عنق الزجاجة القادم للذكاء الاصطناعي: دراسة معمقة حول الاتصال البصري

In-Depth Research Analysis:

Executive Summary:

Optical connectivity is evolving from a supporting component of traditional communications equipment into a core bandwidth layer of AI infrastructure. As AI clusters shift from competition in standalone chip performance toward large-scale system coordination, network bandwidth, latency, power consumption, and connectivity efficiency increasingly determine GPU utilization, training time, and total compute cost. In simple terms, GPUs define the theoretical ceiling of computing power, while optical connectivity determines how much of that capacity can actually be used.

This report begins by explaining the role of optical connectivity inside AI data centers and follows the data path from GPUs and network interface cards to switches, optical transceivers, and fiber. It then examines value-chain economics, competitive positioning, and technology roadmaps. The clearest near-term earnings drivers remain 800G and 1.6T pluggable products, while switching silicon, DSPs, lasers, silicon photonics, optical engines, and test equipment may capture greater value as the industry moves to higher speeds. Co-packaged optics represents an important long-term direction, but commercial adoption still requires further validation.

From an investment perspective, optical connectivity is not a theme in which every company benefits equally. The most strategically attractive companies typically control at least one scarce capability: network architecture, critical silicon, key light sources, hyperscaler access, or difficult-to-replicate mass production. For the second half, core companies such as Broadcom, Arista, and NVIDIA, which combine stronger earnings visibility with greater system control. Marvell, Coherent, and Lumentum offer higher-beta exposure to 1.6T, silicon photonics, DSPs, and lasers, while Cisco, Ciena, Corning, and Keysight provide complementary exposure to systems, DCI, fiber, and testing.

The report’s central conclusion is that the next phase of AI infrastructure competition is moving from “owning more GPUs” toward “connecting more GPUs with lower power and greater efficiency.” As networking captures a larger share of AI-system value, profit pools are likely to concentrate among a limited group of companies controlling technical standards, critical components, and customer access.

1 Why Optical Connectivity Is Becoming a Core AI Infrastructure Bottleneck

Over the past two years, AI infrastructure investment has focused primarily on GPUs, advanced semiconductor manufacturing, and high-bandwidth memory. As AI clusters continue to scale, however, the system’s main constraint is beginning to shift. Individual chip performance remains critical, but the ability to connect thousands of accelerators efficiently is becoming equally important. Optical connectivity is therefore evolving from a supporting component of traditional communications equipment into essential bandwidth infrastructure for AI data centers.

The starting point is the distinction between theoretical compute and effective compute. GPUs determine how much processing a single node can perform, but large AI models do not run on one accelerator. They require thousands of GPUs to operate in parallel, continuously exchanging model parameters, gradients, and intermediate data. If network bandwidth is insufficient, latency is too high, or connections are unstable, GPUs spend more time waiting for data and less time computing. The efficiency of an AI cluster therefore depends not only on how many GPUs it contains, but also on how effectively those GPUs are connected.

This is why AI is gradually turning a computing problem into a communications problem. Traditional cloud traffic was largely driven by data moving between users and servers. In AI training clusters, the dominant traffic increasingly flows between servers, switches, and accelerators. As clusters expand from thousands to tens of thousands of GPUs and beyond, communication demand may grow faster than the number of chips themselves. Networking is no longer a secondary component; it directly influences training time, compute utilization, and the cost of running increasingly complex models.

Optical connectivity becomes more important as bandwidth, distance, and port density increase. Copper remains competitive in short-distance, cost-sensitive applications, but electrical connections face rising signal loss, power consumption, and heat at higher speeds. Optical links can transmit more data over longer distances with lower energy consumption per bit, making them increasingly necessary for rack-to-rack connections, data center fabrics, and campus-level interconnects. As a result, industry demand is shifting away from traditional telecom capital expenditure toward cloud and AI infrastructure spending.

The current optical communications cycle is not simply a volume-growth story. The deeper change is that optical connectivity is capturing a larger share of AI system value. The transition from 400G to 800G and 1.6T requires more advanced lasers, DSPs, photonic components, packaging, testing, and thermal management. At the same time, silicon photonics, co-packaged optics, and higher-speed switching architectures are moving the industry from standalone modules toward tighter system-level integration. This shift is likely to concentrate profits among companies that control high-speed chips, optical components, network architecture, and reliable mass production.

The investment thesis rests on three factors. First, continued AI capital expenditure creates structural demand for high-speed interconnects. Second, larger AI clusters increase the importance and value of networking within the overall system. Third, high-end chips, lasers, packaging, testing, and customer qualification create substantial barriers to entry, limiting supply elasticity and strengthening the position of leading companies.

The central conclusion of this report is that optical connectivity is not merely an extension of the GPU trade. It is the natural consequence of AI infrastructure moving from a competition in individual chip performance to a competition in system efficiency. GPUs determine the theoretical ceiling of computing power; optical connectivity determines how much of that computing power can actually be used. The key industry question is therefore not only how many optical modules will be required, but which companies will control the next bottlenecks in bandwidth, power efficiency, and connectivity.

