AI Data Centers Face a New Bottleneck: Optical Interconnects

Organized & Compiled by Shenchao TechFlow

Host: Nico

AI Optical Interconnect: The Next Trillion-Dollar Sector Overshadowed by GPUs?

Podcast source: Nico Frontier Alpha

Broadcast date: May 8, 2026

Optical interconnects are evolving from mere accessories for GPUs into the core bottleneck of AI data centers. When hundreds or even thousands of GPUs need to work together within a single rack, across racks, or across super-nodes, what truly determines computational utilization is no longer just the chips themselves, but the data transfer capability between GPUs.

This episode of the podcast presents a产业链 research perspective, connecting optical modules, silicon photonics PICs, CPO, external lasers, InP substrates, SOI substrates, foundry services, and packaging and testing into a single diagram, and provides a layered allocation framework ranging from AVGO, MRVL, and GLW to COHR, LITE, TSEM, and further to SIVE, AAOI, AXTI, IQE, and Soitec.

The most noteworthy aspect of this content is not a single stock recommendation, but rather the insight that the AI infrastructure competition is shifting from “who has more GPUs” to “who can secure the scarcer optical interconnect supply chain,” with CPO (Co-Packaged Optics) potentially being the largest incremental variable.

Why has optical interconnect suddenly become important?

• Even if an NVIDIA GB300 GPU accelerator has immense computing power, most of that power would be wasted if it cannot communicate at high speed with thousands of other GPUs.
• Insufficient interconnect bandwidth means spending more money on GPUs will yield diminishing returns.
• Whether for training or inference, when collaborative work is involved, GPUs must exchange data at high speeds—this data pathway is the interconnect.
• Optical interconnects are not hype; the interconnect demands of AI data centers are real, urgent, and irreversible.

Copper cables phase out, fiber optics take over

• The transmission speed of copper cables has approached its physical limit, and the bandwidth a single copper wire can support has reached its maximum.
• Beyond a few meters, copper cables begin to experience signal attenuation and interference, but AI data center connections often span tens to hundreds of meters.
• The bandwidth of fiber optics is dozens of times greater than that of copper cables, works perfectly over distances of several kilometers, and consumes so little energy that it can be ignored.

The industrial nature of optical modules

• The optical module handles communication between different cabinets, not between GPUs within the same cabinet.
• The optical module industry chain and the GPU industry chain are not two separate sectors; rather, GPU shipment volumes directly drive demand for optical modules.
• The manufacturing of an optical module spans two entirely different semiconductor process systems: InP compound semiconductors for the optical chip and silicon for the DSP chip.

The true meaning of CPO

• CPO doesn't disrupt just a component within the optical module—it disrupts the optical module as a product form itself.
• CPO is not an upgrade of existing products, but a structural overhaul.
• A more accurate relationship is that CPO opens up an entirely new market, one that is much larger than the pluggable optical module market, rather than simply replacing the existing one.

Industry chain investment framework

• The optical interconnect industry chain is not dominated by a single company like NVIDIA in the GPU market; it is a highly specialized industry chain with bottlenecks widely distributed across multiple segments.
• The further upstream you go, the smaller the companies, the greater their flexibility, but the lower their certainty; the further downstream you go, the larger the companies, the higher their certainty, but the less their flexibility.
• If you can tolerate high risk and high volatility, the core strategy is to identify bottlenecks; behind each bottleneck, there are often only one or two companies capable of addressing it.

Beyond GPUs, the truly scarce "neural networks" in AI infrastructure

Over the past two to three years, nearly everyone has been discussing GPUs and computing power. Since the emergence of ChatGPT—an AI product launched by OpenAI that ignited a wave of large model applications—and the subsequent explosion of the AI technology revolution, NVIDIA’s stock price has surged fifteenfold in three years, making computing power an indispensable keyword in the realm of AI large models. The semiconductor industry chain centered around GPUs has also entered a golden age that transcends economic cycles.

But over the past year, a component as critical—and even more scarce—as the GPU has been quietly exploding. In large-scale data center deployments, even the most powerful NVIDIA GB300 GPU accelerator card will waste most of its computing power if it cannot communicate at high speed with thousands of other GPUs. Without sufficient interconnect bandwidth, buying more GPUs yields diminishing returns. The component responsible for enabling high-speed communication among thousands of GPUs is optical interconnect.

According to data from LightCounting (a research firm in the optical communications field), the global optical module market size doubled in 2024 to $15.4 billion, and is projected to grow a further 55% in 2025 to $23.8 billion. Under an optimistic scenario, LightCounting expects the total market size of the entire optical interconnect industry chain to exceed $110 billion by 2030.

However, most investors may not have heard of the companies along this supply chain. SIVE/SIVEE, with annual revenue of approximately $30 million, has surged tenfold since the start of 2026; TSEM (Tower Semiconductor, an Israeli specialty foundry) is often called the "TSMC of optical interconnects," with 70% of its capacity pre-booked through 2028; and COHR (Coherent, a vertically integrated company in optics and materials), with annual revenue of about $5.8 billion, has received a $2 billion strategic investment from NVIDIA.

In today’s episode, we’ll break down the entire optical interconnect industry chain from start to finish: What is optical interconnect? What’s inside an optical module? What are the next-generation technology pathways? Where are the key bottlenecks in the supply chain? Where does each company stand? And how can investors allocate across this sector according to their risk preferences.

Training, Inference, and Interconnection: Why GPUs Must Communicate at High Speed

Before discussing specific companies, let’s first address a key question: why has optical interconnect suddenly become one of the most critical and scarce components in AI infrastructure? This begins with understanding how AI works—AI operations consist of two phases: training and inference.

Training involves feeding a model large amounts of text, images, and code so it can continuously learn and evolve based on existing data. The training parameters of a large model may reach trillions in scale, far exceeding the capacity of any single GPU; therefore, the model must be split into thousands of parts and distributed across thousands of GPUs for parallel computation. After each GPU completes its assigned portion, it must transmit intermediate results to other GPUs so they can collaborate to complete the entire task.

