A new 97-page in-depth report from Bernstein points out that copper and optical interconnects in AI data centers are not simply replacing each other but will co-exist long-term in scale-up and scale-out scenarios. Although CPO technology offers advantages in power consumption and cost, its widespread deployment still faces obstacles due to manufacturing and maintenance challenges. Large-scale adoption is unlikely to occur before 2028. Therefore, optical interconnects like LPO/NPO may become the leaders in the transitional period. However, CPO is fundamentally reshaping the value chain, shifting the profit centers from traditional optical module suppliers to chip designers, advanced packaging providers, and system integrators.
It's important to specifically mention the institution, Bernstein (full name: Sanford C. Bernstein), which is a globally renowned investment research firm and asset management institution headquartered in the United States. Founded in 1967, it is currently part of the global asset management giant AllianceBernstein (AB) and is one of the largest and oldest independent sell-side research institutions. Let's break down this Bernstein report in detail.
In mid-February, the underlying logic of AI computing power industry chain bottlenecks and their transmission was detailed, mentioning that optical interconnects are one of the AI thematic shifts the market is experiencing in 2025-2026.
Originally, https://x.com/qinbafrank/status/2015377625167089671?s=20 only started truly paying attention to and researching the field of optical interconnects around the end of last year.
This Bernstein report focuses on three core aspects:
Why is connectivity replacing computing power as the new bottleneck? What is the timeline for CPO materialization? Why are PCB/ABF substrates a more realistic direction for performance delivery in 2026? Let's break it down in detail.
What this report truly wants to convey is not "CPO is about to explode," but rather:
The bottleneck in AI data centers is shifting from GPU/HBM/CoWoS towards the "connectivity system." The future investment theme is not CPO alone winning, but the collective upgrade of optics, electronics, copper, boards, packaging, and testing.
Put more bluntly:
In the past, the market looked at AI mainly through the lens of GPU computing power.
Now, the market is starting to look at how to connect GPUs together.
In the future, we will need to look at whether computing power utilization can be unlocked by the connectivity system.
This is what the report's title refers to as "War for AI Data Center Connectivity."
I. Why is "Connectivity" Becoming the New Bottleneck in AI Data Centers?
An AI cluster is not simply about stacking GPUs together. The real problem is: These GPUs must synchronize at high speeds, exchange parameters, transmit activation values, perform AllReduce, and handle model and data parallelism. No matter how strong the theoretical computing power is, if the communication between GPUs cannot keep up, actual utilization will drop.
An AI cluster can be understood as a massive factory:
Why is connectivity replacing computing power as the new bottleneck?
The root cause of this lies in the training methods of large models. There are two parallel methods for large model training:
One is called tensor parallelism, and the other is expert parallelism. A common characteristic of both methods is the need for frequent, 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 just increase the number of GPUs. Now, the more you add, the greater the communication overhead becomes. Beyond a certain critical point, adding GPUs no longer speeds up training but instead worsens communication congestion—this is the connectivity bottleneck.
Bernstein provides a comparison. In a standard NVIDIA GB30 rack, connections between GPUs use copper cables because they are short-distance, cheap, and stable. However, connections between racks must use fiber optics because copper cables suffer severe signal attenuation beyond 2 meters. Optical modules are needed at both ends of the fiber to convert electrical signals to optical signals and back again.
Here's the problem: A 1.6T optical module consumes about thirty watts of power, with a large portion consumed by a chip called a DSP (Digital Signal Processor). With hundreds of optical modules in a single rack, the power consumption for communication alone becomes difficult to manage.
Therefore, the real problem facing AI data centers today is not insufficient computing power but hitting the power ceiling. NVIDIA itself states that its new CPU switch can save up to 70% of power compared to traditional optical modules. For a 51.2T switch, this alone can save five hundred watts, allowing you to fit in more GPUs with the saved power.
NVIDIA is also reinforcing this narrative. In March 2025, NVIDIA released Spectrum-X Photonics and Quantum-X silicon photonics switches, emphasizing they are designed to connect millions of GPUs in AI factories while reducing energy consumption and operational costs. NVIDIA claims its photonics switches can achieve 1.6Tb/s per port, with 3.5x higher energy efficiency, 63x better signal integrity, and 10x improved network resilience.
The underlying logic of Bernstein's report is: The next phase of AI capital expenditure is not just about buying more GPUs, but buying more "connectivity capability that enables GPUs to work effectively."
