Bernstein’s newly released 97-page in-depth report indicates that copper interconnects and optical interconnects in AI data centers are not mutually exclusive but will coexist long-term in both scale-up and scale-out scenarios. Although CPO technology offers advantages in power consumption and cost, its widespread deployment faces obstacles due to manufacturing and maintenance challenges, making large-scale adoption unlikely before 2028; thus, optical interconnects such as LPO/NPO may lead during the transition period. However, CPO is fundamentally reshaping the value chain, shifting profit centers from traditional optical module suppliers to chip designers, advanced packaging providers, and system integrators.
Special mention should be made of Bernstein, a globally renowned investment research and asset management firm headquartered in the United States. Founded in 1967, Bernstein is now part of AllianceBernstein (abbreviated as AB), a global giant in asset management, and is one of the largest and oldest independent sell-side research firms. Below is a detailed breakdown of Bernstein’s report.
In February, we conducted a detailed breakdown of the underlying logic behind bottlenecks in the AI computing power industry chain and discussed how optical interconnect is one of the key AI trends transitioning in the market over 2025-2026.
I only began truly focusing on and researching optical interconnects at the end of last year, as seen here: https://x.com/qinbafrank/status/2015377625167089671?s=20
In Bernstein's report, the core focuses on three aspects:
Why has connectivity replaced computing power as the new bottleneck? Where is the timeline for CPO realization? Why are PCB/ABF substrates a more realistic performance realization direction for 2026? A detailed breakdown.
What this report really wants to convey is not that "CPO is about to explode," but rather:
The bottleneck in AI data centers is continuing to shift from GPU/HBM/CoWoS toward the "interconnection system." The future investment focus will not be on CPO alone, but on the coordinated advancement of optics, electronics, copper, substrates, packaging, and testing.
Put more simply:
In the past, the market primarily focused on GPU computing power when evaluating AI.
The market is now beginning to focus on how GPUs are connected together.
The future will depend on whether the connectivity of systems can unlock computational utilization.
This is what is referred to in the report title as the "War for AI Data Center Connectivity."
Why is "connection" becoming the new bottleneck for AI data centers?
An AI cluster isn't just about stacking GPUs together. The real challenge is ensuring these GPUs can synchronize at high speed, exchange parameters, transfer activation values, perform AllReduce operations, and handle model and data parallelism. Even with theoretical computing power, if communication between GPUs can't keep up, actual utilization will drop.
You can think of an AI cluster as a massive factory:
Why is connectivity replacing computational power as the new bottleneck?
The root of this issue lies in the training methods of large models. There are two parallel approaches to training large models:
One is called tensor parallelism, and the other is called expert parallelism. Both methods require frequent and large-scale data exchange between GPUs.
The amount of data that must be exchanged between GPUs during a single training round is astronomical. What does this mean? In the past, you could simply add more GPUs to speed up training. Now, the more GPUs you add, the greater the communication overhead between them becomes. At a certain critical point, adding more GPUs no longer speeds up training—instead, it worsens communication congestion, creating a connectivity bottleneck.
Bernstein provided a comparison: inside a standard NVIDIA GB30 rack, GPUs are connected using copper cables because the short distances make copper cost-effective and stable. However, between racks, fiber optic cables are required, as copper cables suffer unacceptable signal degradation beyond two meters. Optical modules are needed at both ends of the fiber to convert electrical signals into optical signals and back again.
Here’s the issue: a 1.6T optical module consumes about 30 watts, with more than half of that power being consumed by a chip called the DSP (Digital Signal Processor). With hundreds of optical modules in a single rack, the power consumption for optical communication alone remains difficult to reduce.
So the real issue AI data centers face today isn't insufficient computing power pushing power consumption to its limit. NVIDIA itself says its new-generation CPU switches can save 70% more power compared to traditional optical modules; a single 51.2T switch can save 500 watts just on this alone—power savings that allow you to add more GPUs.