Inference is when AI applies knowledge it has already learned to generate an answer. When you ask ChatGPT a question and it responds after a few dozen seconds, that’s inference. Many people assume inference is simply a single GPU answering one question without any interconnectivity. This may have been close to the truth in 2023, but by 2026, it will be very different.

AI has evolved from simple question-and-answer interactions to advanced reasoning and agentic AI. The user interaction partner is no longer just a basic chatbot, but a complex agent capable of planning tasks, performing multi-step reasoning, and querying multiple data sources. Behind each interaction, hundreds or even thousands of GPUs may be working in coordination. Whether in training or inference, whenever coordination is involved, GPUs must exchange data at high speed—and this data pathway is the interconnect.

Why copper cables are no longer sufficient

In the past, interconnects primarily used copper cables to transmit electrical signals; now, this pathway is gradually being replaced by optical fibers that transmit light signals. Copper cables are no longer sufficient, mainly for three reasons.

First, the transmission speed of copper cables has approached its physical limit. No matter how much the materials and manufacturing processes are optimized, the bandwidth a single copper wire can carry has reached its maximum—just like a two-lane road, no matter how congested, can only accommodate two cars side by side. Second, the farther the distance, the worse the signal quality. Copper cables begin to experience attenuation and interference after just a few meters, yet connections in AI data centers often span tens or even hundreds of meters—far beyond what copper can handle. Third, copper cables consume more power. Each generation of GPUs has seen increasing power consumption: the H100 consumes 700 watts, the B200 has risen to 1 kilowatt, and the GB300 will be even higher. At these power levels, copper connections between GPUs themselves can consume substantial amounts of electricity.

Fiber optics are completely different. A single fiber optic cable can achieve bandwidth dozens of times greater than copper cable, transmit signals over distances of several kilometers without issue, and consume so little power that it’s negligible. Fiber optics can also transmit multiple optical signals at different wavelengths simultaneously—like a highway divided into eight lanes, each carrying light of a different color without interference. One fiber optic cable can effectively replace dozens of copper cables.

Three stages of optical interconnection

The use of light in data centers is not a sudden new development, but rather one that has progressed through several distinct stages, with each stage bringing optical technology closer to the chip.

The first phase was before 2020. At that time, fiber optics were primarily used between data centers—for example, cloud providers had data centers in Beijing and Shanghai, over a thousand kilometers apart, requiring fiber connections. However, within data centers, servers were still mostly connected using copper cables.

The second phase was from 2023 to 2024. Although ChatGPT ignited the AI technology revolution at the end of 2022, and GPU sales surged the following year, the optical module market did not initially show significant growth. This was because NVIDIA GPU clusters at the time primarily used copper cables, making optical modules non-critical components. Worse still, at the beginning of 2023, cloud providers, fearing an economic downturn, cut capital expenditures; Meta (the parent company of Facebook and one of the world’s largest buyers of cloud and AI infrastructure) even slashed more than half of its optical module deployment plans.

The real turning point came in 2024. Cloud providers expanded their GPU clusters from hundreds to thousands, even tens of thousands, rendering copper cables with transmission distances of just a few meters completely inadequate. NVIDIA replaced copper cables with pluggable optical modules in its reference architecture—a shift at the architectural level that ignited the market, causing the optical module market size to double in 2024.

The third phase spans from 2025 to the present. NVIDIA’s Blackwell (NVIDIA’s next-generation AI GPU architecture) has begun large-scale deployment, driving higher power consumption and greater interconnect bandwidth demands, further surging the need for optical modules. Meanwhile, the combined capital expenditures of the five major cloud providers over the first nine months exceeded $300 billion, setting a new record, causing optical module demand to surpass supply by more than double and creating severe supply-demand imbalances. In March of this year, NVIDIA invested $2 billion each in Lumentum and Coherent. At GTC 2026 (NVIDIA’s annual developer conference), NVIDIA unveiled its CPO solution and optical interconnect design for the next-generation Rubin architecture, effectively signaling that optical interconnects have transitioned from a niche segment to a core narrative in AI infrastructure.

What is an optical module: A translator between electrical and optical signals

Before diving into the investment research content, let’s first cover a few fundamental concepts. The first is the optical module. GPU chips themselves only process electrical signals, while fiber optics transmit optical signals—these two use different “languages,” so a translator is needed to convert electrical signals into optical signals for transmission, and then convert received optical signals back into electrical signals. This translator is the pluggable optical module.

The optical module is about the size of a USB drive, with one end plugged into the server’s network card and the other end connected to fiber optic cable. In large AI data centers, there may be tens of thousands or even hundreds of thousands of these “small boxes.” There’s a commonly misunderstood concept here: optical modules handle communication between different racks, not communication between GPUs within the same rack.

Taking NVIDIA’s GB300 NVL72 (NVIDIA’s rack-scale GPU system) as an example, a single rack contains 72 GPUs interconnected via NVLink and NVSwitch (NVIDIA’s high-speed GPU interconnect technology and switching chips), all using copper electrical signals over distances of only tens of centimeters to one or two meters, requiring no optical transmission. Optical modules are only needed when data must travel between racks over distances of several meters, tens of meters, or more.

In a full AI cluster, optical modules are typically inserted into two locations: the server network card and the switch. Each fiber optic cable requires an optical module at both ends. The more GPUs and racks there are, the greater the inter-rack connectivity demand, and thus the higher the demand for optical modules. The optical module supply chain is not an independent sector from the GPU supply chain; rather, GPU shipment volumes directly drive demand for optical modules.

Five core components of an optical module

A USB drive-sized optical module typically contains five core components: a laser chip, a modulator chip, a detector chip, a DSP chip, and a lens and fiber coupling assembly.