II. The Report's Most Critical Judgment: Not "Optics Replacing Copper," but "Multi-Route Coexistence"
The market often has a simplistic saying: Optics replacing copper.
But this report presents a more nuanced view: Copper and optics are not in a simple substitution relationship. They will coexist long-term under different distances, bandwidths, maintenance requirements, and cost structures. Bernstein believes copper and optical interconnects are not simple substitutes but will develop separately in scale-up and scale-out scenarios. This judgment is crucial.
1. Scale-up: Rack/Short-Distance Interconnect, Copper Remains Strong
Scale-up is closer to high-speed interconnect between GPUs, GPUs and switches, within a rack, or over short distances between nearby racks. Here, the most valued aspects are:
Low latency, low cost, high reliability, maintainability, short-distance transmission capability.
In this scenario, copper is not dead yet.
Jensen Huang has also clearly stated: NVIDIA will not use CPO for main connectivity between flagship GPUs for now because traditional copper connections are currently much more reliable than CPO optical connections. NVIDIA will first use CPO in two new networking chips within the top-of-rack switches.
This statement is very important. It indicates: CPO is the direction, but it's not an immediate, comprehensive replacement for copper.
In other words, at least for the current stage, NVIDIA's logic is:
The switch side can adopt CPO first, while the GPU/XPU side must be more cautious.
The reason is simple: GPUs are the most expensive and critical assets in the system. You cannot sacrifice reliability just for the energy savings of optical interconnects. In an AI training cluster, frequent link failures cause losses not only in hardware costs but also in training task interruptions, decreased GPU utilization, and increased scheduling complexity.
2. Scale-out: Rack/Cluster Interconnect, Optics Hold More Advantages
Scale-out involves larger-scale GPU cluster expansion, typically covering longer-distance, east-west traffic between racks and within data centers.
In this scenario, optical solutions have more obvious advantages:
Longer distance, higher bandwidth, lighter cables, lower power consumption, better cabling density.
Therefore, the future is not "copper being completely replaced by optics," but rather:
The most valuable aspect of Bernstein's report: It doesn't stop at the "CPO concept stock" level but breaks down AI connectivity into multiple technology routes.
III. CPO: The Direction is Important, but 2026 is Not the Year of Full-Scale Breakout
The most easily misunderstood part of this report is CPO.
Many people see CPO and immediately conclude:
Optical modules will be replaced, CPO is exploding immediately, traditional optical module manufacturers are finished.
This understanding is too simplistic.
Bernstein expects small-scale deployment of CPO in scale-out networks may begin in the second half of 2026, primarily to validate real performance and supply chain maturity. However, in the more critical scale-up scenario, CPO adoption may be delayed until the second half of 2028 or later. The industry needs to first validate the long-term reliability of CPO on the switch side before applying it to higher-value, less fault-tolerant XPU systems.
This aligns with Jensen Huang's previous statement: CPO will be used first in network switch chips, not directly in large-scale GPU main connectivity.
Therefore, the timeline should be understood as follows:
LightCounting's viewpoint also supports "gradual evolution" over an "overnight switch." It predicts that traditional retimed pluggables will still dominate over the next five years, although LPO/CPO will account for a significant share of 800G and 1.6T ports in 2026–2028. A summary of industry views by EDN also mentions that Yole believes large-scale CPO deployment may occur between 2028–2030, while LightCounting believes optical modules will still comprise the majority of data center optical links within this decade, but optical components will continue moving closer to ASICs.
Therefore, my judgment is:
CPO is a mid-to-long-term direction, but the more certain revenue in 2026 may not lie in the purest CPO concept stocks, but rather in the light sources, testing, packaging, PCB, ABF, CCL, 1.6T optical modules, and LPO/NPO that must be upgraded before the CPO era.
IV. LPO/NPO: They Are the "Transitional Theme" Before the CPO Breakout
An important point of this report is that it does not simply divide technology routes into "traditional optical modules vs. CPO."
There are also LPO and NPO in between.
1. What is LPO?
LPO, short for Linear Pluggable Optics. It can roughly be understood as: Retaining the pluggable form factor but removing or weakening the DSP, using linear drivers and host-side equalization to reduce power consumption.
Advantages: Lower power consumption, potentially lower cost, retains some maintainability.
Disadvantages: More difficult system debugging, tighter link budget, higher requirements for host-side SerDes and system engineering.
A public summary mentions that LPO, by removing the DSP and handing signal processing to linear components, can significantly reduce power consumption compared to traditional pluggable modules while retaining modular maintenance convenience. Bernstein even believes LPO shipments may surpass CPO by 2030.