NVIDIA is also reinforcing this narrative. In March 2025, NVIDIA launched Spectrum-X Photonics and Quantum-X silicon photonics switches, emphasizing that they are designed to connect millions of GPUs in AI factories while reducing energy consumption and operational costs; NVIDIA claims its photonics switches enable 1.6 Tb/s per port, a 3.5x improvement in energy efficiency, a 63x improvement in signal integrity, and a 10x improvement in network resilience.
The underlying logic of Bernstein's report is that the next phase of AI capital spending will not just involve buying more GPUs, but also investing in more "connectivity to make GPUs work effectively."
II. The Core Insight of the Report: It is not "copper declines, light advances," but rather "multiple pathways coexist."
A common saying in the market is: "Gold rises when copper falls."
However, this report offers a more nuanced perspective: copper and optical interconnects are not simple substitutes, but will coexist long-term under different distances, bandwidth requirements, maintenance needs, and cost structures. Bernstein believes that copper and optical interconnects will evolve separately in scale-up and scale-out scenarios. This assessment is critical.
1. Scale-up: Within-rack or close-proximity interconnects, copper remains strong
Scale-up is closer to high-speed interconnects between GPU and GPU, GPU and switch, within a rack or near-rack range. Here, the most important factors are:
Low latency, low cost, high reliability, maintainability, and short-distance transmission capability.
In this scenario, copper did not die immediately.
Huang previously made it clear: NVIDIA will not initially use CPO for the primary connections between its flagship GPUs, as traditional copper connections are currently far more reliable than CPO optical connections; instead, NVIDIA will first implement CPO in two new network chips for top-of-rack switches.
This statement is very important. It indicates that CPO is the direction, but not an immediate full replacement for copper.
In other words, at least for now, NVIDIA's logic is:
The switch side can adopt CPO first, while the GPU/XPU side requires greater caution.
The reason is simple: GPUs are the most expensive and critical assets in the system. You cannot sacrifice reliability for the sake of power savings in optical interconnects. In an AI training cluster, a frequently failing link doesn't just cost hardware—it leads to interrupted training tasks, reduced GPU utilization, and increased scheduling complexity.
2. Scale-out: Optical solutions offer greater advantages for interconnection between cabinets or clusters.
Scale-out refers to expanding GPU clusters over a broader range, typically involving longer-distance east-west traffic between racks within a data center.
In this scenario, the advantages of the optical solution are more pronounced:
Longer distances, higher bandwidth, lighter cables, lower power consumption, and better cable density.
So the future is not "copper being completely replaced by light," but rather:
The most valuable aspect of Bernstein's report: it goes beyond the "CPO concept stock" level and breaks down the AI connection into multiple technological pathways.
III. CPO: Direction matters, but 2026 is not the year of full-scale breakthrough
The most easily misunderstood part of this report is the CPO.
Many people see CPO and immediately conclude:
Optical modules are being replaced; CPO is immediately surging, and traditional optical module manufacturers are finished.
This understanding is too crude.
Bernstein expects small-scale CPO deployments in scale-out networks to begin in the second half of 2026, primarily to validate real-world performance and supply chain maturity; however, CPO adoption in more critical scale-up scenarios may be delayed until after the second half of 2028, as the industry needs to first verify the long-term reliability of CPO on switches before applying it to higher-value, less tolerant XPU systems.
This aligns with Jensen Huang’s previous statements: CPO will first be used in network switching chips rather than being directly and大规模ly adopted for GPU main connections.
Therefore, the timing should be understood as follows:
LightCounting’s perspective also supports a gradual evolution rather than an overnight switch. It predicts that traditional retimed pluggables will remain dominant over the next five years, although LPO/CPO will account for a significant portion of 800G and 1.6T ports between 2026 and 2028. EDN’s summary of industry views also notes that Yole believes large-scale CPO deployment may occur between 2028 and 2030, while LightCounting expects optical modules to still constitute the majority of data center optical links within this decade, though optical components will continue to move closer to ASICs.