First is the laser chip. Its function is to emit light, producing a stable, continuous laser beam that serves as the carrier for optical signals. The laser is like a tiny flashlight—smaller than a fingernail—but emits light that is highly precise and pure. The most critical aspect of the laser is its material. GPUs and CPOs use silicon, while lasers use indium phosphide (InP) or gallium arsenide (GaAs). Silicon is inherently poor at emitting light; the atomic structure of compound semiconductors like InP and GaAs is better suited for generating photons, which explains why laser chips are not manufactured by silicon-based foundries like TSMC.

Second is the modulator chip. The light emitted by the laser itself carries no information—it is merely “blank light.” The modulator’s function is to encode electrical signals onto the light. The GPU sends binary electrical signals of 0s and 1s, and the modulator controls the laser’s on/off state or intensity to represent these 0s and 1s with light. Continuing the earlier analogy, the laser is a flashlight that remains constantly on, while the modulator is the hand that flips the switch—pressing it tens of billions of times per second. Sometimes, the modulator is integrated onto the same chip as the laser, known as an EML (Electro-Absorption Modulated Laser), effectively combining the flashlight and its switch into a single component.

Third is the detector chip. The modulator converts electrical signals into optical signals for transmission; on the receiving end, the optical signal must be converted back into an electrical signal, which requires a detector. It acts like the receiver’s ear, outputting a 1 when it detects light and a 0 when it does not. Detectors are typically made using InP or GaAs material systems.

Fourth is the DSP chip (Digital Signal Processor). It acts as the brain of the optical module, responsible for error correction, encoding, and signal equalization. During optical signal transmission, noise and distortion can occur—similar to making a phone call on a busy, noisy highway, where the other person’s voice may be hard to hear. The DSP encodes the signal in a specialized way at the transmitter and cleans up noise at the receiver to ensure the restored 0s and 1s match the original data. DSP chips are silicon-based and belong to the same semiconductor manufacturing ecosystem as GPUs and CPOs, typically produced by silicon foundries such as TSMC.

800G and 1.6T refer to the transmission speeds of optical modules. 800G means transmitting 800 gigabits per second, while 1.6T means 1.6 terabits per second—double the speed. Optical modules have evolved from 400G to the current mainstream 800G, and now to the newly deploying 1.6T. As speeds increase, the complexity of chip design rises, along with the cost and design challenges of DSPs, which can sometimes exceed the cost of lasers.

Fifth is the lens and fiber coupling component. It must precisely align the laser's output with the fiber's entrance. The laser beam is extremely narrow, and the fiber core is even finer—only one-tenth the diameter of a human hair—with alignment accuracy required at the micrometer level. Imagine threading a needle with another needle, and doing this automatically millions of times on a factory production line.

Connecting these five components makes the optical module’s workflow clear: The GPU transmits electrical signals, which first enter the DSP for encoding and error correction, then proceed to the modulator; the modulator encodes the electrical signals onto the light emitted by the laser; the light passes through a lens into an optical fiber, traveling tens to hundreds of meters; upon reaching the other end, the light exits the fiber and is focused by a lens onto a detector; the detector converts the light back into electrical signals, which are sent to the receiving DSP for decoding and error correction, and finally delivered to another GPU.

How optical modules are made: Two sets of semiconductor processes coexist

Many people automatically assume that chips are all made by TSMC, so the chips in optical modules must be similar. But the reality is completely different. An optical module contains two entirely different types of chips, made from two distinct materials, and manufactured in two separate types of facilities.

The first category consists of DSP chips, which serve as the "brain" of optical modules and are responsible for error correction coding. These are silicon-based chips, manufactured using processes similar to those used for GPUs and CPO, by silicon foundries such as TSMC. Major DSP design companies include AVGO (Broadcom, a leader in communications and custom AI chips), MRVL (Marvell Technology, a company specializing in data center and networking chips), and CRDO (Credo, a data interconnect chip company).

The second category is optical chips, including lasers, modulators, and detectors, which are made from compound semiconductor materials such as InP. Some companies handle both design and manufacturing in-house, such as LITE (Lumentum, a provider of optical communication devices and lasers), COHR (Coherent, a company specializing in optical materials and devices), and AAOI (Applied Optoelectronics, a U.S.-based company specializing in optical modules and devices). There are also specialized small companies focused solely on laser design, such as SIVE/SIVEE, which push laser technology to its limits before outsourcing manufacturing to foundries.

Optical chips cannot be directly manufactured by TSMC because TSMC’s entire production line, equipment, chemicals, and process parameters are designed for silicon. InP is a completely different material, with different wafer sizes, etching chemicals, and growth temperatures—it simply won’t work on TSMC’s production line. Therefore, optical chips have their own independent manufacturing ecosystem.

Substrate and Epitaxy: The Two Foundations of Optical Chip Manufacturing

To understand optical chip manufacturing, you must first grasp two concepts: substrate and epitaxy. The substrate is the starting point for all optical chip fabrication—a special thin wafer on which all subsequent functional structures are grown. For example, if you wanted to grow a laser tree that emits light, you couldn’t just plant the seed in ordinary sand; you’d need a special soil whose molecular structure matches the seed, allowing it to take root and grow. Ordinary silicon is like sand—it’s unsuitable for light emission—while InP is that special soil.

The quality of the substrate directly determines the quality of all structures built upon it. If there is an atomic-level defect in the substrate, that defect propagates layer by layer like a crack, causing the laser chip to fail specifications and preventing the optical module from entering production. Manufacturing high-purity InP substrates is extremely challenging, and only a handful of factories worldwide can consistently achieve this standard.

Even with a substrate, you cannot directly make a chip—you must grow functional layers layer by layer on top of the substrate, a process called epitaxial growth. The laser emits light not because the substrate itself glows, but because of the specialized structures grown on top of it. When current passes through the epitaxial layers, electrons and holes recombine to release photons—this is the source of the laser light.

Each epitaxial layer is only a few nanometers thick, and dozens of layers stacked together resemble a layered pastry. Each layer must meet extremely precise requirements for composition, thickness, and doping concentration—deviating by even a single atomic layer shifts the wavelength of light, rendering the laser unusable.

