A recent 97-page in-depth report by Bernstein points out that copper interconnects and optical interconnects in AI data centers are not mutually exclusive, but will coexist for a long time in both vertical and horizontal scaling scenarios. While CPO technology has advantages in power consumption and cost, its widespread deployment faces obstacles due to manufacturing and maintenance challenges, and large-scale adoption is unlikely to be achieved before 2028. Therefore, optical interconnect LPO/NPO may become the leader during the transition period. However, CPO is fundamentally reshaping the value chain, shifting the profit center from traditional optical module suppliers to chip designers, advanced packaging companies, and system integrators.
It's worth mentioning Bernstein here. Bernstein (Sanford C. Bernstein) is a globally renowned investment research and asset management firm headquartered in the United States . Founded in 1967, it is currently part of the global asset management giant AllianceBernstein (AB). Bernstein is also one of the largest and oldest independent sell-side research firms. Below is a detailed breakdown of this Bernstein report.
In mid-February, we discussed in detail the underlying logic of the bottleneck transmission in the AI computing power industry chain, and mentioned that optical interconnect is one of the main AI themes that the market is switching to in 2025-2026.
The earliest mention of this was at the end of last year, when we began to truly focus on and research the field of optical interconnects.
Bernstein's report focuses on three main aspects:
Why is connectivity replacing computing power as the new bottleneck? Where is the CPO (Consumer Product Ownership) realization pace? Why is PCB/ABF substrate a more realistic direction for performance realization in 2026? Let's break it down in detail.
This report isn't really trying to say that "CPOs are about to explode," but rather:
The bottleneck in AI data centers is shifting from GPUs/HBM/CoWoS to "connectivity systems." Future investment will not focus solely on CPO (Consumer Product Owner) success, but rather on the coordinated upgrades of optics, electronics, copper, boards, packaging, and testing.
To put it more bluntly:
In the past, the market mainly looked at GPU computing power when it came to AI.
The market is now looking at how GPUs can be connected .
The future depends on whether the utilization rate of computing power can be released by the connected system .
This is what the report title refers to as "War for AI Data Center Connectivity" .
1. Why is "connectivity" becoming a new bottleneck for AI data centers?
AI clusters are not simply a matter of piling up GPUs. The real challenge lies in ensuring these GPUs can synchronize at high speed, exchange parameters, transfer activation values, perform AllReduce, and engage in model and data parallelism. No matter how powerful the theoretical computing power, if the communication between GPUs cannot keep up, the actual utilization rate will plummet.
You can think of an AI cluster as a giant factory:
Why is connectivity replacing computing power as the new bottleneck?
The root of this issue lies in the training method of large models. There are two parallel methods for training large models:
One approach is called tensor parallelism, and the other is called expert parallelism. Both methods share the characteristic of requiring frequent and large-scale data exchange between GPUs.
The amount of data exchanged between GPUs during a single training session is astronomical. What does this mean? In the past, you could simply increase the number of GPUs. Now, the more GPUs you add, the greater the overhead of communication between them. At a certain critical point, adding more GPUs no longer speeds up training; instead, it exacerbates communication congestion—this is the connectivity bottleneck.
Bernstein provided a comparison: in a standard NVIDIA GB30 rack, copper cables are used between GPUs because they are cheap and stable over short distances. However, fiber optic cables must be used between racks because copper cables suffer from signal attenuation beyond 2 meters. Optical modules are needed at both ends of the fiber optic cable to convert electrical signals into optical signals and then back.
The problem is that a 1.6T optical module consumes about 30 watts, and a large portion of that is consumed by a chip called a DSP (Digital Signal Processor). With hundreds of optical modules in a rack, the power consumption of optical communication alone cannot be reduced.
Therefore, the real problem facing AI data centers today is not that insufficient computing power has driven up power consumption. Nvidia itself says that its new generation of CPU switches can save 70% of power compared to traditional optical modules. For a 51.2T switch, this alone can save 500 watts, and the power saved can allow you to add more GPUs.