So my judgment is:
CPO represents a medium- to long-term direction, but more certain revenues in 2026 may not come from the purest CPO-related stocks, but rather from the upstream components that must be upgraded before CPO adoption—such as light sources, testing, packaging, PCBs, ABF, CCL, 1.6T optical modules, and LPO/NPO.
IV. LPO/NPO: They are the "transitional main themes" before CPO explosion
An important point of this report is that it does not oversimplify the technology roadmap as "traditional optical modules vs. CPO."
There are also LPO and NPO in between.
1. What is LPO?
LPO, short for Linear Pluggable Optics, can be understood as retaining the pluggable form factor while removing or reducing the DSP, using linear driving and host-side equalization to lower power consumption.
Advantages include lower power consumption, potentially lower costs, and retained maintainability.
The drawbacks are: more difficult system debugging, tighter link budget, and higher requirements for the host-side SerDes and system engineering.
The public summary notes that LPO significantly reduces 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 by 2030, LPO shipments could surpass those of CPO.
2. What is an NPO?
NPO can be understood as Near-Packaged Optics, which means placing the optical engine closer to the ASIC without fully integrating it as in CPO.
Its value lies in the compromise:
This suggests that in the coming years, it is unlikely to be a direct leap to CPO, but rather:
Traditional pluggable → LPO/NPO → CPO → Optical I/O / optical fabric
This is why in 2026, you can't just look at CPO. The companies that may truly deliver on their performance are those capable of supplying across multiple stages.
In summary, the CPO story won't materialize until 2026; in the second half of 2026, CPO will only be available in small quantities and limited to scale-out scenarios, with widespread deployment across racks not expected until 2028.
Why is it so slow? Bernstein gave three reasons:
The first reason is that cloud service providers prefer not to replace traditional optical modules because when they fail, operations staff can simply unplug and swap them out in minutes. In contrast, CPUs are soldered directly into the switch; if a single optical engine fails, the entire switch must be returned to the factory, resulting in significant downtime and operational costs for cloud providers like Amazon, Google, and Microsoft. Moreover, optical modules have a relatively high failure rate—the industry standard is one failure per 100,000 hours, which translates to approximately nine replacements per year for every 10,000 modules. This accounts only for hard failures, not soft failures.
Integrating the optical engine into the chip requires reliability improvements by several orders of magnitude to reassure cloud service providers. Bernstein explicitly stated that during discussions with Chinese optical module manufacturer InnoLight, InnoLight informed them that no cloud service provider customer plans to deploy CPO at scale between 2026 and 2027. This statement carries significant weight, though the market may not yet have fully absorbed it.
The second reason is that transitional solutions are now available, and CPUs are no longer the only option. Two technologies lie in between: LPO and NPO. LPO removes the most power-hungry DSP chip from the optical module and replaces it with simpler components. This change reduces power consumption to one-third of that of traditional optical modules, while still maintaining pluggable 800G capability—LPO is already in mass production.
NPO places the optical engine on the PCB next to the switch chip, but it remains removable. NVIDIA’s current products referred to as CPUs are, strictly speaking, transitional solutions like NPO that can last two to three years. Therefore, cloud service providers have every reason to say they’ll use LPU for now and wait until CPO becomes truly mature.
The third reason is that in scale-up scenarios, copper cables are still dominant. The connections between GPUs are referred to as scale-up. Currently, no alternative can match the cost and reliability advantages of copper cables.
Bernstein explicitly stated that from 2026 to 2028, scaling up will still be dominated by copper cables, and Luxshare Precision stands to benefit. The company is directly competing with Amphenol on copper cable connectors for NVIDIA’s GP300, and there is also a transitional technology called CPC (co-packaged copper cable) that further extends the lifecycle of copper cables.
Industry research firm LightCounting predicts that by 2029, copper cables will still account for nearly half of the 1.6T connectivity market.