The InP substrate is provided by AXTI (a U.S.-based compound semiconductor substrate supplier), and epitaxy is performed by IQE/IQEE (a U.K.-based compound semiconductor epitaxial wafer supplier). After epitaxy, laser chip manufacturing follows two pathways: one is Fabless (design and manufacturing separated), such as Sweden’s SIVE/SIVEE handling laser design and outsourcing production to Taiwan’s Win Semi (Wingtech Semiconductor, a compound semiconductor foundry); the other is IDM (Integrated Device Manufacturer), where LITE, COHR, and AAOI handle everything in-house—from epitaxy and laser fabrication to modulator and detector production, and finally optical module assembly.

Therefore, the manufacturing of an optical module spans two entirely different semiconductor process systems: InP compound semiconductors for the optical chip and silicon for the DSP chip. These two processes are incompatible and cannot be performed on the same production line. If capacity is constrained at any stage, the entire optical module cannot be shipped.

This also explains why optical companies don’t easily enter the DSP space, and why digital chip companies don’t easily enter the laser market. Optical chip design and digital chip design are two entirely different disciplines. Optical engineers understand laser physics, optical waveguide theory, and quantum well structures; digital chip engineers understand logic circuits and digital signal processing algorithms. Their skill sets do not overlap—just as cardiac surgeons and neurosurgeons are both surgeons, they cannot arbitrarily swap surgical procedures.

Here lies the most interesting aspect of the optical interconnect industry. Unlike GPUs, which are dominated solely by NVIDIA, this is a highly specialized产业链 with bottlenecks distributed across many players. Precisely because of this dispersion, ordinary investors have the opportunity to identify small companies overlooked by the market.

CPO: Move the optical components from the back of the server to beside the chip.

Pluggable optical modules are just the current solution. More notably, this entire industry chain is about to undergo a fundamental restructuring. A next-generation technology called CPO is completely reimagining the optical interconnect architecture.

CPO stands for Co-Packaged Optics, known in Chinese as co-packaged optics. It addresses the issue of optical modules being too far from GPUs. In current standard solutions, optical modules are plugged in as removable units on the back of servers; electrical signals generated by the GPU must travel tens of centimeters over copper traces to reach the back of the server before being converted into optical signals within the module. These tens of centimeters of copper wiring cause energy loss, latency, and heat generation. As AI cluster densities continue to increase, this small amount of loss is amplified hundreds of thousands of times, becoming a serious problem.

The CPO concept involves moving optical components from the back of the server into the chip package, placing them directly next to the GPU or switch chip, reducing the electrical-to-optical conversion distance from tens of centimeters to just a few millimeters. For example, current solutions keep food and soup separate—the GPU is in the rice container, while the optical module is in a separate cup; CPO pours the soup into a dedicated compartment inside the same rice container—food and soup remain distinct, but now share the same box, separated by only a few millimeters.

However, moving optical components inside the chip package faces a major obstacle: traditional optical modules use InP for their optical chips, while GPUs use silicon; the packaging processes for InP and silicon are incompatible, making it impossible to simply integrate an InP chip with a silicon-based GPU in the same package. The solution is to use silicon to make optical chips, which leads to silicon photonics PICs.

PIC stands for Photonic Integrated Circuit, known in Chinese as a photonic integrated circuit. Just as the ICs we are familiar with integrate billions of transistors onto a single chip for computation, PICs follow a similar concept—but integrate optical components instead of transistors. Silicon photonics PICs integrate functions such as modulators, optical waveguides, and detectors onto a single silicon-based chip. Because it is silicon-based, it can be packaged using techniques similar to those used for GPUs, something that InP optical chips cannot achieve.

Silicon photonics PICs do not use standard silicon wafers, but rather a special sandwich-structured wafer called SOI (Silicon-On-Insulator). An insulating layer is added between the substrate and the top silicon layer, allowing optical signals to propagate within the thin top silicon layer without leaking downward. In standard silicon wafers, which are solid blocks of material, light entering the material scatters uncontrollably. In SOI, the insulating layer in the middle acts like a mirror, reflecting the light back into the top layer and guiding it along the designed pathways.

In the SOI substrate segment, France’s Soitec is one of the core suppliers, holding a market position close to monopolistic. The primary foundry for silicon photonic PICs is TSEM (Tower Semiconductor). TSEM fabricates silicon photonic chips on SOI substrates using a modified CMOS process—a process that TSMC is not familiar with—making TSEM the foundry with the highest market share in this niche.

However, silicon has inherent limitations—it does not emit light. Therefore, silicon photonic PICs can only manipulate light but cannot generate it; a light source still requires an InP laser. This forms the core architecture of CPO: a silicon photonic PIC is placed inside the package to handle light manipulation tasks such as modulation, transmission, and detection; it is positioned side-by-side with the GPU on the same package substrate using advanced packaging technologies, at a distance of just a few millimeters, similar to how HBM memory sits next to the GPU.

Next to the silicon photonic PIC, there will be a driver chip responsible for converting between the GPU's electrical signals and the silicon photonic PIC's optical signals. It is also a silicon-based chip, essentially a significantly simplified version of the DSP found in traditional optical modules. Because CPO's electrical-to-optical conversion distance is only a few millimeters, it does not require the complex error-correction coding used by DSPs—a simple driver is sufficient.

For external packaging, an external laser source (ELS) is used. The laser transmits light through an optical fiber into the silicon photonic PIC inside the package. The laser is not integrated directly into the package because InP lasers generate significant heat and would cause issues when placed alongside GPUs and silicon photonic PICs. Additionally, lasers have a limited lifespan; if integrated internally, their failure would render the entire multi-thousand-dollar chip unusable. By designing the laser as an external, pluggable module, it can be easily replaced if it fails, without affecting the chip itself.

What CPO truly disrupts is not a specific component within the optical module, but the optical module itself as a product form. Currently, a pluggable optical module is an independent small box containing a laser, modulator, detector, and DSP. CPO effectively disassembles this box: the silicon photonic PIC is directly packaged inside the chip, the laser becomes a separate external light source, the DSP is significantly simplified or even eliminated, and the small box on the back of the server is no longer needed. This is not an upgrade to the existing product, but a fundamental architectural redesign.

