CPO: Reshaping the Value of AI Interconnect

In-Depth Research Analysis:

Executive Summary:

The core constraint in AI infrastructure is gradually shifting from “whether there is enough compute” toward “how efficiently a growing amount of compute can work together.” GPUs and XPUs remain the largest value pools in AI systems, while HBM and other memory resources continue to gain importance. However, as clusters scale from individual racks to thousands or even larger numbers of accelerators, network bandwidth, data-movement power consumption, and interconnect reliability are becoming increasingly important constraints on overall system efficiency.

This shift is changing the internal value structure of AI capital expenditure. Global AI server shipments are expected to grow by more than 28% in 2026, while Ethernet optical transceiver sales are expected to increase by roughly 65% over the same period. Shipments of 800G modules are expected to more than double, while 1.6T enters the tens-of-millions-of-ports scale from a relatively low base. These datasets are based on different statistical methodologies and should not be used to derive a precise networking share of data-center BOM. Nevertheless, they point to the same underlying trend: connectivity demand is growing faster than server volumes alone.

We therefore do not view the CPO investment thesis simply as “the next generation of optical modules.” Instead, it reflects three layers of value creation. First, larger AI clusters require more connections, creating volume growth. Second, upgrades from 800G to 1.6T and eventually 3.2T increase bandwidth and raise the value content of selected components, creating higher value per connection. Third, NPO and CPO move optics closer to switch ASICs and XPUs, redistributing value among DSPs, TIAs/Drivers, lasers, silicon photonics, fiber connectivity, and advanced packaging, creating a broader value-chain restructuring.

The central research question is therefore not which companies can be labeled as “CPO plays,” but rather:

Which segments can simultaneously benefit from higher connection volumes, rising value content per system, and visible commercialization?

1 AI Data Center Value Structure Is Changing — Why Connectivity Is Becoming More Important

1.1 GPUs Remain the Value Center, but System Bottlenecks Are Expanding from Compute to Data Movement

AI data-center hardware investment is still centered on computing.

In a typical AI system, accelerators and the HBM closely integrated with them remain the largest pools of semiconductor value. The reason is straightforward: model training and inference first require massive matrix-computing capacity, while GPUs and XPUs themselves incorporate advanced process nodes, HBM, and complex packaging, resulting in substantially higher unit value than conventional server chips.

As the number of AI chips continues to increase, the second bottleneck is increasingly memory. Larger models and longer context windows require higher-capacity and higher-bandwidth HBM and DRAM. Industry forecasts suggest AI server shipments could grow by more than 28% in 2026, while HBM consumption could rise by more than 70%, implying that memory content per compute platform is also continuing to increase.

Once computing and memory expand, however, a third question becomes increasingly important:

Can these computing resources exchange data fast enough?

Higher performance from an individual GPU does not automatically translate into proportional improvements in cluster-level performance.

During training, accelerators frequently synchronize parameters. During inference, data must also continuously move among compute, memory, storage, and networking resources. As systems expand from dozens of GPUs to thousands or more, latency, power consumption, and reliability associated with data movement become increasingly significant.

The evolution of AI infrastructure bottlenecks can therefore be simplified as:

Compute → Memory → Interconnect

This does not mean compute or memory are becoming less important. Rather, it means connectivity is evolving from supporting infrastructure into a more important determinant of effective compute utilization.

1.2 Connectivity Value Growth Comes from Three Variables, Not Merely More Optical Modules

AI connectivity can be understood through a simple framework:

AI Connectivity Value
≈ Number of Accelerators × Connections per Accelerator × Bandwidth per Connection × Value Content per Connection

All four variables are currently changing.

The first is the growing number of accelerators.

The second is the increasing number of resources connected to each accelerator. As AI systems expand from single-rack to multi-rack architectures, both scale-up and scale-out networks grow.

The third is the increase in bandwidth per connection. High-speed data centers are progressing from 400G and 800G toward 1.6T, while preparing for 3.2T.

The fourth, and most important for CPO, is that the semiconductor and optical content embedded in each connection is changing.

The future growth of the connectivity industry should therefore not be interpreted simply as:

More GPUs → proportionally more optical modules.