2 How Optical Connectivity Works — Following Data from One GPU to Another

The most effective way to understand optical communications is not to begin with a long list of technical terms, but to follow how data moves through an AI cluster. During AI training and inference, information must continuously travel among GPUs, servers, switches, and data centers. Optical connectivity provides the links that allow these computing nodes to exchange increasingly large volumes of data within strict limits on time, power, and physical space.

A simplified data path looks like this:

GPU or Server → Network Interface Card → Switch → Optical Transceiver → Fiber → Optical Transceiver → Switch → Another Server or GPU

GPUs perform the computation. Switches determine where data should go. Optical transceivers convert electrical signals into optical signals and back again. Fiber carries those optical signals across the network. Most products and technologies in the optical communications industry are built around this transmission path.

2.1. Optical Transceivers: Converting Electricity into Light

Servers and networking chips process electrical signals, but electrical signals become increasingly inefficient over longer distances and at higher speeds. An optical transceiver converts the electrical output of a server or switch into an optical signal, sends it through fiber, and converts it back into an electrical signal at the receiving end.

A transceiver is typically inserted into a server or switch port. Inside the module are lasers, photonic components, detectors, driver chips, amplifiers, DSPs, packaging, and thermal-management structures. Although the external form factor is standardized, the internal technology becomes substantially more complex as transmission speeds increase.

Terms such as 400G, 800G, and 1.6T refer to the total amount of data a module can theoretically transmit each second. An 800G module supports 800 gigabits per second, while a 1.6T module supports 1.6 terabits per second. Higher speeds increase the volume of data that can move through each port, but they also require more advanced chips, lasers, signal processing, packaging, power management, and cooling.

2.2. Why Convert Signals from Electrical to Optical and Back?

If servers operate electronically, why not connect every device with copper cables?

Copper has clear advantages. It is relatively inexpensive, simple, and efficient over very short distances, including connections inside servers and racks. However, as speed and distance increase, copper faces rapidly rising signal loss, interference, power consumption, and heat.

Optical links are better suited to high-bandwidth and longer-distance transmission. They offer lower signal loss, immunity to electromagnetic interference, and lower energy consumption per transmitted bit. The industry is therefore not replacing copper in every application. Instead, copper and optics serve different parts of the network:

Copper remains competitive close to chips and across very short distances.

Optical links become increasingly important between racks and throughout data centers.

For campus and inter-data-center connections, optics are generally indispensable.

Copper and optical connectivity are therefore complementary rather than entirely competing technologies. As AI clusters become larger and faster, the distance and number of connections where optics provide the better solution generally increase.

2.3. Lanes, PAM4, and DSPs: How High-Speed Transmission Is Achieved

An 800G transceiver usually does not carry the full 800G over a single channel. Instead, the total bandwidth is divided across several high-speed channels, known as lanes. An 800G module may use eight 100G lanes, while more advanced designs may use four 200G lanes.

Increasing the speed per lane can reduce the number of optical components and help improve density. However, it also raises design complexity, signal-integrity requirements, and manufacturing difficulty. The industry’s migration from 100G per lane toward 200G per lane is an important foundation for 1.6T products and a major area of technological competition.

PAM4 is a modulation format that increases the amount of information carried by each signal. Instead of using two signal levels, PAM4 uses four, allowing each transmitted symbol to carry more data. The trade-off is that the signal becomes more sensitive to noise and distortion.

A digital signal processor, or DSP, compensates for transmission errors, noise, and signal degradation. As speeds rise, DSPs become increasingly important, but they also add cost and consume power. Broadcom and Marvell hold important positions in high-speed DSPs and networking silicon because these chips directly affect whether high-speed optical links can operate reliably.

2.4. Lasers, Photonic Chips, and Detectors

Lasers generate the optical signal. Modulators encode data onto that light, while detectors at the receiving end convert the optical signal back into an electrical one. Different speeds, distances, and power requirements require different laser technologies and optical architectures.

Common technologies include VCSELs, EMLs, continuous-wave lasers, and silicon photonics. VCSELs are generally suited to shorter distances and cost-sensitive applications. EMLs offer strong high-speed and longer-distance performance. Continuous-wave lasers can serve as external light sources for silicon-photonics and co-packaged-optics systems. Silicon photonics uses semiconductor manufacturing techniques to integrate more optical functions onto silicon-based chips.

These technologies do not follow a simple winner-takes-all replacement path. Each architecture balances cost, distance, power consumption, manufacturing yield, and integration. The industry’s technical progress is fundamentally a search for higher bandwidth with lower power, greater density, and better reliability.

2.5. Switches and Network Architecture: Where Optical Demand Begins

Optical transceivers do not determine where data travels. That role belongs to switch chips, networking systems, and network operating software. Switches connect large numbers of servers and GPUs and direct traffic to the appropriate destination.

In traditional data centers, networks primarily carried general computing and storage traffic. AI clusters require thousands of accelerators to exchange information simultaneously, creating much stricter requirements for throughput, latency, congestion control, and reliability. An inadequate network forces GPUs to wait for data, reducing cluster utilization.

Optical demand therefore depends not only on the number of GPUs deployed, but also on the architecture used to connect them. Each expansion in accelerator capacity usually requires additional network interface cards, switch ports, transceivers, and fiber. More network layers, greater redundancy, and larger cluster sizes increase the required number of optical links.