Why CPO becomes an investment theme in 2026

The concept of CPO has existed for many years—why has it suddenly become a hot investment theme in 2026? Goldman Sachs released a report stating that the potential market size for optical interconnects will expand from approximately $15 billion today to $154 billion by 2028, a ninefold increase, with CPO accounting for $91 billion. The core reason is simple: NVIDIA’s next-generation architecture has turned CPO from an option into a necessity.

In the current GB300 NVL72 system, 72 GPUs are housed in a single rack, with GPUs within the rack still connected via copper cables. However, as AI clusters scale to hundreds or even thousands of GPUs, inter-rack network connectivity becomes a bottleneck. NVIDIA is introducing a CPO solution for inter-rack network switches on its next-generation Rubin platform (NVIDIA’s upcoming AI platform codename), replacing traditional pluggable optical modules. This marks NVIDIA’s first official adoption of CPO on its own platform.

Moving to the next-next-generation Feynman (NVIDIA’s subsequent AI platform codename), CPO may even enter inside the rack to connect GPUs directly. This means light is gradually advancing from between racks to between GPUs. Lumentum’s CEO also confirmed on the latest earnings call that CPO will face a significant supply-demand imbalance, with demand far exceeding supply; CPO is Lumentum’s largest single growth driver and remains in a very early stage.

According to industry data, the current actual shipment volume of the CPO market is still very small, estimated at only $1.6 billion in 2026, primarily consisting of samples and small batches. However, if Goldman Sachs’ forecast materializes, the market could surge to $91 billion by 2028—a trajectory from zero to hundreds of billions in explosive growth. NVIDIA has already begun mass production of CPO switches in early 2026, Broadcom delivered CPO-related products to customers in October 2025, and TSMC has launched its COUPE (TSMC CPO Advanced Packaging Solution). The adoption of CPO by both NVIDIA and Broadcom indicates that it is no longer a distant concept, but is rapidly becoming a reality.

However, CPO will not fully replace pluggable optical modules in the short term. CPO primarily addresses the ultra-high-density interconnect requirements within AI clusters, such as GPU interconnects inside NVIDIA super nodes; data centers still have numerous other connectivity scenarios—including rack-to-switch, switch-to-switch, and data center-to-data center—that will continue to rely on pluggable optical modules for the foreseeable future. Therefore, a more accurate characterization is that CPO is creating a new market, potentially much larger than the pluggable optical module market, rather than simply replacing the existing one. Both technologies will coexist in different use cases.

Five beneficiaries after the CPO surge

If the CPO truly experiences a surge in the future, even entering a super cycle, the five most benefited segments of the industry chain are likely to be:

First is silicon photonics PIC foundry services. The CPO architecture mandates the use of silicon photonics PICs, as only silicon-based chips can be advanced-packaged with GPUs. Very few manufacturers are capable of silicon photonics PIC foundry services, making capacity one of the most critical bottlenecks.

Second is the silicon photonics substrate. Each silicon photonics PIC requires an SOI substrate; the surge in demand for silicon photonics PICs driven by CPO will also cause a sharp increase in demand for SOI substrates, which are nearly a globally monopolized market.

Third is the external laser and its upstream supply chain. CPO creates a new product category: whereas traditional pluggable optical modules integrate the laser inside the module, under the CPO architecture, the laser must be separated and implemented as an external light source. This market barely existed before.

There is another critical mismatch in production capacity. Major laser manufacturers currently have their production capacity primarily dedicated to manufacturing EML traditional lasers, which integrate light emission and modulation onto a single chip for use in pluggable optical modules; orders and contracts have already been secured through 2027 and 2028. However, CPO requires simpler lasers that only emit light and do not perform modulation, as the modulation function is handled by the silicon photonic PIC within the package. Although both types of lasers use InP, their designs and production lines differ significantly, making seamless switching impossible. The large manufacturers’ capacities are locked into traditional laser contracts, forcing even Lumentum itself to source CPO-compatible lasers on the open market—excess demand will flow to independent laser suppliers.

The surge in demand for lasers will continue to ripple upstream. More lasers mean greater demand for InP substrates and epitaxial wafers. Goldman Sachs' report warns that the supply shortage of InP substrates could persist until 2027.

Fourth is packaging and assembly. CPO is essentially a packaging challenge, requiring precise integration of silicon photonic PICs and electronic chips with extremely high accuracy. Manufacturers capable of achieving CPO-level packaging and assembly will be scarce in the future.

Fifth is testing and inspection. Each silicon photonic PIC must undergo optical performance testing and reliability verification before leaving the factory. CPO testing is more complex than traditional optical modules because it involves mixed optical and electronic validation, and this phase will grow rapidly as CPO volumes increase.

In summary, after the surge in CPO demand, the bottleneck areas that benefit the most are silicon photonics foundry services, silicon photonics substrates, external lasers, InP substrates and epitaxy, packaging and assembly, and testing and inspection.

Upstream substrates: AXTI and Soitec

From upstream to downstream, the two most important companies in the substrate layer are AXTI and Soitec. The two companies serve different technological pathways and are not competitors, but rather complementary partners. AXTI supports the laser industry chain by generating light, while Soitec supports the silicon photonics industry chain by controlling light. Optical interconnects require collaboration between both.

AXTI is a U.S.-based company specializing in InP and GaAs substrates. Its process involves purifying, synthesizing, and growing single-crystal ingots from rare elements such as indium, phosphorus, gallium, and arsenic, which are then sliced into thin wafers. AXTI’s competitive advantage lies in the fact that only a handful of companies worldwide can produce high-quality InP substrates—besides AXTI, these include Sumitomo Electric Industries in Japan and Freiberger in Germany. AXTI’s moat stems from its decades of accumulated expertise in material purity processes and lengthy customer qualification cycles. For downstream customers, switching suppliers requires revalidating entire product lines, resulting in very high switching costs.