NVIDIA itself is also reinforcing this narrative. In March 2025, NVIDIA released Spectrum-X Photonics and Quantum-X silicon photonics switches, emphasizing that they are designed to enable AI factories to connect millions of GPUs and reduce power consumption and maintenance costs; NVIDIA claims that its photonics switches can achieve 1.6Tb/s per port, 3.5x energy efficiency improvement, 63x signal integrity improvement, and 10x network resilience improvement.
The underlying logic of Bernstein's report is that the next phase of AI capital expenditure is not just about buying more GPUs, but about buying more "connectivity that makes GPUs work effectively".
II. The report's core judgment: It's not "copper retreat and photovoltaic advancement," but rather "coexistence of multiple approaches."
There's a common saying in the market: "Light advances, copper retreats ."
However, this report offers a more nuanced perspective: copper and optical interconnects are not simple substitutes, but rather will coexist for a long time under different distances, bandwidths, maintenance requirements, and cost structures. Bernstein argues that copper and optical interconnects are not simple replacements, but will develop separately in scale-up and scale-out scenarios. This judgment is crucial.
1. Scale-up: Copper remains a strong option for rack-mount/near-distance interconnects.
Scale-up refers to high-speed interconnects between GPUs, between GPUs and switches, and within or near server racks. The most important aspect here is:
Low latency, low cost, high reliability, maintainability, and short-distance transmission capability.
In this scene, copper does not die immediately.
Huang has previously stated that NVIDIA will not use CPO for the main connection between flagship GPUs for the time being, because traditional copper connections are currently much more reliable than CPO optical connections; NVIDIA will first use CPO in two new network chips in top-of-the-line switches for servers.
This statement is very important. It means that CPO is a direction, but it will not immediately and completely replace copper.
In other words, at least for now, NVIDIA's logic is:
On the switch side, CPO can be implemented first, but more caution is needed on the GPU/XPU side.
The reason is simple: GPUs are the most expensive and critical asset in the system. You can't sacrifice reliability just because optical interconnects are energy-efficient. In an AI training cluster, frequent link failures result not only in hardware costs, but also in interrupted training tasks, decreased GPU utilization, and increased scheduling complexity.
2. Scale-out: Optical interconnection offers advantages between racks/clusters.
Scale-out is a larger-scale expansion of GPU clusters, typically involving longer east-west traffic between racks or within a data center.
In this scenario, the advantages of optical solutions are more obvious:
Longer distance, higher bandwidth, lighter cables, lower power consumption, and better wiring density.
Therefore, the future will not be one where "copper is completely replaced by light," but rather:
The most valuable aspect of Bernstein's report is that it doesn't stop at the level of "CPO concept stocks," but breaks down AI connectivity into multiple technological routes.
III. CPO: Direction is important, but 2026 is not the year of full-scale explosion.
The part of this report that is most easily misinterpreted by the market is the CPO.
Many people see CPO and immediately draw the following conclusion:
Optical modules are about to be replaced, CPOs will immediately surge, and traditional optical module manufacturers will be finished.
This understanding is too superficial.
Bernstein anticipates that small-scale deployments of CPO in scale-out networks may begin in the second half of 2026, primarily to validate real-world performance and supply chain maturity. However, in more critical scale-up scenarios, CPO adoption may be delayed until after the second half of 2028, as the industry needs to first validate the long-term reliability of switch-side CPO before applying it to higher-value, more fault-tolerant XPU systems.
This aligns with Jensen Huang's previous statement: CPO will first be used in network switching chips, rather than being directly applied to GPU main connections on a large scale.
Therefore, the rhythm of time should be understood as follows:
LightCounting also supports a "gradual evolution" rather than an "overnight switch." It predicts that traditional retimed pluggables will remain dominant for the next five years, although LPO/CPO will account for a significant proportion of 800G and 1.6T ports between 2026 and 2028. EDN's summary of industry perspectives also mentions that Yole believes large-scale CPO deployment may occur between 2028 and 2030, while LightCounting believes that optical modules will still account for the majority of data center optical links within this decade, but optical components will continue to move closer to ASICs.