This is why companies such as Arista Networks, Broadcom, NVIDIA, and Cisco are central to an optical-connectivity investment framework. They may not manufacture every optical module, but their switch chips, network systems, and architectures determine required speeds, port counts, and connectivity standards.

2.6. Inside the Data Center and Between Data Centers

Optical demand can be divided into connections inside data centers and connections between data centers.

Intra-data-center links connect servers, racks, and switches across distances ranging from a few meters to several hundred meters. This market is highly sensitive to cost, power consumption, port density, and large-scale manufacturing. It is the primary market for 800G and 1.6T high-speed data-center transceivers.

Data-center interconnect, or DCI, connects separate buildings, campuses, or regional facilities. These longer-distance links often require coherent optics, wavelength-division multiplexing, and optical transport systems to carry more data over each fiber. As AI infrastructure expands from individual data centers into multi-building campuses and distributed clusters, DCI demand is also increasing.

These two markets have different technical requirements and competitive structures. Intra-data-center connectivity emphasizes high-speed pluggable modules, Ethernet switching, and power efficiency. DCI emphasizes transmission distance, spectral efficiency, coherent DSPs, and network management. Investors should distinguish which market each company primarily serves rather than treating all optical revenue as economically identical.

Section Conclusion

Optical connectivity is not a standalone hardware product. It is a complete system linking GPUs, servers, switches, and data centers. Transceivers perform electrical-to-optical conversion, lasers and photonic chips generate and manipulate light, DSPs correct errors in high-speed transmission, switches direct traffic, and fiber provides the physical transmission medium.

Once this data path is understood, the industry can be analyzed through three central questions: how much data the network must carry, how technology upgrades change the value of each connection, and which companies control the most difficult-to-replicate chips, optical components, and system capabilities.

3 The Value Chain and Profit Pools — Who Actually Captures the Economics?

Understanding optical connectivity as an investment theme requires more than identifying which companies sell the most transceivers. The more important question is who controls technical standards, critical chips, customer access, and reliable mass production. Although optical products are ultimately delivered as transceivers, switches, or transport systems, profits are not distributed evenly across the value chain.

In general, companies closer to core silicon, system architecture, and customer standards tend to enjoy stronger barriers and pricing power. Companies concentrated in standardized assembly and manufacturing are more exposed to price competition and product-cost declines.

For investment purposes, the optical-connectivity value chain can be divided into five layers: electrical and photonic chips, optical modules and engines, network equipment and system platforms, fiber and connectivity infrastructure, and the cloud customers that ultimately fund the industry.

3.1. DSPs and Switching Silicon: The Intelligence Behind the Network

Optical transceivers contain not only lasers and photonic components, but also DSPs, drivers, transimpedance amplifiers, and clock-recovery chips. These components process high-speed signals, compensate for distortion, reduce error rates, and ensure interoperability across different systems.

At the network level, switch chips determine port counts, bandwidth, and congestion-management capability. Their importance rises as AI clusters scale because network performance directly affects GPU utilization.

This layer generally has strong pricing power because chip development is difficult, product cycles are long, performance and power requirements are demanding, qualification periods are lengthy, and the number of credible suppliers is limited.

Broadcom and Marvell are major listed representatives. Broadcom spans switching silicon, custom AI accelerators, optical DSPs, and co-packaged-optics platforms, giving it a broad position across AI networking. Marvell focuses on high-speed electro-optical connectivity, optical DSPs, PAM4, and custom compute.

Their value is not measured by how many transceivers they manufacture. Their silicon defines what performance the entire network can achieve. As a result, DSPs and switching chips are among the most technically protected and economically attractive areas of the optical value chain.

3.2. Lasers, Photonic Chips, and Silicon Photonics

Lasers and photonic chips generate, modulate, and receive optical signals. Their performance affects transmission distance, speed, power consumption, reliability, and manufacturing yield.

Traditional modules use separately manufactured optical components that are later assembled into a package. As speeds move from 400G to 800G and 1.6T, this approach faces rising component counts, packaging complexity, and power consumption. Silicon photonics seeks to integrate more optical functions onto a silicon platform, potentially improving density and long-term manufacturing economics while supporting architectures such as co-packaged optics.

Coherent and Lumentum are important listed suppliers in this segment. Coherent offers a broad portfolio spanning lasers, photonic components, materials, silicon photonics, and optical engines. Lumentum has significant exposure to high-speed lasers, EMLs, and cloud-data-center optics.

Their competitiveness depends on the ability to deliver next-generation products, maintain yields, pass hyperscaler qualification, and expand capacity reliably.

The investment thesis has two components. First, faster product generations increase the value of advanced lasers and photonic components. Second, long qualification cycles and difficult manufacturing limit supply elasticity. If 1.6T and silicon-photonics adoption accelerate, suppliers controlling critical light sources and photonic integration may gain stronger pricing power.

3.3. Optical Modules and Engines: Direct Demand Exposure, but Intense Competition

Optical modules provide the most visible exposure to AI-network demand. More AI servers and switches require more high-speed ports, and the transition from 400G to 800G and 1.6T raises both product complexity and selling prices.