CPO will not bypass the InP substrate; instead, it will amplify demand. Under the CPO architecture, each GPU requires an external laser, and the number of lasers is directly tied to the number of GPUs. More lasers mean more InP substrates. Therefore, CPO is clearly beneficial to AXTI. AXTI’s investment profile features a small market cap and high volatility, with demand transmission experiencing a lag—but once this translates into orders, the stock’s price elasticity could be substantial.

Soitec is a publicly listed company based in Paris, France, specializing in SOI silicon photonics substrates. Soitec holds a dominant market position in the field of SOI substrates dedicated to silicon photonics and invented the patented Smart Cut technology for manufacturing SOI wafers. The core of CPO is the silicon photonics PIC, and each silicon photonics PIC requires an SOI substrate, making Soitec one of the most certain beneficiaries of the CPO supercycle. At the time, its valuation was approximately 1.4 times book value, which was low for a global monopolist. Note that Soitec is listed on the Paris Stock Exchange, not on U.S. markets.

Epilayer: IQE/IQEE

Below is the epitaxy layer. The globally significant independent epitaxy supplier is IQE/IQEE, which is listed in London. IQE’s moat lies in the extreme difficulty of epitaxy itself. Epitaxy involves growing functional layers like a layered cake on a substrate, with each layer only a few nanometers thick—any minor deviation in material, temperature, or growth time can render a laser defective. These parameter combinations constitute the epitaxy recipe, and IQE has accumulated decades of expertise in these recipes, which cannot be replicated in the short term simply by spending money.

After the CPO surge, IQE follows a logic similar to AXTI: CPO drives increased demand for lasers, and more lasers require more epitaxial wafers. IQE’s risk lies in its high customer concentration, with LITE being one of its key clients. If LITE decides in the future to produce epitaxial wafers in-house and pursue vertical integration, IQE’s largest revenue source could be adversely affected—this is a single-point risk that must be carefully considered before investing.

Laser Layer: SIVE/SIVEE, LITE, COHR, AAOI

Moving deeper into the chip layer, the most scarce component is the laser. Key companies include SIVE/SIVEE, LITE, COHR, and AAOI.

SIVE/SIVEE is one of the top-performing optical interconnect stocks over the past year. It is a small Swedish-listed company with a market capitalization of approximately $1.5 billion and annual revenue of around $30 million. It follows a fabless model, possessing its own InP100 platform and a small wafer fab in Glasgow, UK, giving it some manufacturing capability. It also collaborates with Taiwan’s Win Semi, outsourcing laser design to established foundry capacity to scale up high-power laser production.

SIVE/SIVEE has five core advantages. First, the InP100 standardized platform enables laser core modules to be standardized, allowing different product specifications to be rapidly assembled like building blocks. Second, wafer-level testing eliminates the need to dice chips before individual testing; instead, each chip is tested directly on the wafer, improving yield and reducing costs. Third, it covers both current and next-generation technologies, offering products for pluggable optical modules and CPO external light sources. Fourth, it pursues multiple parallel markets—beyond AI data center optical interconnects, it also serves LiDAR, satellite communications, and defense, diversifying risk from a single market. Fifth, it adopts a lightweight expansion model: a small facility handles core validation and low-volume production, while large-scale manufacturing leverages Win Semi’s capacity, avoiding heavy capital investment in factories while retaining core manufacturing capabilities.

SIVE/SIVEE is a highly elastic asset within the CPO supercycle. One reason is that large manufacturers' production capacity is locked in by traditional laser orders, leaving external light source demand from CPOs to be absorbed by independent laser suppliers. Another reason is that it has been integrated into the supply chains of multiple CPO projects. AMD’s CPO solution is driven through the GlobalFoundries platform, and SIVE is one of the few laser suppliers within its ecosystem; customers also include Celestial AI (a silicon photonics interconnect startup under Marvell) and Ayar Labs (a CPO/silicon photonics interconnect startup).

However, the risks of SIVE/SIVEE are also clear: revenue is too low, and most customers are still in the development and validation stages, not yet entering full-scale mass production. If just two or three customers finalize their orders, the stock price could continue to rise; if customers delay or cancel, the stock price could also drop significantly. Think of it as a high-payout lottery ticket.

LITE, or Lumentum, is a representative of the laser IDM model. It handles laser design, manufacturing, and full optical module assembly. LITE’s most critical highlights are NVIDIA’s $2 billion strategic investment and multi-billion-dollar procurement commitments, which directly secure its production capacity. Additionally, LITE is deeply integrated with Google’s TPU (Google’s proprietary AI accelerator ecosystem), with Google’s AI data centers extensively utilizing LITE’s optical switching technologies and lasers.

The LITE CEO made three key statements during the earnings call: CPO will experience a significant supply-demand imbalance; CPO is Lumentum’s largest single growth driver; and CPO is still in a very early stage. This amounts to direct confirmation from a top industry CEO that a CPO supercycle is underway. LITE’s production capacity is already booked through 2028, with its competitive moat built on dual major customer commitments from NVIDIA and Google. The risk lies in the fact that NVIDIA’s lock-in of capacity also caps short-term upside—revenue is largely dependent on NVIDIA’s orders, leaving the company with limited autonomy and a growth trajectory less steep than that of SIVE/SIVEE.

COHR, or Coherent, is a highly rare full-stack player in the optical interconnect space, capable of covering the entire value chain—from materials and InP lasers to silicon photonics PICs and optical modules. Its market share in optical modules ranks among the top globally, at approximately 20%. Like LITE, COHR has received a $2 billion strategic investment from NVIDIA and multi-billion-dollar procurement commitments.

COHR’s advantage lies in its ability to avoid being left behind regardless of how technological pathways evolve. It can produce silicon photonics PICs required for CPO, it can manufacture lasers, and it can also continue producing pluggable optical modules. This is the value of full-stack coverage. COHR is more like a mid-cap optical interconnect stock with higher safety, offering high certainty, though with less upside potential than SIVE/SIVEE, while exhibiting lower volatility and reduced risk.