Therefore, my judgment is:
CPO is a medium- to long-term trend, but the more certain revenue in 2026 may not be in the purest CPO concept stocks, but in the light source, testing, packaging, PCB, ABF, CCL, 1.6T optical modules and LPO/NPO that must be upgraded on the eve of CPO.
IV. LPO/NPO: They are the "transitional mainstay" before the CPO boom.
A key point of this report is that it does not simply divide the technology routes into "traditional optical modules vs. CPO".
There are also LPO and NPO in between.
1. What is LPO?
LPO stands for Linear Pluggable Optics. It can be roughly understood as: retaining the pluggable form factor, but removing or weakening the DSP, and using linear drive and host-side equalization to reduce power consumption.
The advantages are: lower power consumption, potentially lower cost, and still some maintainability.
The disadvantages are: system debugging is more difficult, the link budget is tighter, and the requirements for host-side SerDes and system engineering are higher.
The published abstract mentions that LPO can significantly reduce power consumption compared to traditional pluggable modules by eliminating the DSP and delegating signal processing to linear components, while retaining the convenience of modular maintenance; Bernstein even believes that LPO shipments may exceed CPO shipments by 2030.
2. What is an NPO?
NPO can be understood as Near-Packaged Optics, which means placing the optical engine closer to the ASIC, but not completely encapsulated like CPO .
Its value lies in compromise:
This suggests that the next few years are unlikely to be a "one-step path to CPO," but rather:
Traditional pluggable → LPO/NPO → CPO → Optical I/O / optical fabric
This is why you can't just look at CPOs in 2026. The companies that can truly deliver results are likely those that can supply across multiple stages.
In summary, the CPO story will not come true in 2026. CPO can only be shipped in small batches in the second half of 2026, and is only used for scale-out scenarios. In other words, large-scale deployment between racks will not happen until 2028.
Why so slow? Bernstein gave three reasons:
The first reason is that cloud service providers are unwilling to replace traditional optical modules. If a module fails, maintenance can simply remove it and replace it with a new one – a quick fix in minutes. The CPU, however, is soldered into the switch. If one optical engine fails, the entire switch has to be returned to the factory. The downtime and maintenance costs are a major problem for cloud service providers like Amazon, Google, and Microsoft. Furthermore, the failure rate of optical modules is not low; the industry standard is one failure per 100,000 hours. This translates to nine replacements per year for 10,000 optical modules – this is a hardware failure, not even considering software failures.
CPO integrates the optical engine into the chip, requiring a reliability improvement of several orders of magnitude to reassure cloud service providers. Bernstein stated directly that they communicated with InnoLight Technology, a Chinese optical module manufacturer, and InnoLight told them that no cloud service provider customer plans to deploy CPO on a large scale between 2026 and 2027. This statement carries significant weight, and the market may not have taken it to heart.
The second reason is that transitional solutions have emerged, and CPUs are no longer the only option. There are two technologies in between: LPO and NPO. LPO removes the most power-consuming DSP chip in the optical module and replaces it with a simpler component. This reduction cuts the power consumption to one-third of that of traditional optical modules, while retaining the pluggable 800G LPO, which is now in mass production.
NPO places the optical engine on the PCB next to the switch chip, but it is still removable. Nvidia's current CPU products are, strictly speaking, actually NPOs. These two transitional solutions can last for 2 to 3 years. Therefore, cloud service providers have every reason to say, "I'll use LPUs first, and wait until CPOs are truly mature."
The third reason is that in scale-up scenarios, copper cables are still viable; the connection between GPUs is called scale-up. Currently, no alternative can match the cost and reliability advantages of copper cables.
Bernstein explicitly stated that from 2026 to 2028, copper cabling will still dominate scale-up, and Luxshare Precision will benefit from this. Luxshare Precision is directly competing with Nvidia's GP300 copper connector and Amphenol. There is also a transitional technology called CPC co-packaged copper cabling, which further extends the lifespan of copper cabling.