However, the module layer is not automatically the most profitable. External interfaces and some product specifications are standardized, major customers possess substantial purchasing leverage, and prices tend to decline as products mature and volumes increase.

Long-term competitiveness depends on whether a supplier can:

Enter leading hyperscaler or equipment-vendor supply chains;
Complete next-generation product qualification quickly;
Maintain yields, power efficiency, and reliability;
Deliver at large scale with competitive costs;
Secure critical upstream lasers, DSPs, and components.

Chinese manufacturers play an important global role in high-speed-module production, packaging, cost control, and large-scale delivery. Because this report focuses on US-listed equities, they are discussed primarily as part of the competitive backdrop rather than as individual investment cases.

For US equity investors, the more relevant question is whether listed companies control high-end optical components, silicon, network platforms, or customer standards rather than simply comparing module shipment volumes.

The module business therefore combines strong demand elasticity with rapid product cycles, customer concentration, and pricing pressure. Investors should monitor product mix, average selling prices, gross margins, customer exposure, and the contribution of next-generation products—not shipment growth alone.

3.4. Network Equipment and System Platforms

Optical modules only create value when installed in switches, routers, or transport systems. Network-equipment companies therefore possess important system-definition power. They determine port speeds, network topologies, protocols, interoperability requirements, and module specifications.

Arista Networks, Cisco, NVIDIA, and Broadcom are important participants in AI-network architecture.

Arista’s strengths lie in high-performance Ethernet switches, network operating software, and deep relationships with large cloud customers. As AI clusters increasingly adopt faster Ethernet fabrics, Arista may benefit from both higher port counts and greater value per port.

Cisco combines a broad enterprise and service-provider customer base with switching silicon, coherent optics, and DCI capabilities through Silicon One and Acacia. Its advantage is system breadth and integration, although it faces competition from more specialized cloud architectures.

NVIDIA combines InfiniBand, Spectrum-X, network interface cards, switches, and full-system designs. Its objective is not merely to sell networking hardware, but to optimize compute, networking, and software as one AI infrastructure platform.

System-platform companies may capture more value than component suppliers because they influence how entire networks are designed. The companies that define AI-network architecture can shape future demand for switch silicon, optical modules, and connectivity technologies.

3.5. Fiber, Connectors, and Test Equipment

High-speed optical networks also depend on fiber, connectors, cabling systems, and test equipment.

Corning is a major supplier of fiber, cable, and data-center connectivity systems. As AI campuses expand, fiber counts, connection density, and cabling complexity all increase. Fiber technology may evolve more slowly than high-speed silicon, but large-scale data-center construction creates substantial physical-connectivity demand.

Connectors, fiber arrays, and passive components may appear to carry less value, but they directly affect insertion loss, reliability, and serviceability. At higher speeds, small connection errors can undermine system stability, making qualification and manufacturing consistency increasingly important.

Test equipment is a prerequisite for mass production of 1.6T and future generations. Higher speeds require more demanding verification of error rates, signal integrity, optical performance, and consistency. Companies such as Keysight may benefit from the rising complexity of product development and manufacturing even though they do not sell optical modules directly.

Relative to general manufacturing, test equipment often offers stronger technical barriers and customer stickiness.

3.6. Hyperscalers: The Industry’s Real Pricing Center

The optical industry is ultimately funded by hyperscalers and large AI-infrastructure operators. Capital expenditure by Microsoft, Amazon, Alphabet, Meta, and major AI platforms determines demand for high-speed switches, optical modules, DCI systems, and related components.

These customers are not passive buyers. They increasingly participate in network-architecture design, develop custom silicon, define module specifications, influence protocols, and shape supplier selection. Hyperscalers are therefore both customers and major drivers of industry standards and technology roadmaps.

This has two important implications.

First, qualification by a leading cloud customer is highly valuable and can result in large, multi-year revenue opportunities.

Second, customer concentration creates meaningful risk. If one hyperscaler changes architecture, delays capital expenditure, or switches suppliers, revenue and valuation can move sharply.

Company analysis must therefore consider customer mix, order visibility, and positioning within next-generation cloud-network architectures—not only product specifications.

3.7 Section Conclusion

Profit pools in optical connectivity do not simply follow shipment volumes. High-speed modules provide direct exposure to AI demand, but the strongest long-term pricing power often belongs to companies controlling critical silicon, lasers, system architecture, test capability, or hyperscaler qualification.

From a US equity perspective, the value chain can be grouped into three broad categories:

Technology and standards controllers, including Broadcom, Marvell, NVIDIA, Arista, and Cisco;
Critical optical-component suppliers, including Coherent and Lumentum;
Infrastructure and supporting suppliers, including Corning, Ciena, and Keysight.

The central question is not whether a company belongs to the optical-communications industry. It is what scarce resource the company controls: bandwidth standards, critical silicon, light sources, system platforms, customer access, or merely a highly competitive manufacturing process. Different positions in the value chain lead to materially different growth quality, margins, and valuation frameworks.

4 Competition and Technology Roadmaps — From Pluggable Optics to Silicon Photonics and CPO

The optical-connectivity market contains many products and technical terms, but the competitive question is relatively straightforward: which companies can define next-generation products, secure customer qualification, and deliver them reliably at scale?