AAOI is one of the few vertically integrated optical interconnect companies based in the United States. It uses MBE (Molecular Beam Epitaxy) equipment to grow epitaxial layers on InP substrates, manufactures its own laser chips, packages optical subassemblies, and assembles complete optical modules. Its current core business focuses on 800G and 1.6T pluggable optical modules. According to the transcript, AAOI secured its first large-volume order for 1.6T data center optical modules in March, with an initial order value exceeding $200 million, followed by an additional $71 million order for 800G modules in April.

AAOI is not necessarily at risk from CPO. First, pluggable optical modules will not disappear due to the rise of CPO; CPO addresses internal connections within super nodes, while large-scale connections between cabinets still require pluggable optical modules. Second, AAOI is entering the CPO supply chain. Under the CPO architecture, lasers cannot be placed inside the package and must instead be externalized into a small module that delivers light via fiber optics. AAOI’s newly introduced product is precisely such an external laser source designed for CPO. Overall, AAOI’s advantages include vertical integration, supply chain security through U.S.-based manufacturing, and the expansion potential of its laser technology into external light sources for CPO. However, it is also a small-cap, high-beta stock with high volatility, high upside potential, and elevated risk.

Contract Manufacturers: Win Semi and TSEM

After discussing lasers, let’s look at the foundries. The two most critical companies are Win Semi and TSEM.

Win Semi is one of the world’s largest compound semiconductor foundries, offering GaAs and InP foundry services. Mass production of SIVE/SIVEE lasers is primarily carried out through Win Semi. As next-generation CPO architectures drive increased demand for external lasers, Win Semi has become the most critical foundry partner for these laser design companies. Regardless of which laser design company ultimately succeeds, they will most likely rely on Win Semi for manufacturing.

TSEM is an Israeli specialty foundry, often referred to as the TSMC of optical interconnects. It may be one of the companies that stands to benefit most directly from the CPO supercycle. At the core of CPO is the silicon photonic PIC, and TSEM is the foundry with the largest market share in silicon photonic PIC manufacturing. The mandatory use of silicon photonic PICs in CPO effectively elevates TSEM’s silicon photonics foundry business from a niche segment to the center of the supply chain.

Most of TSEM's production capacity has been booked through 2028, yet its expected P/E ratio remains only at 16 to 18 times, leaving upside potential given the strong growth outlook for CPO. The primary risk is geopolitical, as it is an Israeli company based in the Middle East and could be affected by regional conflicts.

Win Semi and TSEM are both foundries, but their core difference lies in the materials used and the components they manufacture. Win Semi uses InP and GaAs to produce lasers, responsible for light generation, while TSEM uses SOI substrates to manufacture silicon photonic PICs, responsible for light manipulation. The two material systems are incompatible, and they are not competitors, but rather foundries operating at different stages of the supply chain.

DSP and Switch Chip Layer: Broadcom and Marvell

Below that are the DSP and switching chip layers, primarily consisting of Broadcom and Marvell.

Broadcom (AVGO) is a trillion-dollar U.S. tech giant whose business spans switching chips, custom AI acceleration chips, and enterprise software. Two of its operations are directly related to optical interconnects. First, DSP chips—the "brain" of optical modules responsible for error correction coding; Broadcom is one of the most important suppliers in this field. Second, CPO switches: Broadcom’s third-generation CPO switch has entered mass production, integrating the optical engine directly alongside the switching chip. In terms of CPO commercialization, Broadcom is even ahead of NVIDIA.

However, from an investment perspective, optical interconnect is just one of Broadcom’s many business segments and represents a relatively small portion of its overall revenue. Its stock price will not multiply severalfold simply due to a surge in CPO adoption. Investing in Broadcom means betting on the comprehensive certainty of AI infrastructure, not just the isolated upside potential of the optical interconnect industry.

MRVL, or Marvell Technology, is another diversified semiconductor company involved in custom AI acceleration chips, data center networking chips, and storage chips. Two areas directly related to optical interconnects are: first, DSP chips, where Marvell and Broadcom are the two leading suppliers and direct competitors; second, CPO. Marvell’s acquisition of Celestial AI has significantly strengthened its capabilities in silicon photonics interconnects.

The core logic of this update is replacing copper cables with optical connections between GPUs. Celestial AI is also pursuing this direction, but at a shorter distance: replacing copper with optical connections inside chip packages. Through this acquisition, Marvell has significantly strengthened its strategic position in CPO.

Compared to Broadcom, Marvell has a more concentrated exposure to optical interconnects. Broadcom is a trillion-dollar company where optical interconnects represent just one segment; Marvell is smaller, with revenue of $8.2 billion in the last fiscal year, up 42% year-over-year, and management expects revenue to approach $15 billion over the next two fiscal years. Optical interconnects and CPO account for a larger and more elastic share of Marvell’s overall revenue. While Marvell is not a pure-play in optical interconnects, it may be one of the best overall choices with balanced exposure to both DSP and CPO.

Underlying fiber: Corning

Finally, there is the underlying company, Corning (GLW). Corning is the global leader in optical fiber. Many people are familiar with Corning because of its glass used in Apple iPhone screens; however, optical communications has become one of Corning’s largest and fastest-growing divisions. Since inventing communications fiber in 1970, Corning has deployed millions of miles of fiber optic cable.

Regardless of which optical module company wins or whether the technology path is pluggable or CPO, Corning’s fiber is required. Under the CPO architecture, fiber still connects the laser to the silicon photonic PIC, and fiber continues to be used between different cabinets. Fiber is one of the few components in the entire supply chain unaffected by the technology debate.

Corning has recently strengthened its customer relationships. In January this year, Meta announced it would invest up to $6 billion to help Corning expand its fiber optic cable manufacturing facilities; NVIDIA also announced a multi-year partnership with Corning, investing $500 million to acquire warrants in Corning. Corning has committed to increasing its optical connectivity capacity in the U.S. tenfold, boosting fiber production by more than 50%, and building three new factories.