2. What is NPO?
NPO can be understood as Near-Packaged Optics, meaning placing the optical engine closer to the ASIC, but not fully integrated into it like CPO.
Its value lies in compromise:
This indicates that the next few years are likely not a "direct jump to CPO," but rather:
Traditional Pluggable → LPO/NPO → CPO → Optical I/O / Optical Fabric
This is also why you cannot just focus on CPO in 2026. Companies that can supply across multiple stages are the ones likely to deliver performance.
To summarize, the CPO story won't materialize in 2026. CPO can only ship in small batches in the second half of 2026, only for scale-out scenarios (between racks). Large-scale rollout will have to wait until 2028.
Why so slow? Bernstein provides three reasons:
The first reason is cloud service providers' reluctance to switch. With traditional optical modules, if there's a failure, you can just pull it out and replace it—done in minutes. With CPO, the optical engine is soldered into the switch. If an optical engine fails, the entire switch needs to be sent back to the factory. Downtime and maintenance costs are major concerns for cloud service providers like Amazon, Google, and Microsoft. Moreover, optical module failure rates aren't low—the industry standard is one failure per 100,000 hours. Translated, out of 10,000 optical modules, nine might need replacing in a year, and that's just for hard failures, not counting soft failures.
CPO integrates the optical engine into the chip, requiring reliability improvements by orders of magnitude for cloud service providers to feel comfortable. Bernstein explicitly states they communicated with the Chinese optical module manufacturer InnoLight, which told them that none of their cloud service provider customers plan to deploy CPO on a large scale in 2026-2027. This statement is significant, and the market may not have fully absorbed it.
The second reason is that transitional solutions have already emerged. CPO is not the only option. There are two intermediate technologies: one called LPO and the other called NPO. LPO removes the most power-hungry DSP chip from the optical module and replaces it with simpler components. This single move reduces power consumption to one-third of traditional optical modules while retaining pluggability. 800G LPO is already in mass production.
NPO places the optical engine on the PCB near the switch chip but remains detachable. What NVIDIA currently calls CPO products are, strictly speaking, actually NPO. These transitional solutions can last for 2 to 3 years. Therefore, cloud service providers have every reason to say, "We'll stick with LPO for now and wait for CPO to truly mature."
The third reason is that in scale-up scenarios, copper cables aren't dead yet. Connections between GPUs are called scale-up. Here, the cost and reliability advantages of copper cables currently have no comparable alternatives.
Bernstein clearly states that from 2026 to 2028, scale-up will still be dominated by copper cables. Luxshare Precision is a beneficiary here, competing head-on with Amphenol for NVIDIA's GP300 copper cable connectors. There's also a transitional technology called CPC (Co-Packaged Copper Cable), which further extends the lifecycle of copper cables.
Industry consultancy LightCounting predicts that in the 1.6T connectivity market by 2029, copper cables will still hold nearly half the share.
V. CPO's Biggest Impact: Not Simply Reducing Cost, But Redistributing Profit Pools
The industrial significance of CPO is not just energy savings or simply replacing optical modules.
Its real change is: Where profits are generated.
In the era of traditional pluggable optical modules, the value chain was roughly:
DSP / Optical Chips / TOSA/ROSA / Module Packaging / Optical Module Manufacturers / Switch Manufacturers / Cloud Providers.
In the CPO era, it will become:
Switch ASIC / Optical Engine / External Light Source / FAU / Advanced Packaging / Wafer Fabrication / Testing / System Integration.
Bernstein performed a cost breakdown for the NVIDIA Quantum-X800 CPO switch: This switch is configured with four switch ASICs, each integrated with 18 optical engines, and there are 18 external light source modules. The estimated cost per Quantum-X800 CPO switch is about $570,000. The summary also points out that under the CPO architecture, the DSP is eliminated, and the optical engine is co-packaged with the switch chip, shifting the value center towards chip design, advanced packaging, and wafer manufacturing.
This is why the report is favorable to these directions:
Relatively speaking, traditional optical module manufacturers will face an issue:
If value shifts from module packaging to ASIC, packaging, optical engines, and system integration, their profit pools may be restructured.
But this doesn't mean traditional optical module manufacturers have no immediate value. Because from 2026–2028, there will still be substantial demand for 800G, 1.6T, and LPO/NPO. Cignal AI also notes that high-speed datacom modules, especially 800GbE and emerging 1.6TbE designs, will remain the main growth engine in 2026.