Competition is not simply about production capacity or the ability to demonstrate a prototype. It is a combination of technology, customer relationships, supply-chain control, and manufacturing execution.

The industry is transitioning from 400G toward 800G and 1.6T while exploring silicon photonics, LPO, NPO, and co-packaged optics. Near-term revenue remains driven by high-speed pluggable transceivers. The medium-term battleground is silicon-photonics adoption and the move toward 200G per lane. Over the longer term, optics may move closer to switching and computing silicon to address power, bandwidth-density, and thermal constraints.

4.1. Competitive Structure: Layered Control Rather Than a Single Leader

No company controls the entire optical-connectivity value chain. Different participants dominate silicon, light sources, modules, network systems, and customer access.

At the silicon and network-architecture layer, Broadcom, Marvell, NVIDIA, Arista, and Cisco exert significant influence. Through switching silicon, DSPs, custom AI accelerators, network software, and system design, they determine required port counts, speeds, and optical architectures. Long development cycles and high switching costs generally support strong pricing power.

At the optical-component and engine layer, companies such as Coherent and Lumentum control lasers, EMLs, silicon-photonics platforms, photonic chips, and packaging capabilities. Their competitiveness depends on next-generation product development, manufacturing yields, and supply availability. As speeds rise, optical complexity increases, raising the value of qualified leading suppliers.

Chinese suppliers have established important positions in high-speed module manufacturing, packaging, cost control, and large-scale delivery. Because this report focuses on US-listed equities, they are treated primarily as part of the global competitive backdrop rather than as individual investment cases.

At the systems layer, Arista, Cisco, NVIDIA, and Ciena are closer to end customers. Their network architectures and integrated platforms can shape optical specifications and supplier choices.

The industry therefore cannot be reduced to a simple conclusion that module companies will defeat chip companies, or that CPO will replace every pluggable product. Competition is fundamentally a contest for control over different layers of the value chain.

4.2. Industry Barriers: A Prototype Is Not a Commercial Product

One of the most underappreciated characteristics of optical communications is the distance between producing a laboratory prototype and generating large-scale revenue. Demonstrating an 800G or 1.6T product does not guarantee access to hyperscaler supply chains.

Sustained commercial success normally requires four capabilities.

First, next-generation product development. A supplier must keep pace with switch-silicon and hyperscaler architectures while meeting requirements for power, reach, error rates, and form factor.

Second, customer qualification. Cloud and equipment customers conduct extended interoperability, stability, and reliability testing. Successful qualification can create a durable supplier relationship, while missing one generation may cost several years of growth.

Third, reliable mass production. The challenge is not only product design, but also yield, consistency, and supply-chain management. As speeds increase, packaging and testing become more difficult.

Fourth, control of upstream bottlenecks. Companies with internal lasers, photonic chips, DSPs, silicon-photonics platforms, or secure supply arrangements are better positioned to maintain delivery and margins during tight supply.

Investors should therefore evaluate whether new capacity is tied to qualified products, customer orders, and acceptable yields—not simply whether capacity is being expanded.

4.3. Pluggable Transceivers: The Main Near-Term Earnings Driver

Pluggable transceivers remain the most mature and commercially visible architecture. Their advantages include standardized form factors, ease of maintenance, field replaceability, and compatibility with existing switch platforms.

800G is becoming an important product for high-speed AI-data-center interconnects, while 1.6T represents the next generation. The transition to 1.6T is more than a doubling of bandwidth. It typically requires faster lanes, more advanced DSPs, higher-performance lasers and photonic chips, and more demanding packaging, testing, and thermal design.

The investment thesis has three layers:

AI-cluster expansion increases the number of ports and modules;
The transition from 800G to 1.6T raises value per port;
Early-generation supply is limited and customer qualification is concentrated among a smaller number of suppliers.

The principal risk is price erosion. As products mature and industry capacity expands, module prices typically decline. If customers adopt multi-supplier strategies or capacity grows too rapidly, shipment growth may not translate into equivalent profit growth. Investors must therefore track average selling prices, cost reductions, gross margins, and the contribution of next-generation products.

4.4. Silicon Photonics: A More Practical Medium-Term Upgrade Than CPO

Silicon photonics is one of the industry’s most important technology directions. It uses semiconductor manufacturing techniques to integrate optical functions such as modulators, waveguides, and splitters onto silicon-based chips.

Its investment relevance does not depend on immediately replacing every traditional optical architecture. As bandwidth and port density rise, the industry needs greater optical integration. Silicon photonics can potentially reduce assembly complexity, improve density, and support architectures that place optics closer to switching silicon.

However, silicon photonics has limitations. Silicon is not an efficient light-emitting material, so systems generally still require external lasers or indium-phosphide light sources. Manufacturing also requires solutions for coupling, packaging, yield, and heat.

The likely winners may therefore be companies capable of integrating silicon photonics, lasers, packaging, and testing rather than firms that merely possess a silicon-photonics design.

From an investment-timing perspective, silicon photonics is closer to meaningful revenue than CPO. It can already be used in high-speed pluggable products and may gain share as 800G and 1.6T scale. The key question is whether it improves product cost, power, and manufacturing yield—not whether it generates attractive technology demonstrations.