V. CPO's Greatest Impact: Not Simply Reducing Costs, But Reallocating the Profit Pool
The industrial significance of CPO goes beyond energy savings and merely replacing optical modules.
What it truly changes is: where the profits come from.
In the era of traditional pluggable optical modules, the value chain was roughly:
DSP / Photonic Chip / TOSA/ROSA / Module Packaging / Optical Module Manufacturer / Switch Manufacturer / Cloud Provider
The CPO era will become:
Switch ASIC / Optical Engine / External Laser Source / FAU / Advanced Packaging / Wafer Manufacturing / Testing / System Integration
Bernstein conducted a cost breakdown of the NVIDIA Quantum-X800 CPO switch: the switch is configured with four switch ASICs, each integrating 18 optical engines and 18 external light source modules; the estimated cost of a single Quantum-X800 CPO switch is approximately $570,000. The summary also notes that under the CPO architecture, the DSP is eliminated, with optical engines and the switching chip co-packaged, shifting the value center toward chip design, advanced packaging, and wafer manufacturing.
This is why the report will be bullish for these areas:
Relatively speaking, traditional optical module manufacturers face a challenge:
If value shifts from module packaging to ASICs, packaging, optical engines, and system integration, their profit pools may be restructured.
But this does not mean traditional optical module manufacturers will immediately lose value, as there will still be significant demand for 800G, 1.6T, LPO/NPO solutions between 2026 and 2028. Cignal AI also notes that high-speed datacom modules, particularly 800GbE and emerging 1.6TbE designs, will remain the primary growth drivers in 2026.
So the correct understanding is:
CPO will alter profit distribution across the optical module industry chain, but it will not eliminate pluggable optical modules immediately by 2026.
Six: Why does the report emphasize PCB, ABF, and CCL as more realistic directions for 2026?
This is the area I believe deserves your closest attention.
CPO has significant growth potential, but its realization timeline is further out. In comparison, upgrades to PCB, ABF, and CCL are more closely aligned with current orders.
The reason is that AI servers and switches are already being upgraded, even though CPO has not yet been commercially deployed at scale.
Rubin, Rubin Ultra, GB300, cloud provider ASICs, and next-generation switch ASICs are all improving:
Board data rate, package area, power density, signal integrity requirements, thermal dissipation requirements, and low-loss material requirements.
This is the most contrarian yet most overlooked point in this report: the real money in 2026 will be made in the traditional sectors of PCB, HDI, ABF, and substrates.
Why call it "anti-consensus"? Because this sector is extremely traditional. PCBs are a decades-old industry, with a global market size of $85 billion by 2025—sounds anything but exciting. Everyone is focused on CPOs, optical modules, and NVIDIA, with no one willing to spend time studying printed circuit boards. But Bernstein’s data tells us this sector has quietly taken off by 2025.
Bernstein provided a set of figures: Shenghong Technology, which produces HDI high-density interconnect boards, achieved a 63% year-over-year revenue growth in 2025. WUS (Jiangsu Huadian Electronics) saw a 45% revenue increase from supplying NVIDIA’s GB300 MPCB. Gold Circuit (Jinxiang Electric) experienced a 40% year-over-year growth in annual supply to AWS Trinium, and Shengyi Electronic, another supplier in AWS’s supply chain, also saw a 40% growth. These are actual, realized performance figures—not projections—already delivered. Why is this sector rising? There are three dimensions to consider:
The first layer is that the PCB content in AI servers has doubled. Previously, for NVIDIA H10 servers, the total value of the PCB per GPU was around $100 to $150. With the GB200 NVL72 rack, this figure has directly increased to $300 per GPU. What does this mean? For the same GPU sold, PCB manufacturers now earn twice as much.