A more useful framework is:

Quantity Growth + Speed Upgrade + Architecture Migration

Together, these three factors determine the evolution of connectivity value.

1.3 Optical Communications Are Showing Greater Growth Elasticity Than Server Volumes

Global optical transceiver and related product sales reached approximately USD 26.8 billion in 2025, representing year-over-year growth of around 74%. Sales in 1Q26 approached USD 10 billion, increasing by more than 90% year over year. Industry forecasts indicate that the Ethernet optical transceiver market could still grow by roughly 65% in 2026.

Within the speed mix, 800G module shipments are expected to more than double in 2026, while 1.6T ports could expand from a relatively small base to the tens-of-millions range. Sales of 1.6T-related chipsets are expected to exceed USD 2 billion.

By comparison, AI server shipments are expected to grow by around 28% in 2026.

These are not directly comparable datasets, and it would be inappropriate to mechanically divide 65% by 28% to infer a precise change in connectivity’s share of data-center value.

The difference is nevertheless directionally meaningful:

AI networking growth is being driven not only by more servers, but also by more bandwidth required per server and higher-speed connections.

This is the core reason we believe connectivity could capture an increasing share of incremental AI infrastructure value.

2 Why Traditional Pluggable Optics Are Beginning to Encounter Limits

2.1 The Problem Begins Not in the Fiber, but Between the ASIC and the Optical Module

The signal path in a conventional data-center switch can be simplified as:

Switch ASIC → SerDes → PCB → DSP → Pluggable Optical Module → Fiber

The switch ASIC generates high-speed electrical signals, which travel across the PCB to front-panel pluggable modules, where they are converted into optical signals and transmitted through fiber.

At 100G per lane, this architecture has been highly economical.

Pluggable modules are standardized, supported by multiple suppliers, and relatively easy to service. If a module fails, it can be replaced without replacing the entire switch.

As per-lane speed rises toward 200G and beyond, however, the economics begin to change.

High-speed electrical signals suffer increasing loss as they travel through PCB traces, connectors, and longer electrical paths. Compensating for that loss requires more capable DSPs, retimers, and analog front-end circuitry.

The result is:

Higher speed → Greater electrical loss → More complex compensation → Higher power and thermal pressure

As SerDes moves to 200G-class signaling and beyond, the electrical path from ASIC to front-panel optics becomes increasingly difficult to scale economically.

2.2 CPO Fundamentally Moves the Electrical-to-Optical Conversion Closer to the ASIC

CPO, or Co-Packaged Optics, is not simply a smaller version of a traditional optical module.

Its core idea is:

Move the optical engine much closer to the switch ASIC or XPU, sharply shorten the high-speed electrical path, and convert the signal into light earlier.

The conventional architecture:

ASIC → long electrical path → optical module

gradually becomes:

ASIC → very short electrical link → optical engine → fiber

As the electrical path shortens, less high-power signal compensation is required.

Broadcom’s 102.4T Tomahawk 6 Davisson integrates optical engines with the switching ASIC through advanced packaging. The company has stated that its architecture can reduce optical-interconnect power consumption by approximately 70% relative to conventional pluggable optics.

TSMC has stated that its COUPE on-substrate solution entered production in 2026 and can deliver approximately 2x better power efficiency and roughly 10x lower latency than board-level pluggable architectures.

These figures are based on different vendor architectures and test conditions and should not be directly compared with one another. Their primary significance is that power consumption and latency have become real engineering constraints driving CPO commercialization.

2.3 Pluggable, NPO, and CPO Are Unlikely to Follow a Simple Linear Replacement Path

The more realistic industry path is not:

Pluggable → NPO → CPO → traditional modules disappear.

NPO, or Near-Packaged Optics, moves optical modules closer to the ASIC while generally preserving a higher degree of modularity and serviceability.

CPO goes further in terms of integration.

The two approaches therefore represent different trade-offs among:

Power, Bandwidth Density, Cost, and Serviceability

NPO shortens the electrical signal path while preserving better maintenance and modularity.