4.5. LPO and Linear Optics: Lower Power with Less System Tolerance

Traditional high-speed transceivers rely on DSPs to correct signal degradation. DSPs improve reliability and interoperability, but they also add cost, power consumption, and latency. Linear pluggable optics, or LPO, seeks to reduce or remove DSPs within the module and shift more signal-processing responsibility to the switch silicon.

The main attraction is lower module power and latency, which are increasingly important in dense AI data centers. However, LPO places greater demands on system design, link quality, and component consistency. It also provides less tolerance for imperfect operating conditions.

LPO may therefore gain adoption first in short-reach, tightly controlled environments designed by a single hyperscaler, rather than across every general-purpose data center.

For investors, LPO could reduce the value of optical DSPs inside modules while increasing the strategic importance of switching silicon, system coordination, and high-quality analog and photonic components. It represents a potential migration of profit pools, not simply a lower-cost module.

4.6. Co-Packaged Optics: A Clear Long-Term Direction, but Not Yet a Near-Term Earnings Story

Co-packaged optics places optical engines near switching silicon or within the same package. As switch capacity and port speeds rise, the electrical connection between the chip and external transceiver consumes more power and creates greater signal loss and thermal complexity. CPO seeks to perform electrical-to-optical conversion closer to the chip, reducing electrical reach and improving bandwidth density.

The long-term logic is compelling. Larger AI clusters and faster switch silicon increase the value of bringing optics closer to the chip. CPO may become an important component of future high-capacity network platforms.

Commercial adoption, however, still faces significant challenges. Combining optical components with high-power switching silicon complicates thermal management and reliability. Pluggable modules can be individually replaced, while CPO may create more difficult maintenance economics. External light sources, packaging standards, and supply-chain coordination also require further development.

Our conclusion is that CPO represents an important medium- to long-term technology option, but near-term earnings remain driven primarily by 800G and 1.6T pluggable products and silicon-photonics adoption. Investors should focus on customer validation, system deployment, and revenue contribution rather than valuing every CPO-related company as though mass adoption were already assured.

4.7. How Technology Roadmaps Reshape Profit Pools

Different architectures will not benefit every company equally. They redistribute value across the industry.

The scaling of 800G and 1.6T supports high-speed modules, DSPs, lasers, photonic chips, packaging, and test equipment.

Greater silicon-photonics adoption benefits companies with optical integration, external-laser, advanced-packaging, and mass-production capabilities.

LPO could reduce the value of module-level DSPs while increasing the importance of switching silicon, system design, and high-quality analog components.

CPO could shift more value from standalone modules toward switch silicon, optical engines, advanced packaging, and integrated systems.

Technology migration therefore does more than expand the market. It changes who captures the economics. Investors need to identify whether a company sits on the expanding or contracting side of the profit pool.

Competition in optical connectivity is not primarily a race to build the most capacity. It is a race to define the next network generation, qualify with leading customers, and deliver reliably at scale.

In the near term, 800G and 1.6T pluggable products remain the clearest earnings drivers. Over the medium term, silicon photonics and 200G-per-lane architectures may raise the value of photonic chips, lasers, and advanced packaging. Over the longer term, LPO and CPO may redistribute value among DSPs, modules, switching silicon, and system platforms.

The purpose of technology analysis is therefore not to predict that one architecture will eliminate all others. It is to determine where each architectural transition moves value—and which companies can convert technical leadership into orders, margins, and cash flow.

5 Key Companies and Investment Strategy — Selecting the Real Profit Pools

At the investment stage, the key question is no longer whether the optical-connectivity industry will grow. It is which companies can convert that growth into revenue, margins, and cash flow.

Companies across the value chain earn different types of returns. Switching-silicon providers monetize architectural upgrades and standards control. Optical-component suppliers monetize higher speeds and supply scarcity. System vendors monetize customer access and platform control. Module manufacturers are more directly exposed to shipment growth and product-generation transitions.

Optical investing should therefore move beyond simple industry beta and instead focus on profit-pool exposure, earnings visibility, and technology timing. Our central view is that the most attractive near-term opportunities are companies already converting 800G, 1.6T, and AI-network orders into revenue. Medium-term opportunities lie in silicon photonics, 200G-per-lane architectures, high-speed DSPs, lasers, and coherent connectivity. CPO should be treated as a long-term technology option rather than the primary source of current earnings.

5.1. Broadcom: One of the Most Complete AI-Network Assets

Broadcom should not be viewed merely as a supplier of chips used in optical modules. It spans switching silicon, custom AI accelerators, optical DSPs, high-speed SerDes, connectivity, and CPO platforms. This gives it broad control over the transition from standalone compute toward system-level AI networking.

Its investment thesis has three layers.

First, larger AI clusters require continuously higher switching capacity. More GPUs require not only additional accelerators, but also faster and more complex network fabrics. Port count, bandwidth, and congestion-management capability directly influence cluster efficiency, creating structural demand for switching silicon.

Second, hyperscaler custom accelerators expand Broadcom’s role in AI compute. Large cloud customers are seeking lower cost and power consumption for specific workloads and greater control over their infrastructure. Broadcom’s ability to provide both custom compute and networking creates multiple revenue channels within AI capital expenditure.