And that’s not all—the upcoming Vera Robin platform will adopt a new architecture called the midplane, replacing copper cable connections with multilayer PCBs. This midplane is a 44-layer board made with the highest-grade M8 copper-clad laminate, and the next-generation Rubin Ultra may use a 78-layer M9-grade board. With double the layers and upgraded materials, the value proposition doubles again.
The second layer involves a bottleneck in upstream materials. A key material in ABF substrates is T-glass, a low-coefficient-of-thermal-expansion glass fiber, which prevents substrate deformation and solder joint failure in AI chips under high temperatures.
Currently, only one company worldwide can achieve the highest specifications for T-glass: Nittobo, with a CTE value of 2.8%. No other manufacturers can reach this level. Nittobo’s new production capacity won’t come online until the end of 2026, with official shipments expected in 2027, meaning T-glass will remain in short supply throughout 2026.
What is the T-glass shortage? It allows ABF substrate manufacturers to legitimately raise prices. Unimicron Advanced Electronics has already renegotiated prices with its customers. Bernstein’s model predicts that the ASP of ABF substrates will increase quarter-over-quarter by 5% to 7% in 2026, with a potential cumulative annual increase exceeding 20%.
The third layer is the invisible monopolist of ABF films. ABF film is one of the core materials in ABF substrates, and this material was invented by Agenomoto, a Japanese food company known for selling monosodium glutamate. During their research on MSG in the 1990s, they accidentally discovered a unique amino acid-derived film suitable as a thermal expansion layer for semiconductor substrates. Since then, 95% of global ABF films have come from Agenomoto.
According to Bernstein's data, Ajinomoto's ABF business has a gross margin of 60%, grew at a rate of 32% in fiscal year 12026, and is expected to accelerate to 45% by fiscal year 2027. This company's ABF business has remained untouchable for 30 years.
So what's more certain in 2026 is not "CPO exploding overnight," but:
High-speed PCBs need to be upgraded; ABF substrates need to be upgraded; CCL must be upgraded to lower-loss materials; copper foil, glass fiber fabric, and low-Dk/low-Df materials need to be upgraded; testing and validation processes need to be upgraded.
Therefore, a more realistic strategy for 2026 is to first focus on three areas of certainty: optical demand driven by 1.6T and the LPO/NPO transition, PCB/ABF/CCL upgrades brought by Rubin/ASIC, and the necessary investments in testing/FAU/light sources/advanced packaging required before CPO pilot production.
Because capital markets often make one mistake:
Love to buy the most distant concepts, but the ones that actually deliver results first are often the infrastructure that must be built before those distant concepts can become reality.
CPO is like the high-speed rail station of the future.
Before the high-speed rail station operates fully, the companies that may profit first are those involved in road construction, track laying, power supply, signaling systems, and inspection equipment.
Seven: The Order of Benefit to the Industry Chain in This Report
If we divide the AI industrial chain into four layers:
Layer 1: The Strongest Platform-Level Winners
These companies don't just sell a single component—they control the architecture.
NVIDIA
NVIDIA’s advantage is not just its GPUs, but the combination of GPUs, NVLink, InfiniBand, Ethernet, Spectrum-X, Quantum-X, and its software ecosystem. NVIDIA’s officially disclosed silicon photonics networking switches have already integrated TSMC, Coherent, Corning, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo Electric, and TFC Communication into its ecosystem.
This indicates that NVIDIA is doing something:
Not only selling GPUs, but also bringing the network architecture of AI factories under their own platform control.
TSMC is the invisible hub of this entire story.
The COP platform combines electronic chips and photonic chips using hybrid bonding technology. All major clients—NVIDIA, Broadcom, and AI labs—are migrating to TSMC. This company doesn't make substantial profits directly from CPO, but CPO strengthens TSMC's dominance in advanced packaging and wafer foundry services.
Broadcom
Broadcom's logic is different. It's more like:
Ethernet switch ASIC + custom ASIC + CPO + cloud provider customized chip ecosystem.