CPO offers stronger advantages in power and density but imposes greater requirements on packaging yield, thermal management, testing, and serviceability.

As of September 2026, industry data do not indicate that traditional pluggable modules are entering rapid decline. Ethernet modules still account for roughly 80% of the optical transceiver market by sales, and the majority remain conventional retimed pluggable products.

We therefore believe:

Pluggable, NPO, and CPO are likely to coexist for several years, serving different bandwidth, distance, and system-architecture requirements.

3 First Investment Logic — The Number of Connections Continues to Grow

The most basic foundation of the CPO thesis is straightforward:

AI requires an increasing number of physical connections.

3.1 Scale-Up: New Optical Demand Is Emerging Inside the Computing System

Historically, optical networking primarily connected servers, racks, and data centers.

As AI systems scale, optics are moving closer to the compute domain.

Scale-up networking focuses on high-bandwidth, low-latency communication among GPUs/XPUs, CPUs, and memory resources.

Once copper interconnects within a rack reach physical limits in distance and power consumption, optical connectivity becomes increasingly relevant for cross-rack scale-up.

This potentially creates a market that optical communications historically had limited exposure to:

Data movement previously handled by PCB traces and copper interconnects may increasingly migrate toward optical links.

The most important long-term incremental opportunity for CPO/NPO may therefore not be replacing existing pluggable optics, but rather:

Extending optics from rack-to-rack connectivity toward chip-to-chip and intra-rack/cross-rack connectivity.

3.2 Scale-Out: High-Speed Ethernet and InfiniBand Remain the Largest Commercial Markets Today

Scale-out networks connect servers and racks.

This market already supports large-scale 400G and 800G deployment and is transitioning toward 1.6T.

In 2026, 1.6T moved beyond the sampling phase into actual volume production. For example, Applied Optoelectronics disclosed in March that it had received an initial volume order exceeding USD 200 million for 1.6T transceivers from a major hyperscale customer, with shipments expected to begin in 3Q26.

This demonstrates that 1.6T is no longer merely a technology roadmap.

At the same time, 800G demand remains strong and is not being displaced quickly by 1.6T.

The next several years are therefore more likely to feature simultaneous growth across multiple generations rather than a simple one-generation replacement cycle.

3.3 Scale-Across: AI Clusters Are Expanding Across Campuses and Regions

As power availability and physical space increasingly constrain individual data centers, AI compute may be distributed across multiple campuses or regions.

This should support demand for longer-reach optical connections such as 800G ZR/ZR+ and 1.6T coherent links.

AI connectivity is therefore expanding from:

inside the server

toward:

inside the rack → rack-to-rack → data-center-to-data-center

The proliferation of connection layers is itself an important driver of optical value growth.

4 Second Investment Logic — From 800G to 3.2T, Value per Connection Is Changing

If the first investment logic is primarily about Q, the second is about P and content.

4.1 Higher Speeds Require More Channels and More Complex Signal Processing

Data-center optical communications are moving through:

400G → 800G → 1.6T → 3.2T

In September 2026, MACOM announced a 3.2T front-end chipset based on eight 448Gbps PAM4 channels, incorporating TIAs, photodiodes, and drivers.

Semtech has also introduced 224Gbps-per-lane TIAs and drivers targeting 800G, 1.6T, 3.2T, as well as NPO and CPO applications.

This indicates that the analog front end is transitioning from 100G/lane toward 200G/lane and eventually 400G/lane.

Higher speeds raise requirements for:

bandwidth,

linearity,

noise performance,

power efficiency,

and signal compensation.

Speed upgrades therefore affect not only module pricing, but also the technical complexity and value of selected semiconductor components.

4.2 NPO Could Significantly Increase the Value Density of TIAs and Drivers

In conventional retimed pluggable architectures, DSPs perform substantial signal recovery and compensation.

NPO seeks to reduce power consumption by reducing or removing traditional DSP functionality.

Signal loss, however, does not disappear.

Some functions previously handled by DSPs must instead be supported by:

SerDes + Driver + TIA

This can increase the importance of the analog front end.