Lightcounting, an industry consulting firm, predicts that copper cables will still account for nearly half of the 1.6T connection market by 2029.
V. The biggest impact of CPO: It's not simply about reducing costs, but about redistributing the profit pool.
The industrial significance of CPO is not just about energy saving, nor is it simply about replacing optical modules.
What it truly changes is: where profits are generated.
In the era of traditional pluggable optical modules, the value chain was roughly as follows:
DSP / Optical Chip / TOSA / ROSA / Module Packaging / Optical Module Manufacturer / Switch Manufacturer / Cloud Vendor.
The CPO era will become:
Switch ASIC / Optical Engine / External Laser Source / FAU / Advanced Packaging / Wafer Manufacturing / Testing / System Integration
Bernstein broke down the cost of the NVIDIA Quantum-X800 CPO switch: this switch is equipped with four switch ASICs, each integrating 18 optical engines, and has 18 external light source modules; the estimated cost of a single Quantum-X800 CPO switch is approximately $570,000. The abstract also points out that in the CPO architecture, the DSP is eliminated, the optical engines are co-packaged with the switch chip, and the value center shifts to chip design, advanced packaging, and wafer manufacturing.
This is why the report will be favorable to these areas:
Comparatively speaking, traditional optical module manufacturers face a problem:
If value shifts from module packaging to ASICs, packaging, optical engines, and system integration, their profit pools may be restructured.
However, this doesn't mean traditional optical module manufacturers will immediately become worthless. Because from 2026 to 2028, there will still be significant demand for 800G, 1.6T, and LPO/NPO. Cignal AI also points out that high-speed datacom modules, especially 800GbE and emerging 1.6TbE designs, will remain the main growth engine in 2026.
Therefore, the correct understanding is:
CPO will change the profit distribution in the optical module industry chain, but will not immediately eliminate pluggable optical modules in 2026.
VI. Why does the report emphasize that PCB, ABF, and CCL are more realistic directions for 2026?
This is what I think is most worthy of your attention.
CPO has great potential, but its realization cycle is relatively long. In contrast, upgrades to PCB, ABF, and CCL are closer to current orders.
The reason is that even though CPO has not yet been commercially deployed on a large scale, AI servers and switches are already being upgraded.
Rubin, Rubin Ultra, GB300, cloud vendor ASICs, and next-generation switch ASICs are all improving:
Single board speed, package area, power supply density, signal integrity requirements, heat dissipation requirements, and low material loss requirements.
This is the most counterintuitive point in the report, yet it's the easiest to overlook. The real money-makers in 2026 will be in the established sectors of PCB, HDI, ABF, and substrates.
Why is this considered contrary to consensus? Because this industry is too traditional. PCB is a decades-old industry, with a global market of $85 billion by 2025. It doesn't sound sexy at all. Everyone is focused on CPO, optical modules, and Nvidia. Nobody is willing to spend time researching printed circuit boards. But Bernstein's data tells us that this industry has already quietly taken off in 2025.
Bernstein provided some figures: Sheng Hong Technology, which manufactures HDI high-density interconnect boards, saw a 63% year-on-year revenue increase in 2025. WUS Electric's revenue from NVIDIA's GB300MPCB increased by 45%. Gold Circuit's annual supply to AWS Trinium increased by 40%, and another supplier in Shengyi Electronics' AWS supply chain saw a 40% increase. These are real, actual results, not expectations. Why is this sector rising? There are three dimensions to consider:
The first layer is that AI servers have doubled the amount of PCB content. Previously, an NVIDIA H10 server with 80 GPUs and an HDI (High-Intensity Distributor) PCB had a total value of about $100 to $150 per GPU. Moving to a GB200VL72 rack, that figure jumps to $300 per PCB. What does this mean? For the same GPU, PCB manufacturers earn twice as much money.
And that's not all. The upcoming Vera Robin platform will adopt a new structure called midplane, replacing the original copper cable connections with multi-layer PCBs. This midplane is a 44-layer board, using the highest-end M8 grade copper-clad laminate. The next-generation Rubin Ultra may use a 78-layer M9 grade. Doubling the number of layers, upgrading the materials, and doubling the value again.