Thus, the correct understanding is:
CPO will change the profit distribution in the optical module industry chain, but it won't immediately eliminate pluggable optical modules in 2026.
VI. Why Does the Report Emphasize PCB, ABF, CCL as More Realistic Directions for 2026?
This is what I believe deserves your attention the most.
CPO has a large imagination space, but its realization cycle is further out. In comparison, the upgrades for PCB, ABF, and CCL are closer to current orders.
The reason: Even before CPO is commercially deployed on a large scale, AI servers and switches are already being upgraded.
Rubin, Rubin Ultra, GB300, cloud provider ASICs, next-generation switch ASICs—all are increasing:
Per-board data rates, packaging area, power supply density, signal integrity requirements, thermal dissipation requirements, low-loss material requirements.
This is the most contrarian yet easily overlooked point in this research report. The real money in 2026 will be made in the old-money sectors of PCB, HDI, ABF, and substrates.
Why contrarian? Because this sector is too traditional. PCB is a decades-old industry, with a global market of $85 billion in 2025, sounding not sexy at all. Everyone is focused on CPO, optical modules, NVIDIA. No one wants to spend time studying printed circuit boards. But Bernstein's data tells us this sector has quietly taken off in 2025.
Bernstein provided a set of numbers: Shengyi Technology, which makes HDI (High-Density Interconnect) boards, saw its 2025 revenue grow 63% year-over-year. WUS (likely referring to Shengyi Electronics' board business, or a similar firm) supplying PCB for NVIDIA's GB300 saw revenue growth of 45%. Gold Circuit (likely referring to Kingboard or a similar company) supplying AWS's Trinium saw year-on-year supply growth of 40%. Shengyi Electronics, another supplier in the AWS supply chain, grew 40%. These are real, already realized performance numbers, not expectations.
Why is this sector rising? We can look at it from three dimensions:
The first layer is that the content value of PCB in AI servers has doubled. In the past, for an NVIDIA H100 server, the total value of HDI plus PCB per GPU board was about $100 to $150. Switching to the GB200 VL72 rack, this number directly jumps to $300 per GPU. What does this mean? For selling the same GPU, PCB manufacturers earn double the money.
And that's not all. The upcoming Vera Rubin platform will adopt a new structure called a midplane, replacing parts originally connected with copper cables with multilayer PCBs. This midplane is a 44-layer board using the highest-grade M8 level copper-clad laminate. The next-generation Rubin Ultra might use a 78-layer board with M9 grade material. Doubling the layers and upgrading materials again doubles the value.
The second layer is upstream material bottlenecks. ABF substrates have a critical material called T-glass (low coefficient of thermal expansion glass fiber). Its purpose is to prevent the substrate from deforming under high temperatures in AI chips, which can cause solder joint failures.
Currently, only one company globally can produce T-glass at the top specification: Nittobo (Japan). Its CTE (Coefficient of Thermal Expansion) value is 2.8 ppm/°C (note: article says 2.8%, likely a typo/interpretation error; CTE is usually in ppm). Other manufacturers cannot achieve this level. Nittobo's new capacity won't come online until the end of 2026, with formal shipments in 2027. This means T-glass will remain in short supply throughout 2026.
What does T-glass shortage mean? It means ABF substrate manufacturers can legitimately raise prices. Unimicron (Taiwan) has already renegotiated prices with its customers. Bernstein's model predicts that the ASP (Average Selling Price) of ABF substrates will increase 5% to 7% quarter-over-quarter in 2026, with cumulative annual growth possibly exceeding 20%.
The third layer is the hidden monopolist of ABF film. ABF film is one of the core materials for ABF substrates. The inventor of this material is Ajinomoto—yes, the Japanese food company that sells seasoning. In the 1990s, while researching seasoning, they accidentally discovered a special amino acid-derived film that could serve as a thermal expansion layer for semiconductor substrates. Since then, 95% of the global ABF film has come from Ajinomoto.
Bernstein's data shows that Ajinomoto's ABF business has a gross margin of 60%, with growth of 32% in FY2026 and an expected acceleration to 45% in FY2027. This company's ABF business has been unshakable for 30 years.
Therefore, what is more certain for 2026 is not "CPO exploding overnight," but rather:
High-speed PCBs need to be upgraded; ABF substrates need to be upgraded; CCL needs to be upgraded to lower-loss materials; Copper foil, glass fiber cloth, 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 target three types of certainty: 1) Optical demand driven by the transition to 1.6T and LPO/NPO; 2) PCB/ABF/CCL upgrades driven by Rubin/ASICs; 3) Testing/FAU/Light Source/Advanced Packaging investments required before CPO trial production.