Based on Sahm Capital Research’s indicative supply-chain research, TIAs and Drivers together could account for close to 30% of total BOM in certain 3.2T NPO architectures. This figure is an industry-research estimate rather than an official vendor quotation and can vary materially depending on channel count, modulator choice, and system architecture.

The key takeaway is:

Removing the DSP does not necessarily reduce total electrical semiconductor value; instead, value may migrate from digital signal processing toward high-performance analog front-end components.

4.3 The TIA/Driver Opportunity Is Highly Architecture-Dependent

It would be incorrect to conclude:

Faster CPO adoption → uniformly higher independent TIA/Driver demand.

Certain highly integrated CPO architectures can integrate the Driver—and potentially parts of the receive path—more deeply into the EIC/PIC system, reducing the need for discrete external analog chips.

NPO, by contrast, generally retains a more independent optical engine and may therefore rely more heavily on external TIAs and Drivers.

Accordingly:

The TIA/Driver opportunity may be more directly linked to NPO adoption than to CPO in general.

This is one of the most important principles in CPO investment analysis:

Architecture matters more than headline market size.

5 Third Investment Logic — CPO Is Redistributing Value Across the Optical Supply Chain

CPO does not simply create a new product category.

Its deeper effect is to change:

Where the value previously embedded inside pluggable optical modules ultimately resides.

5.1 DSP: One of the Most Exposed Areas to Potential Value Migration

DSPs in high-speed pluggable modules are primarily used to compensate for signal degradation between the switch ASIC, PCB traces, connectors, and front-panel module.

As the optical engine moves closer to the ASIC, that electrical distance falls sharply.

CPO can therefore reduce the need for high-power DSP compensation.

Broadcom’s 102.4T CPO product uses directly driven optical engines, illustrating this architectural direction.

CPO therefore does not imply that every optical semiconductor category will gain value simultaneously.

Conventional retimed DSPs could face some of the clearest long-term value migration pressure.

However, this process is unlikely to happen immediately.

Large volumes of 800G and 1.6T networking are still expected to use conventional pluggable modules for years, and fully retimed Ethernet transceivers remain one of the largest near-term growth drivers in the optical semiconductor market.

5.2 TIA and Driver: Value May Migrate from Digital Compensation Toward the Analog Front End

In NPO, LPO, and other linear architectures, TIAs and Drivers carry a greater share of the burden for signal integrity.

This is one reason MACOM and Semtech are accelerating development of 224G and 448G analog front-end products.

MACOM has demonstrated 112G- and 226G-per-lane Drivers, TIAs, and photodiodes for NPO applications, as well as an 8×448G 3.2T front-end solution.

Semtech explicitly positions its latest 224G Linear TIA/Driver products across a range of low-power architectures from LPO to NPO and CPO.

The analog semiconductor opportunity is therefore not simply about “more optical modules,” but rather:

Higher analog performance requirements per lane, combined with a shift of certain DSP functions into the front end.

5.3 Laser: Value Does Not Disappear — It Moves Toward External Laser Sources

CPO places optical engines close to high-heat ASICs, while lasers themselves are temperature-sensitive.

Many CPO architectures therefore locate the laser externally and deliver continuous-wave light to the silicon photonics engine through fiber.

This creates the External Laser Source, or ELS.

The implication is:

CPO does not reduce the importance of lasers; it may increase the strategic value of high-power, highly reliable CW lasers.

Lumentum has introduced ELSFP external laser modules for CPO and NPO applications and demonstrated 800mW-class high-power lasers and multi-wavelength DWDM external light sources in 2026.

More importantly, NVIDIA announced a USD 2 billion investment in Lumentum in March 2026 alongside a multiyear strategic agreement, including multi-billion-dollar purchase commitments for advanced lasers and access to future capacity.

This suggests high-power laser demand is no longer merely a long-term technical concept. It is already entering real capacity planning and procurement arrangements.

5.4 Silicon Photonics: Optical Functions Are Moving Closer to the Compute Core

As optical engines move closer to the ASIC package, the importance of silicon photonics naturally increases.

Silicon photonics can integrate modulators, waveguides, and other optical functions within a chip-based platform and combine them with advanced packaging.