In October 2025, Broadcom announced the Tomahawk 6 Davisson, its third-generation CPO Ethernet switch, featuring a 102.4 Tbps switching capacity and stating that it is already in shipping; Broadcom claims that by integrating TSMC’s COUPE optical engine and advanced multi-chip packaging, it reduces optical interconnect power consumption by 70%, while supporting scale-up to 512 XPUs and over 100,000 XPUs across a two-layer network.
This indicates that TSMC and Broadcom are also critical companies in the AI network and CPO value chain, aside from NVIDIA.
Layer 2: Optical and high-speed interconnects with higher certainty
This includes:
1.6T optical modules, LPO/NPO, silicon photonics, lasers, external light sources, FAU, optical connectors.
Representative companies include Coherent, Lumentum, Fabrinet, Innolight, Eoptolink, SENKO, Corning, and Sumitomo. NVIDIA’s official ecosystem list includes multiple companies related to optics, packaging, and connectivity.
The focus of this layer is not "who looks most like a CPO," but rather:
Who can simultaneously meet the requirements for 800G/1.6T, LPO/NPO, CPO pilot production, external light source, and FAU?
Companies that span multiple stages have a higher success rate than single-concept companies.
Layer 3: PCB, ABF, CCL, Materials
This is the place most likely to be underestimated in 2026.
The public paraphrase mentioned that the original report covered or referenced companies such as Chroma, Luxshare, Unimicron, NVIDIA, Broadcom, TSMC, and Ibiden.
Companies in the substrate/PCB supply chain, such as Unimicron and Ibiden, are particularly noteworthy, as the increasing complexity of AI servers has transformed PCBs and packaging substrates from passive components into critical performance constraints themselves.
Layer 4: Test Equipment, Yield, and Reliability
The biggest challenge for CPO is not the PPT, but mass production.
Mass production needs to address:
Optocoupler yield;
Stability of the external laser source;
Reliability in high-temperature environments;
Packaging stress;
On-site maintenance;
Test time;
Consistency;
Repair mode after expiration.
Therefore, testing equipment and reliability verification could be excellent "sellers of shovels."
These companies may not be the most glamorous, but if the CPO enters pilot production, they are often the first to see orders.
Eight: Investment Implications of This Report: Don't buy what's "most like a concept"; buy what's "hardest to avoid."
The biggest insight this report offers for investors is:
AI connectivity is not a single-point technological revolution, but a shift in bottlenecks. Invest in shared bottlenecks, not just one path.
What is a common bottleneck?
Regardless of whether the final outcome is CPO, LPO, NPO, or continued upgrades to traditional pluggables, this is something that cannot be avoided. For example:
In contrast, single-path risk comparison
For example, if you only invest in "pure CPO concepts," the risk is:
The CPO mass production timeline has been delayed, orders are not being fulfilled, and the valuation has been slashed.
The risk of buying only traditional optical modules is:
CPO/NPO/LPO restructure the value chain, with the long-term profit pool captured by platform manufacturers and chip/packaging manufacturers.
Buying only PCBs/materials carries the risk of:
The customer expanded production too rapidly, supply was released in a concentrated manner, and gross profit margins reversed.
So a better combination is:
Buy certainty in 2026, buy order elasticity in 2027, and buy architectural options after 2028.
Nine: Personal Evaluation of the Reasonableness of This Report
A reasonable place
- First, expanding the AI bottleneck from GPUs to the interconnected system is the right direction—this is being validated by product releases from NVIDIA and Broadcom.
- Second, it is crucial to reject the oversimplified narrative of "copper declining, light advancing." Reuters' coverage of Jensen Huang has clearly indicated that copper still holds a reliability advantage in GPU/XPU core connections in the short term.
- Third, it is also reasonable to view CPO as the right direction but to wait for reliability validation before scaling. Industry analysts such as LightCounting and Yole/EDN favor a gradual transition rather than immediate full replacement.