Because capital markets often make a mistake:
They like to buy the most distant concept, but the real performance often comes first from the "infrastructure that must be built before the distant concept."
CPO is like the future high-speed rail station.
But before the high-speed rail station fully operates, those who make money first are likely the ones building the roads, laying the tracks, providing power and signaling systems, and inspection equipment.
VII. The Order of Beneficiaries in This Report's Industry Chain
If we divide the AI connectivity industry chain into four layers:
First Layer: Strongest Platform-Level Winners
These companies don't just sell a single component; they control the architecture.
NVIDIA
NVIDIA's advantage isn't just GPUs, but GPU + NVLink + InfiniBand + Ethernet + Spectrum-X + Quantum-X + software ecosystem. NVIDIA's official disclosure of silicon photonics networking switches has already included TSMC, Coherent, Corning, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo Electric, TFC Communication, etc., into its ecosystem.
This indicates NVIDIA is doing one thing:
Not just selling GPUs, but also bringing the AI factory's network architecture under its platform control.
TSMC, it is the invisible hub of this entire story.
The COUPE platform combines electronic and photonic chips using hybrid bonding technology. All major clients—NVIDIA, Broadcom, AI labs—are migrating to TSMC. This company may not earn much from CPO itself, but CPO strengthens TSMC's dominance in advanced packaging and wafer foundry.
Broadcom
Broadcom's logic is different. It is more like:
Ethernet switch ASIC + custom ASIC + CPO + cloud provider custom chip ecosystem.
In October 2025, Broadcom announced the Tomahawk 6 (Davisson), its third-generation CPO Ethernet switch, with 102.4Tbps switching capacity, claiming it's already shipping. Broadcom states 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 of 512 XPUs and over 100,000 XPUs in a two-tier network.
This indicates that TSMC and Broadcom are crucial companies in the AI network and CPO value chain alongside NVIDIA.
Second Layer: Optics and High-Speed Interconnects with Stronger Certainty
This includes:
1.6T optical modules, LPO/NPO, silicon photonics, lasers, external light sources, FAU (Fiber Array Unit), optical connectors.
Representative directions include Coherent, Lumentum, Fabrinet, Innolight (中际旭创), Eoptolink (新易盛), SENKO, Corning, Sumitomo Electric, etc. NVIDIA's official ecosystem list includes many optical, packaging, and connectivity-related companies.
The focus of this layer is not "who is most like CPO," but:
Who can simultaneously capture demand from 800G/1.6T, LPO/NPO, CPO trial production, external light sources, and FAU.
Companies capable of spanning multiple stages have higher odds of winning than single-concept companies.
Third Layer: PCB, ABF, CCL, Materials
This is the area where the report suggests being most easily underestimated in 2026.
The public summary mentions that the original report covers or mentions companies like Chroma (致茂电子), Luxshare (立讯精密), Unimicron (欣兴电子), NVIDIA, Broadcom, TSMC, Ibiden (揖斐电), etc.
Among these, substrate/PCB chain companies like Unimicron and Ibiden are particularly noteworthy. Because after AI server complexity increases, PCBs and packaging substrates are no longer just passive components but become performance constraints themselves.
Fourth Layer: Test Equipment, Yield, Reliability
The biggest challenge with CPO is not the PowerPoint presentation, but mass production.
Mass production needs to solve:
Opto-electronic coupling yield;
External light source stability;
High-temperature environmental reliability;
Packaging stress;
Field maintenance;
Testing time;
Consistency;
Repair models after failure.
Therefore, test equipment and reliability verification may be excellent "pick-and-shovel" plays.
These companies might not be the sexiest, but if CPO enters trial production, they are often the first to see orders.
VIII. Investment Implications of This Report: Don't Buy "The Most Concept-Like," Buy "The Hardest to Bypass"
This report's biggest inspiration for investment is:
AI connectivity is not a single-point technological revolution, but a bottleneck migration. Invest in common bottlenecks, not in single routes.
What are common bottlenecks?
They are things that are unavoidable regardless of whether the final solution is CPO, LPO, NPO, or continued upgrades of traditional pluggables. For example:
Conversely, single-route risks are relatively higher.
For example, if you only buy "pure CPO concept" stocks, the risks are:
CPO mass production timeline gets delayed, orders don't materialize, valuations get cut first.