Third, CPO and optical integration provide a long-term technology option. Even if near-term CPO revenue remains limited, Broadcom already occupies a position in switch-silicon and packaging architecture. If optics move closer to switching silicon, profit pools may shift further toward integrated system platforms.

The primary risks are customer concentration, elevated market expectations, and exposure to a potential slowdown in hyperscaler capital expenditure. Nevertheless, Broadcom remains one of the most strategically comprehensive listed exposures to AI networking.

5.2. Arista Networks: Direct Exposure to AI Ethernet Expansion

Arista’s core strength is not optical-module manufacturing, but high-performance Ethernet switching and network operating software. As AI clusters scale, Ethernet is improving its latency, congestion-management, and reliability capabilities relative to specialized alternatives such as InfiniBand. Arista provides relatively direct listed exposure to the expansion of Ethernet-based AI fabrics.

Its investment thesis has two components.

First, AI data centers increase both switch-port counts and value per port. Larger clusters require additional leaf, spine, and back-end networking systems. Even if GPU growth moderates, greater network complexity can continue to increase system value.

Second, Arista’s EOS software creates meaningful customer stickiness. AI networks require congestion control, traffic management, automation, and stability—not simply switching hardware. Software and hardware integration support stronger margins and customer retention than conventional equipment manufacturing.

The main risks are customer concentration, hyperscaler in-house solutions, competition from NVIDIA’s Spectrum-X, and a valuation that already reflects strong growth. Key indicators include AI-network revenue, order growth, cloud capital expenditure, and Ethernet penetration in large training clusters.

5.3. Marvell: High-Beta Exposure to Connectivity and Custom Compute

Marvell is a higher-beta asset within AI networking and optical connectivity, with exposure to optical DSPs, PAM4, SerDes, high-speed connectivity, and custom compute. Compared with Broadcom, Marvell is smaller and more sensitive to individual AI programs, creating greater upside potential but also higher volatility.

Its central thesis is that AI-cluster upgrades create two simultaneous demand pools: faster electro-optical connectivity and more custom silicon. The transition toward 1.6T modules and 200G per lane raises the technical requirements for DSPs and connectivity silicon, while hyperscaler ASIC investment creates a second growth engine.

Marvell’s positioning is attractive if these programs enter volume production, because revenue and margins can respond strongly. However, dependence on a limited number of projects and customers creates quarterly volatility. In addition, faster adoption of LPO or other DSP-light architectures could reduce module-level DSP value.

5.4. Coherent: A Broad Optical Platform Dependent on High-End Product Execution

Coherent is one of the broadest US-listed optical platforms, spanning lasers, photonic components, materials, silicon photonics, transceivers, and optical engines. Its opportunity is not simply industry volume growth, but a favorable shift in product mix.

As 800G, 1.6T, and silicon photonics scale, a greater contribution from advanced lasers, optical engines, and integrated products could improve earnings quality. If Coherent converts its technology breadth into reliable production, higher yields, and hyperscaler orders, margin upside may exceed that of traditional low-speed optical businesses.

The company may also benefit from CPO and external-laser demand. Even silicon-photonics systems generally require reliable light sources, creating long-term value for suppliers with laser technology and manufacturing capability.

The main risks are portfolio complexity, debt associated with historical acquisitions, and uneven conditions across end markets. Investors should focus on data-communications revenue, 1.6T progress, silicon-photonics and laser orders, margin improvement, and balance-sheet repair.

5.5. Lumentum: A More Focused Laser and Optical-Component Play

Lumentum offers relatively focused exposure to lasers, EMLs, photonic chips, and cloud-data-center optical components. Its investment thesis rests on rising value per optical link and limited supply elasticity in high-end laser products.

The transition from 800G to 1.6T increases requirements for optical power, stability, and modulation performance. Silicon photonics does not eliminate laser value. It may increase demand for external continuous-wave light sources used in silicon-photonics and CPO systems.

Lumentum therefore remains strategically relevant across multiple next-generation architectures. Its potential upside comes from product scarcity and high-speed optical demand, while risks include customer concentration, technology-route changes, and capacity-cycle volatility.

Order visibility, utilization rates, customer qualification, and product mix are the most important variables.

5.6. NVIDIA: A System-Level Optical-Network Beneficiary

NVIDIA is not a traditional optical company, but it cannot be excluded from the industry framework. Through InfiniBand, Spectrum-X, ConnectX, switches, NVLink, and full-system architectures, NVIDIA is turning networking into an integrated component of its AI platform.

Its core advantage is compute-network co-design. Because NVIDIA controls accelerators, systems, and software, it can optimize the complete cluster. Customers are therefore not simply buying network hardware, but a more integrated performance solution.

Its optical-connectivity exposure comes through rising networking revenue and its influence over future interconnect standards and architecture. However, NVIDIA’s valuation remains primarily driven by GPUs, AI systems, and overall capital spending. It should be viewed as a system-level core asset rather than a high-beta optical-component substitute.

