaging (TSMC), external laser modules (value increased 3-4 times), and FAU/MPO (both volume and price rising).

Long-term (2029–2032+)

CPO is penetrating scale-up (within racks), OIO technology is being commercially deployed in GPU interconnection scenarios, and copper cables are being largely replaced by optical interconnects. CPO is expected to achieve a 35% penetration rate in AI data center optical communication modules by 2030.

Key beneficiary segments: OIO-related manufacturers (Ayar Labs), silicon photonics platforms, and the entire optical interconnection industry chain.

If GPUs are the "brain" of AI, HBM is the "memory," and electricity is the "food," then optical interconnects are AI's "nervous system"—without it, even the most powerful brain cannot connect with the world.

Huang Renxun made it clear: energy is our most important resource, and the core value of CPO lies in fundamentally reducing energy consumption in data transmission by replacing electricity with light.

On this赛道, the U.S. holds the authority to define the architecture (NVIDIA, Broadcom) and dominates high-end optical chips (Lumentum, Coherent), while TSMC controls the critical aspects of packaging and manufacturing. Chinese companies have established strong competitive barriers in optical module assembly (InnoLight, Eoptolink), fiber optic connectivity components (TFF Communications), CW lasers (Source Photonics), and optical fiber and cable (FiberHome).

Over the coming years, the investment logic for this trillion-dollar sector will gradually shift from selling shovels (optical modules) to building highways (CPO/OIO infrastructure), and the ultimate winners will be companies that can keep pace with technological iteration while securing critical bottlenecks in the supply chain.

Disclaimer: This article is solely for the purpose of organizing industry chain knowledge and does not constitute any investment advice. The companies and assets mentioned herein are not recommended; investing carries risks, and participants should proceed with caution.