TSMC’s COUPE initiative demonstrates that foundries are increasingly participating directly in this part of the value chain.

The company plans to begin production of its COUPE on-substrate CPO architecture in 2026.

This suggests that future CPO competition is no longer simply:

competition among optical module vendors

but increasingly:

system-level competition among ASICs, foundries, silicon photonics, and advanced packaging.

5.5 Fiber and Connectors: Moving the Optical Engine into the Package Increases Fiber-Management Complexity

CPO removes the traditional front-panel optical module, but it does not remove fiber.

On the contrary.

Traditional front-panel modules connect directly to standardized fiber connectors.

In CPO systems, fiber must travel from optical engines located near the ASIC package to the front panel. This raises demand for:

Fiber Array Units,

high-density connectors,

fiber shuffles,

small-bend-radius fiber,

and detachable fiber-to-chip connections.

Corning has demonstrated a complete passive connectivity platform for CPO, including Fiber Array Units, fiber-to-chip connectors, and pre-assembled fiber-management systems.

In May 2026, NVIDIA and Corning announced a long-term partnership under which Corning plans to expand U.S. optical-connectivity manufacturing capacity by 10x and increase U.S. optical-fiber capacity by more than 50%.

This provides another important industry signal:

Moving optics closer to the chip does not reduce fiber content; instead, it raises the requirements for high-density connectivity and precision assembly.

5.6 Advanced Packaging: CPO May Ultimately Be as Much a Packaging Revolution as an Optical One

The fundamental challenge of CPO is that it must integrate:

Electronics + Photonics + Fiber

Traditional advanced semiconductor packaging primarily solves electrical connections between chips.

CPO must additionally address:

optical coupling,

fiber alignment,

thermal management,

optical-engine testing,

Known Good Die / Known Good Optical Engine,

and overall package yield.

This moves advanced packaging from a back-end manufacturing function toward a more integral part of CPO system performance.

As NVIDIA’s Spectrum-X Photonics enters production, its ecosystem already spans TSMC silicon-photonics manufacturing, SPIL packaging and testing, laser modules, and system integration.

We therefore view CPO not simply as “optical modules becoming chips,” but rather as:

the convergence of optical communications and advanced semiconductor packaging.

6 Fourth Investment Logic — System Integration Is Becoming More Valuable

Historically, the optical supply chain could be highly specialized.

One supplier manufactured lasers.

Another produced PICs.

Another supplied fiber.

A module vendor then assembled these components into a transceiver.

As CPO and NPO move optics closer to the ASIC, system tolerance becomes tighter.

Customers increasingly need suppliers that can solve multiple issues simultaneously:

light generation,

modulation,

detection,

fiber coupling,

packaging,

testing,

and reliability.

Competitive advantage may therefore expand from:

Component Performance

toward:

Integration + Manufacturing Scale

6.1 Coherent: Moving from Component Supplier Toward Integrated Optics

Coherent launched its PhotonLink platform in 2026, covering CPO, NPO, and chip-to-chip optical interconnects.

The company has disclosed more than 10 CPO customer engagements, more than 10 NPO engagements, and more than five chip-to-chip engagements, with PhotonLink revenue expected to begin ramping from 4Q26.

PhotonLink spans InP lasers, VCSELs, silicon photonics, fiber, micro-optics, and detectors.

This indicates a business model extending beyond individual components toward more complete optical-system solutions.

Coherent has previously shipped more than 300 million InP lasers and more than 1 billion InP/GaAs photodetectors. That manufacturing scale could become increasingly relevant as CPO moves toward volume deployment.

6.2 Lumentum: High-Power Lasers Are Among the Most Visible Early CPO Value Pools

Compared with a complete optical engine, the commercialization path for lasers is currently more visible.

NVIDIA has explicitly incorporated Lumentum into its silicon-photonics ecosystem and secured future advanced-laser capacity through a long-term strategic agreement and a USD 2 billion investment.

ELS and CW lasers therefore represent one of the CPO value-chain segments with relatively high near-term commercial visibility.

Long-term economics will still depend on:

laser ASPs,

optical power required per system,

the number of WDM channels,

and how different silicon-photonics architectures affect external-laser content.