NVIDIA previously invested $2 billion each in LITE and COHR, and now has invested $500 million in Corning, demonstrating its strategy to extend the AI infrastructure competition from chips to fiber optics, systematically securing the entire optical interconnect supply chain. Corning is the most certain, yet least elastic, player across the entire optical interconnect value chain.

Three configuration strategies: conservative, balanced, aggressive

After discussing so many companies, the final question is: “How to invest?” The most important principle is: the further upstream you go, the smaller the companies, the higher their potential upside, but the lower their certainty; the further downstream you go, the larger the companies, the higher their certainty, but the lower their potential upside. Upstream companies like AXTI and IQE, which produce substrates and epitaxial wafers, have small market caps and delayed demand transmission, but once demand surges, their upside potential can be substantial. In contrast, large downstream companies like AVGO offer high certainty, but it’s unrealistic to expect them to quintuple in value within a year.

The first portfolio is a conservative allocation, with core holdings in AVGO, MRVL, and GLW. All three companies are large-cap firms, with Broadcom’s market capitalization now reaching approximately $2 trillion, placing it among the top ten U.S. stocks; Marvell and Corning are also in the hundreds of billions of dollars range. Broadcom and Marvell have diversified business models, with optical interconnect being just one component; Corning is more focused, but optical fiber remains a fundamental, technology-agnostic necessity. This portfolio offers limited downside risk—even if optical interconnect development falls short, other business segments can support stock performance, making it ideal for long-term investors seeking to avoid significant volatility.

The second portfolio is a balanced allocation, centered on COHR, LITE, and TSEM. Each company is a market leader in its respective segment, with mid-sized scale, offering both stability and upside potential. COHR is a full-stack optical company that remains well-positioned regardless of industry direction, with NVIDIA’s $2 billion investment providing a margin of safety. LITE is a core supplier of lasers with capacity locked in by NVIDIA, and its CEO has personally confirmed a supply-demand imbalance in CPO. TSEM is the largest foundry in silicon photonics PIC manufacturing, with relatively attractive valuation. This combination is well-suited for those looking to gain exposure to optical interconnects while accepting some volatility.

The third portfolio is an aggressive allocation, consisting of SIVE/SIVEE, AAOI, SOI/Soitec, AXTI, and IQE. All five companies operate at critical upstream bottlenecks in the supply chain. SIVE/SIVEE is a scarce supplier of external light sources for CPO and is already integrated into multiple CPO supply chains; AAOI is a high-beta player in pluggable optical modules with the capability to enter the CPO external light source market; Soitec holds a dominant market position in silicon photonics substrates; AXTI provides InP substrates essential for laser manufacturing; and IQE produces the critical epitaxial wafers used in laser fabrication. If the CPO supercycle unfolds at the pace predicted by Goldman Sachs, this portfolio offers the highest upside potential—but also carries the greatest risk.

It’s normal for these low-market-cap assets to drop 20% to 30% in a single day; it’s best to keep your position size within 5% to 10% of your total investment portfolio. Also note that many photonic interconnect low-market-cap assets are not listed on U.S. exchanges. Soitec is listed on the Paris Stock Exchange, IQE on the London Stock Exchange, SIVE in Sweden, and Win Semi in Taiwan. Most of these can be traded through Interactive Brokers, but you’ll need to enable access to the respective markets.

Track risk: CPO progress, NVIDIA's selection, small-cap volatility

The entire sector also carries significant investment risks.

First, the commercialization timeline for CPO is uncertain. Goldman Sachs’ prediction of a $91 billion CPO market is a highly aggressive estimate. Achieving this figure requires NVIDIA’s next-generation architecture to launch on schedule, CPO yield targets to be met, InP substrate supply to keep pace, cloud providers to maintain high capital expenditures, and continuous inflows of capital into the supply chain. Any bottleneck in these areas would significantly reduce the actual market size.

Second, NVIDIA’s choice is critical. The optical interconnect solution adopted by NVIDIA’s next-generation Rubin platform will directly impact the entire supply chain landscape. NVIDIA has already included CPO in the Rubin reference architecture, but the final supplier selection and production timeline remain uncertain.

Third, small-cap assets carry inherent risks. Many companies in the optical interconnect industry have small market capitalizations; therefore, you should not heavily weight these assets in your portfolio, and you should never use leverage on them.

Three Core Judgments and Conclusion

Finally, here are my three key assessments of the optical interconnect sector.

First, optical interconnects are not a conceptual hype. The interconnection demands of AI data centers are real, urgent, and irreversible. The more GPUs sold, the greater the demand for optical interconnects—this is a highly certain sector tightly bound to the GPU industry chain.

Second, CPO represents the largest future growth driver in this space. Goldman Sachs forecasts that the optical interconnect market could grow ninefold, with CPO accounting for $91 billion; Lumentum’s CEO has personally confirmed that CPO supply and demand are severely imbalanced and still in the early stages; NVIDIA has already incorporated CPO into its next-generation architecture, demonstrating that this is not a distant prospect but something happening now.

Third, if you can tolerate high risk and high volatility and are seeking high returns, the core strategy is to identify bottlenecks. Unlike the GPU industry, which is dominated by NVIDIA, the optical interconnect产业链 is characterized by extremely fine specialization and widely dispersed bottlenecks. Behind each bottleneck, there are often only one or two companies capable of addressing it. Finding these bottlenecks means finding the largest alpha in this sector.

In summary, GPUs are the brain of AI, but it’s the neural network connecting these brains that determines how fast the entire system can run. Optical interconnects are AI’s neural network—without them, no matter how many GPUs you have, they remain isolated islands. This产业链, long overshadowed by the glow of GPUs but potentially worth trillions in the future, may be brewing the next major investment opportunity.

Of course, the volatility and risks associated with the optical interconnect sector can also be substantial; the above content does not constitute investment advice. Before investing, be sure to carefully consider the potential returns and risks, and make decisions based on your actual position and cash flow.