- Fourth, emphasizing that "upstream" segments such as PCB/ABF/CCL, testing, and light sources are more likely to realize value by 2026 is more helpful for investors. This is because capital markets often overvalue distant, speculative narratives while undervaluing near-term segments that are already securing actual orders.
Things to note
First, publicly paraphrasing may turn Bernstein’s views into investment hype or clickbait. For example, the statement “The real battlefield for AI isn’t chips, but connectivity” is catchy, but strictly speaking, GPU/HBM/CoWoS remain the core bottlenecks—connectivity’s marginal importance has increased, not chip importance.
Second, the direction of value transfer for CPO is correct, but its adoption speed may be overestimated by the market. CPO addresses challenges in manufacturing, packaging, on-site maintenance, failure replacement, and reliability—it is not a technology that will immediately scale up after a product launch.
Third, the transition value of LPO/NPO is significant, but the system debugging complexity is also high. 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, although the PCB/ABF/CCL segment has strong certainty, be cautious of expansion cycles. Once the materials and substrate industries see high demand, they tend to rapidly expand capacity; if customer demand later slows down, gross margins will suffer as a result.
Ten: Over the next 2–3 years, you can track progress according to this timeline.
2026: Don't just look at CPO—focus on three certainties
In 2026, the focus won't be on a massive surge in CPO, but rather on:
Is the 1.6T pluggable optical module being scaled up?
Have LPO/NPO obtained more certifications from cloud providers/switching platforms?
Will PCB/ABF/CCL continue to increase prices or expand production?
Have there been any actual orders for CPO-related test equipment, FAUs, or external light sources?
If these occur, it indicates that the report's logic has entered its realization phase.
2027: Watch CPO pilots transition from “prototypes” to “customer deployments”
Key metrics are:
Real customer deployments of NVIDIA Quantum-X/Spectrum-X Photonics;
Customer expansion for Broadcom Davisson/Tomahawk CPO;
Do CoreWeave, Lambda, Meta, Google, Microsoft, Amazon, and others use it?
Have the CPO external light source, FAU, and testing equipment been recognized as revenue?
After 2028: Monitor whether CPO enters the scale-up phase
The most critical turning point is:
Is the CPO moving from the switch side toward the XPU/GPU?
Is optical I/O being integrated into high-end ASIC/GPU packages?
Is OCS/optical fabric beginning to transform data center network topologies?
At this point, CPO is not just about replacing optical modules, but about a shift in AI computing architecture.
Eleven: Investment Framework Based on This Report: Four Asset Classes, Four Logics
If I were to use this report to guide investments in U.S., Hong Kong, or A-shares, I would categorize them into four types.
My most trusted strategy is:
Buy the core warehouse winner, have confidence in the optical and PCB flexible warehouse, and make a small-position purchase of CPO forward direction via options.
It’s not recommended to invest all your funds upfront in the “purest CPO concept stocks.”
Twelve: The five most critical points of this report
- First, the bottleneck in AI data centers is shifting from "computing fast" to "connecting fast, connecting stably, and connecting with low power consumption."
- Second, light won't instantly eliminate copper, and copper won't permanently dominate all scenarios; different solutions will be chosen based on varying distances and system levels.
- Third, CPO is the direction, but more realistic revenue in 2026 lies in 1.6T, LPO/NPO, light sources, testing, PCB, ABF, and CCL.
- Fourth, the true 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, invest in AI connectivity—don’t buy the hottest trends, buy the hardest-to-avoid bottlenecks.
- This is a highly valuable report on "AI Layer-2 Infrastructure." It reminds the market that after GPUs, the next asset to be repriced won't be a single component, but the entire AI connectivity stack.
But it cannot be simply interpreted as "CPO will explode immediately." A more accurate interpretation is:
In 2026, watch for pluggable/LPO/NPO/PCB/ABF/testing;
Watch for CPO pilot orders in 2027;
Watch whether CPO and optical I/O truly enter the core AI computing architecture after 2028.