6.3 Vertical Integration Does Not Mean Every Supplier Gains Value

Higher integration can also produce the opposite outcome for certain vendors.

If ASIC suppliers, foundries, and large optical companies increasingly provide complete solutions, some standalone component suppliers could face greater pricing pressure.

The competitive question in the CPO era is therefore not:

Does a company make an optical component?

It is:

Does the company own a difficult-to-replace capability, and can that capability remain embedded in the final system architecture?

7 Commercialization — CPO Has Entered Early Production, but the Industry Is Still at an Early Stage

Evaluating the CPO opportunity requires a strict distinction among:

Technology Demo

Qualification

Initial Production

Volume Deployment

This is essential to avoid overstating future revenue.

7.1 Broadcom: CPO Has Entered Actual Shipment

Broadcom announced in October 2025 that it had begun shipping its 102.4T Tomahawk 6 Davisson CPO switch.

The platform supports configurations including up to 64×1.6TbE or 128×800GbE and incorporates 512 lanes of 200G optical I/O.

This means CPO has moved beyond the technology-demonstration stage.

However, Broadcom still classifies the relevant product lifecycle as Limited Release, meaning that “shipping” should not yet be interpreted as broad-scale market penetration.

7.2 NVIDIA: CPO Is Becoming Part of the AI System Platform

In May 2026, NVIDIA announced that Spectrum-X Ethernet Photonics had entered production alongside the Vera Rubin platform.

Spectrum-X Photonics uses 200Gbps SerDes and targets scale-out and scale-across networks.

NVIDIA’s SN6800 CPO switch offers 409.6Tbps of bandwidth and can support 512×800G ports or 2,048×200G ports. The company states that the architecture can reduce network power consumption per 1.6T port by approximately 5x relative to conventional pluggable solutions.

This marks an important transition: CPO is no longer simply an independent networking product, but is beginning to form part of a complete AI infrastructure platform.

7.3 TSMC: Foundry Production Capability Is Beginning to Mature

TSMC’s COUPE entered the on-substrate CPO production stage in 2026, suggesting silicon-photonics manufacturing is moving from pilot lines toward industrialization.

The broader significance is:

Large-scale CPO ultimately requires mature foundry and advanced-packaging platforms rather than isolated proprietary processes.

7.4 Marvell: Strong Technology Readiness, but Demo and Revenue Must Still Be Distinguished

At ECOC 2026, Marvell demonstrated a 102.4T CPO platform based on 200G/lane silicon photonics, while also presenting next-generation 400G/lane optical technology.

These demonstrations show that the company is actively participating in next-generation CPO development.

However, current public information remains centered primarily on product demonstration and technology readiness.

It would therefore be inappropriate to treat these disclosures as evidence of already recognized large-scale CPO revenue.

The discipline required in CPO research remains:

Demo ≠ Design Win ≠ Production ≠ Revenue

7.5 NPO Commercialization May Come Later, but Its Value-Chain Effects Deserve Early Attention

Compared with CPO, NPO offers advantages in serviceability and open ecosystem participation, but the industry is still developing specifications, system validation, and supply-chain maturity.

Based on public information and Sahm Capital Research’s industry research, we believe 2027 is more likely to represent an initial, small-scale NPO deployment phase, while broader volume production may occur more meaningfully from 2028 onward. This timing is an industry-research estimate rather than formal guidance from any single company.

This suggests current investment research should focus on:

Which parts of the supply chain are already receiving orders, expanding capacity, or entering qualification before broader revenue is fully realized.

8 Industry Mapping — Value Capture Differs Significantly Across Segments

From an investment-research perspective, we believe the appropriate framework is not to build a list of “CPO concept stocks,” but rather to evaluate:

Value Content Change × Commercialization Visibility × Financial Materiality

The table below is intended solely as an industry-mapping framework and does not constitute stock ratings or investment recommendations.

The most important distinction is:

Industry exposure is not the same as financial sensitivity.

A large semiconductor company may play a strategically important role in defining the CPO architecture, while CPO-related revenue remains a relatively small portion of its total business.