5.7. Cisco, Ciena, Corning, and Keysight: Complementary Exposure

Cisco offers exposure to enterprise networking, Silicon One, Acacia coherent optics, and system integration. It may provide less AI-network beta than Arista or NVIDIA, but its customer base and product breadth offer more balanced exposure. Key variables include AI Ethernet orders, Silicon One adoption, and Acacia growth.

Ciena is primarily exposed to DCI and coherent optical transport. As AI infrastructure expands across campuses and regions, demand for high-capacity inter-data-center links should rise. Its revenue cycle is more project-driven and sensitive to carriers and large network deployments.

Corning provides exposure to fiber, connectivity, and cabling systems. Larger AI campuses require greater fiber counts, higher connection density, and more complex physical networks. Its upside may be lower than that of high-speed silicon, but demand is tied to tangible infrastructure construction.

Keysight benefits from the increasing complexity of testing 800G, 1.6T, 200G-per-lane, and CPO systems. Test equipment offers technical barriers and customer stickiness without requiring investors to select a single module winner.

Optical connectivity remains one of the most important structural themes in AI infrastructure, but the opportunity is no longer simply about rising module demand. The most attractive companies control at least one scarce capability: network architecture, critical silicon, light sources, customer access, or difficult-to-replicate production.

For the second half, core allocations should prioritize companies with the clearest earnings conversion and strongest strategic positions. Higher-beta exposure can focus on 1.6T, silicon photonics, lasers, and high-speed DSPs. CPO and other longer-term architectures retain option value, but require customer deployment and revenue validation.

The central investment thesis is not that every optical company benefits equally. It is that as networking captures a larger share of AI-system value, profit pools will concentrate among a limited number of technology and system controllers.

6 Key Risks

The principal risks discussed in this report include:

Lower-than-expected AI capital expenditure.
Optical-connectivity demand is highly dependent on capital spending by North American hyperscalers and major AI-infrastructure operators. If AI monetization develops more slowly than expected, or cloud providers reduce investment in data centers, switching systems, and networking equipment, demand for high-speed optical modules, components, and related silicon could fall below expectations.

Delays in 800G and 1.6T product deployment.
High-speed optical products require extended customer qualification, interoperability testing, and manufacturing ramp-up. Delays in 1.6T adoption or slower-than-expected yield improvement in technologies such as 200G per lane could postpone revenue and earnings realization.

Technology-roadmap risk.
Silicon photonics, LPO, CPO, conventional pluggable optics, and high-speed copper remain competing or complementary architectures. If the market adopts a different technical route from the one prioritized by a company, current R&D investment and product positioning may not translate into expected revenue.

Intensifying competition and price erosion.
As suppliers expand capacity and more competitors enter the high-speed optical market, optical modules, photonic components, and certain silicon products may face pricing pressure. If selling prices decline faster than production costs, gross margins and profitability could weaken.

High customer concentration.
High-speed data-center demand is concentrated among a limited number of North American hyperscalers and large networking customers. Changes in procurement timing, supplier replacement, or broader multi-sourcing strategies could lead to significant volatility in orders, revenue, and market share.

Supply-chain and capacity constraints.
High-end lasers, DSPs, photonic chips, advanced packaging, and test equipment may remain supply constrained. Limited availability of critical components or slower-than-expected yield improvement at new facilities could affect shipment volumes and revenue recognition.

Valuation and market-volatility risk.
Some optical-connectivity and AI-networking companies already reflect high growth expectations. If orders, earnings, or cash flow fall short of market expectations, their share prices may experience valuation compression and significant volatility even if the long-term industry trend remains intact.

Macroeconomic and policy risk.
Higher interest rates, slower economic growth, trade restrictions, export controls, and geopolitical developments could affect hyperscaler capital expenditure, global supply chains, and technology-sector valuations, creating downside risk for optical-connectivity equities.

Disclaimer:

Sahm is subject to the supervision and control of the CMA, pursuant to its license no. 22251-25 issued by CMA.

The Information presented above is for information purposes only, which shall not be intended as and does not constitute an offer to sell or solicitation for an offer to buy any securities or financial instrument or any advice or recommendation with respect to such securities or other financial instruments or investments. When making a decision about your investments, you should seek the advice of a professional financial adviser and carefully consider whether such investments are suitable for you in light of your own experience, financial position and investment objectives. The firm and its analysts do not have any material interests or conflicts of interest in any companies mentioned in this report.

Performance data used in this report are obtained or compiled from company filings and disclosures, official stock exchanges and regulatory authorities, recognized industry associations, reputable financial databases, and other third-party sources that Sahm Capital Financial Company believes to be reliable.

IN NO EVENT SHALL SAHM CAPITAL FINANCIAL COMPANY BE LIABLE FOR ANY DAMAGES, LOSSES OR LIABILITIES INCLUDING WITHOUT LIMITATION, DIRECT OR INDIRECT, SPECIAL, INCIDENTAL, CONSEQUENTIAL DAMAGES, LOSSES OR LIABILITIES, IN CONNECTION WITH YOUR RELIANCE ON OR USE OR INABILITY TO USE THE INFORMATION PRESENTED ABOVE, EVEN IF YOU ADVISE US OF THE POSSIBILITY OF SUCH DAMAGES, LOSSES OR EXPENSES.