The second bottleneck is the supply of upstream materials. ABF substrates contain a key material called T-glass, a low coefficient of thermal expansion glass fiber. Its function is to prevent the AI chip from deforming at high temperatures, which could lead to solder joint failure.
Currently, only one company globally can produce T-glass at the top-tier specifications: Nittobo. Their CTE value is 2.8%, a level unmatched by other manufacturers. Nittobo's new production capacity won't be operational until the end of 2026, with shipments not expected until 2027. This means a continued shortage of T-glass throughout 2026.
What is the T-glass shortage? It means that ABF substrate manufacturers can legitimately raise prices. Unimicron Emerging Electronics has already renegotiated prices with its customers. Bernstein's model predicts that the ASP of ABF substrates will increase by 5% to 7% quarter-on-quarter in 2026, with a cumulative annual increase potentially exceeding 20%.
The third layer is the hidden monopolist of ABF film. ABF film is one of the core materials of ABF substrates. The inventor of this material is Agenomoto, the Japanese food company that sells MSG. In the 1990s, during their research and development of MSG, they accidentally discovered a special amino acid-derived thin film that could be used as a thermal expansion layer for semiconductor substrates. Since then, 95% of the ABF films worldwide have come from Agenomoto.
According to Bernstein's data, Ajinomoto's ABF (Alternative Flavor) business has a gross margin of 60%, a growth rate of 32% in fiscal year 2026, and is projected to accelerate to 45% in fiscal year 2027. This company's ABF business has been unchallenged for 30 years.
Therefore, what is more certain in 2026 is not a "CPO boom overnight," but rather:
High-speed PCBs need to be upgraded; ABF substrates need to be upgraded; CCLs need to be upgraded to lower loss materials; copper foil, fiberglass cloth, and low Dk/low Df materials need to be upgraded; testing and verification processes need to be upgraded.
Therefore, a more realistic strategy for 2026 is to focus on three types of certainties : the optical demand brought about by the transition from 1.6T to LPO/NPO, the PCB/ABF/CCL upgrades brought about by Rubin/ASIC, and the testing/FAU/light source/advanced packaging that must be invested in before CPO trial production.
Because the capital market often makes a mistake:
They like to buy the most distant concepts, but what truly produces results first is often the "infrastructure that must be built before the long-term concept."
CPO is like the high-speed rail station of the future.
However, before the high-speed rail station is fully operational, the companies that will likely make money first are those involved in road construction, track laying, power supply, signaling systems, and testing equipment.
VII. The order in which the industry chain benefits from this report
If we divide the AI-connected industry chain into four layers:
Tier 1: The Strongest Platform-Level Winner
These companies don't just sell a single component; they sell the entire control architecture.
NVIDIA
NVIDIA's advantage isn't just in GPUs, but rather in its GPU + NVLink + InfiniBand + Ethernet + Spectrum-X + Quantum-X + software ecosystem. NVIDIA's officially disclosed silicon photonics networking switches already include TSMC, Coherent, Corning, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo Electric, TFC Communication, and others in its ecosystem.
This shows that NVIDIA is doing one thing:
It's not just about selling GPUs; it's about bringing the network architecture of AI factories under its platform control.
TSMC is the invisible linchpin of this whole story.
The CPO platform combines electronic chips and photonic chips using hybrid integration technology. All major customers, including Nvidia, Broadcom, and AI Labs, are migrating to TSMC. While TSMC doesn't profit much from the CPO itself, it strengthens TSMC's dominance in advanced packaging and wafer foundry.
Broadcom
Broadcom's logic is different. It's more like:
Ethernet switch ASIC + custom ASIC + CPO + cloud vendor customized chip ecosystem.
In October 2025, Broadcom announced the Tomahawk 6 Davisson, its third-generation CPO Ethernet switch with a switching capacity of 102.4Tbps, and stated that it is already shipping. Broadcom claims that by integrating the TSMC COUPE optical engine and advanced multi-chip packaging, it reduces optical interconnect power consumption by 70%, while supporting scale-up of 512 XPUs and 100,000+ XPUs in two-layer networks.