Conversely, a smaller optical component supplier may participate in only one part of the system, but if CPO/NPO revenue becomes a rapidly increasing share of its business, the financial impact may be more material.

Company-level analysis must therefore ultimately return to:

Revenue Exposure, not merely Technology Exposure.

9 Translating Industry Growth into Financial Models

The financial contribution of CPO/NPO can be simplified as:

Revenue
= System Shipments × Content per System × Market Share

All three variables matter.

The first is system shipments.

If CPO switches remain limited to small-scale deployments, even high per-system value will not create meaningful industry revenue.

The second is content per system.

Architectures differ significantly. NPO may increase the value of discrete TIAs and Drivers, while highly integrated CPO architectures may reduce those same components.

The third is market share.

Technical capability does not automatically translate into production orders.

Profitability can then be simplified as:

Incremental Profit
= CPO/NPO Revenue × Gross Margin − Incremental R&D / Capacity Cost

This creates several very different possible company outcomes.

One company may have exposure to a very large addressable market, but CPO may remain immaterial to its total revenue.

Another may have high content per system, but products may not yet be in volume production.

A third may already have production revenue, but face falling ASPs as competition increases.

A fourth may simultaneously benefit from higher volumes, higher value content, and market-share gains.

The last combination is the one most likely to generate a sustained financial impact.

10 Risks and Key Validation Indicators

CPO remains at an early stage of industrialization, and its commercial value still requires ongoing validation.

Key risks include slower-than-expected CPO and NPO deployment; traditional pluggable architectures extending their lifecycle through DSP improvements and LPO; advanced-packaging yield, thermal-management, and maintenance costs remaining elevated; greater integration of TIAs, Drivers, or lasers into system architectures reducing value available to standalone suppliers; and a slowdown in AI infrastructure capital expenditure.

From an industry-validation perspective, the most important indicators are not simply the number of CPO products announced by vendors, but rather:

actual deployment volumes of Broadcom and NVIDIA CPO switches;

shipment growth in 1.6T and 3.2T ports;

the transition of NPO products from qualification to volume production;

orders and capacity for ELS/CW lasers;

optical-engine and advanced-packaging yields;

and actual CPO-related revenue contribution at relevant companies.

Research Conclusion

The importance of CPO does not come from the emergence of another optical-communications technology acronym.

The deeper structural change is:

AI data centers are evolving from an investment structure dominated almost entirely by compute toward an architecture in which compute, memory, and interconnect jointly determine system efficiency.

GPUs/XPUs and HBM still represent the largest absolute value pools, but as AI clusters scale, connectivity is showing greater marginal growth.

That growth is driven by three forces.

First, more accelerators require more connections, increasing the number of links.

Second, upgrades from 800G to 1.6T and 3.2T raise bandwidth per connection and increase the value of selected components.

Third, NPO and CPO change the traditional optical-module architecture, redistributing value from DSPs toward TIAs/Drivers, lasers, silicon photonics, fiber connectivity, and advanced packaging.

The most important investment logic is therefore not:

“Optical module demand is growing, so the entire optical supply chain benefits.”

Rather, it is:

“The value of AI connectivity is increasing while value is simultaneously being redistributed across the supply chain; the most relevant industry opportunities are those where volume, value content, and architectural share improve together.”

From the current commercialization perspective, CPO has moved from technology demonstration into early production. Broadcom and NVIDIA have CPO platforms entering shipment or production, TSMC’s silicon-photonics manufacturing capabilities are progressing toward industrialization, while optical suppliers such as Coherent, Lumentum, and Corning are entering the next stage through orders, strategic partnerships, and capacity expansion.

At the same time, traditional pluggable modules remain in strong growth, while NPO is still in an early adoption phase.

We therefore expect AI interconnect infrastructure over the coming years to feature:

continued growth in Pluggable + gradual NPO adoption + increasing deployment of CPO in core AI systems

rather than a single technology rapidly replacing the others.

For investment research, the key question should therefore evolve from:

“Who has CPO technology?”

toward:

“Who can convert the rising value content of new architectures into sustainable revenue, profit, and cash flow?”

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.