This demonstrates that TSMC and Broadcom are crucial companies in the AI networking and CPO value chain, besides NVIDIA.
The second layer: deterministic optics and high-speed interconnects
This includes:
1.6T optical module, LPO/NPO, silicon photonics, laser, external light source, FAU, optical connector.
Representative companies in this sector include Coherent, Lumentum, Fabrinet, Innolight, Eoptolink, SENKO, Corning, and Sumitomo. NVIDIA's official ecosystem list includes numerous companies in optics, packaging, and connectivity.
The focus at this level is not "who is most like a CPO", but rather:
Who can simultaneously meet the demands of 800G/1.6T, LPO/NPO, CPO trial production, external light source, and FAU?
Companies that can navigate multiple stages have a higher success rate than companies with only a single concept.
Third layer: PCB, ABF, CCL, materials
This is the area most likely to be underestimated in 2026.
The public account mentions that the original report covered or mentioned companies such as Chroma, Luxshare, Unimicron, NVIDIA, Broadcom, TSMC, and Ibiden.
Companies like Unimicron and Ibiden, which are part of the substrate/PCB supply chain, are particularly noteworthy because as AI servers become more complex, PCBs and packaging substrates are no longer just follower components, but rather the performance constraints themselves.
Fourth layer: Testing equipment, yield, reliability
The biggest challenge for a CPO is not the PowerPoint presentation, but mass production.
Mass production needs to address:
Optocoupler yield;
Stability of external laser source;
Reliability in high-temperature environments;
Encapsulation stress;
On-site maintenance;
Test time;
consistency;
Repair mode after failure.
Therefore, testing equipment and reliability verification can be a good "shovel seller".
These types of companies may not be the most attractive, but if the CPO enters the trial production phase, they are often the first to see orders.
8. Investment Implications of This Report: Don't buy the "most concept-driven" stocks; buy the "most unavoidable" ones.
The biggest takeaway from this report for investment is:
AI connectivity is not a single-point technological revolution, but rather a shift in bottlenecks. Investment should focus on common bottlenecks, not on a single path.
What is a common bottleneck?
This is something that can't be avoided regardless of whether the final product is a CPO, LPO, NPO, or a continued upgrade from traditional pluggable devices. For example:
Conversely, a single route carries greater risk.
For example, if you only buy "pure CPO concept" stocks, the risks are:
CPO mass production is delayed, orders are not fulfilled, and valuations are already being driven down.
The risks of only buying traditional optical modules are:
CPO/NPO/LPO are restructuring the value chain, with long-term profits being taken by platform manufacturers and chip/packaging manufacturers.
The risks of only buying PCBs/materials are:
Customers expanded production too quickly, supply was released in a concentrated manner, and gross profit margins reversed.
Therefore, a better combination is:
Buy certainty in 2026, buy order flexibility in 2027, and buy structural options after 2028.
IX. My personal assessment of the reasonableness of this report
Very reasonable
- First, shifting the AI bottleneck from GPUs to connected systems is a very correct direction. Product releases from NVIDIA and Broadcom are validating this.
- Second, it is crucial to oppose the simplistic narrative of "copper going backwards and optical technology advancing." Reuters ' report on Jensen Huang has clearly stated that copper still holds a reliability advantage in GPU/XPU core interconnects in the short term.
- Third, the assessment that CPO is the right direction but large-scale deployment requires reliability verification is also reasonable. The industry assessments of LightCounting and Yole/EDN both lean towards "gradual migration rather than immediate, comprehensive replacement."
- Fourth, it emphasizes that "front-end processes" such as PCB/ABF/CCL, testing, and light sources are more likely to deliver results by 2026 , which is more helpful for investment. This is because the capital market tends to over-trade on the most distant future while underestimating the near-term processes that actually bring in orders.
Points to note
First, publicly relaying Bernstein's views might turn them into "investment-related" or sensationalist headlines. For example, the statement "The real battleground for AI is not in chips, but in connectivity" has viral potential, but strictly speaking, GPUs/HBM/CoWoS are still the core bottlenecks. It's just that the marginal importance of connectivity is increasing, not that chips are unimportant.
Second, while the value transfer direction of CPO is correct, the speed may be overestimated by the market. CPO needs to solve problems such as manufacturing, packaging, field maintenance, failure replacement, and reliability; it is not a technology that can be mass-produced immediately after a press conference.
Third, while LPO/NPO offers significant transitional value, its system debugging is also quite challenging. LPO is not simply a "low-power version of pluggable"; it shifts much of the complexity to the host side and system-level debugging.
Fourth, while the PCB/ABF/CCL line offers strong certainty, we must be wary of the capacity expansion cycle. Once the materials and substrate industries see a period of high prosperity, they easily expand production; however, if customer platforms slow down later, gross margins will suffer.
10. This timeline can be used to track progress over the next 2-3 years.
2026: Don't just look at the CPO, look at three certainties.
The focus in 2026 will not be a surge in CPOs, but rather:
Is the production volume of 1.6T pluggable optical modules being increased?
Do LPO/NPO obtain more cloud vendor/switch platform certifications?
Will PCB/ABF/CCL prices continue to rise or production capacity be expanded?
Have there been any actual orders for CPO-related testing equipment, FAU, and external light sources?
If these events occur, it indicates that the report's logic has entered the realization phase.
2027: Witnessing the CPO Pilot Program's Journey from "Prototype" to "Customer Deployment"
The key metrics are:
Real customer deployments of NVIDIA Quantum-X / Spectrum-X Photonics;
Broadcom Davisson/Tomahawk CPO's customer expansion;
Whether CoreWeave, Lambda, Meta, Google, Microsoft, Amazon, etc. adopt it;
Should CPO external light source, FAU, and testing equipment be included in revenue recognition?
After 2028: Check if the CPO has entered scale-up mode.
The most crucial turning point is:
Does the CPO move from the switch side to the vicinity of the XPU/GPU?
Whether optical I/O is packaged in high-end ASIC/GPU packages;
Does the OCS/optical fabric begin to change the data center network topology?
If it reaches this stage, CPO will not just be a replacement for optical modules, but a change in AI computing architecture.
XI. Investment Framework Based on This Report: Four Asset Classes, Four Logics
If I were to use this report to guide my investments in US stocks, Hong Kong stocks, and A-shares, I would categorize them into four types.
My personal favorite strategy is:
The core position is to buy the platform winner, the flexible position is to buy optics and PCB for certainty, and the option position is to buy CPO forward direction with a small proportion.
It is not advisable to put all your funds into "purest CPO concept stocks" right away.
12. The five most important points of this report
- First, the bottleneck of AI data centers is shifting from "fast computing" to "fast, stable, and energy-efficient connectivity".
- Second, light will not immediately destroy copper, nor will copper remain in all scenarios forever; different distances and system levels will choose different solutions.
- Third, CPO is a direction, but more realistic revenue in 2026 will come from 1.6T, LPO/NPO, light source, testing, PCB, ABF, and CCL.
- Fourth, the real impact of CPO is not to make optical modules cheaper, but to shift the profit pool from traditional module packaging to chips, packaging, optical engines, light sources, testing, and system platforms.
- Fifth, when investing in AI connectivity, don't buy the hottest concepts; buy the bottlenecks that are hardest to overcome.
- This is a valuable report on "AI Layer 2 Infrastructure." It reminds the market that after GPUs, the next thing to be repriced is not a single component, but the entire AI connectivity stack.
However, it cannot be simply read as "CPO immediately erupts." A more accurate way to read it is:
Looking ahead to 2026: Pluggable/LPO/NPO/PCB/ABF/Testing;
Looking at CPO pilot orders in 2027;
We'll see after 2028 whether CPO and optical I/O truly become part of the core AI computing architecture.