If you only buy traditional optical modules, the risks are:
CPO/NPO/LPO restructures the value chain, and long-term profit pools are taken by platform companies and chip/packaging manufacturers.
If you only buy PCB/materials, the risks are:
Clients over-expand production capacity, supply gets concentrated and released, gross margins reverse.
Therefore, a better portfolio is:
Buy certainty in 2026, order elasticity in 2027, and architecture options for 2028 and beyond.
IX. Personal Evaluation of the Report's Rationality
Very Reasonable Aspects
- First, expanding the AI bottleneck from GPU to the connectivity system is a very correct direction. Product releases from NVIDIA and Broadcom validate this.
- Second, opposing the simplistic narrative of "optics replacing copper" is a very important judgment. Reuters' report on Jensen Huang already explicitly states that copper still has short-term reliability advantages in core GPU/XPU connectivity.
- Third, viewing CPO as the direction but requiring reliability validation before scaling is also a reasonable judgment. Industry assessments from LightCounting and Yole/EDN lean towards "gradual migration, not immediate comprehensive replacement."
- Fourth, emphasizing that "pre-requisite" links like PCB/ABF/CCL, testing, light sources are more likely to deliver in 2026 is more helpful for investment. Because capital markets tend to over-trade the most distant story while undervaluing links that receive actual orders in the near term.
Aspects to Note
First, the public summary may have "investment-ized" or sensationalized Bernstein's views. For example, the phrase "AI's real battlefield is not in chips, but in connectivity" has viral appeal, but strictly speaking, GPU/HBM/CoWoS are still core bottlenecks. It's just that connectivity's marginal importance is rising; it's not that chips are unimportant.
Second, the direction of CPO's value transfer is correct, but the speed may be overestimated by the market. CPO needs to solve manufacturing, packaging, field maintenance, failure replacement, reliability issues—it's not a technology that scales immediately after a press conference.
Third, the transitional value of LPO/NPO is significant, but their system debugging difficulty is also not low. LPO is not simply a "low-power version of pluggable optics"; it shifts a lot of complexity to the host side and system-level debugging.
Fourth, while the PCB/ABF/CCL line is highly certain, one must also be wary of the expansion cycle. Once the materials and substrate industries see high prosperity, they are prone to over-expansion. If client platform rhythms slow down later, gross margins will suffer.
X. Over the Next 2–3 Years, Track According to This Timeline
2026: Don't Just Focus on CPO; Look at Three Certainties
The focus for 2026 is not a CPO breakout, but:
Whether 1.6T pluggable optical modules ramp up volume;
Whether LPO/NPO gain more certifications from cloud providers/switch platforms;
Whether PCB/ABF/CCL continue to see price increases or capacity expansions;
Whether CPO-related test equipment, FAU, external light sources start receiving actual orders.
If these happen, it means the report's logic is entering its realization phase.
2027: Watch CPO Pilots Transition from "Prototypes" to "Customer Deployments"
Key indicators are:
Real customer deployments of NVIDIA Quantum-X / Spectrum-X Photonics;
Customer expansion of Broadcom Davisson/Tomahawk CPO;
Adoption by CoreWeave, Lambda, Meta, Google, Microsoft, Amazon, etc.;
Whether CPO external light sources, FAU, and test equipment enter revenue recognition.
2028 Onwards: Watch if CPO Enters Scale-up
The most critical turning point is:
Whether CPO moves from the switch side to near XPU/GPU;
Whether optical I/O enters high-end ASIC/GPU packaging;
Whether OCS/Optical Fabric starts changing data center network topology.
If we reach this point, CPO is no longer just about replacing optical modules, but about changes in AI computing architecture.
XI. Investment Framework Based on This Report: Four Asset Types, Four Logics
If using this report to guide US/HK/A-share investments, I would categorize into four types.
The strategy I personally find most credible is:
Core holdings in platform winners; flexible holdings in optical and PCB certainties; option holdings (small proportion) in CPO long-term direction.
It is not advisable to put all funds into "the purest CPO concept stocks" from the start.
XII. The Five Most Essential Points of This Report
- First, The bottleneck of AI data centers is shifting from "computing fast" to "connecting fast, connecting stably, connecting power-efficiently."
- Second, Optics will not immediately eliminate copper, nor will copper hold all scenarios forever; different distances and system layers will choose different solutions.
- Third, CPO is the direction, but more realistic revenue in 2026 lies in 1.6T, LPO/NPO, light sources, testing, PCB, ABF, CCL.
- Fourth, CPO's real impact isn't making optical modules cheaper, but shifting profit pools 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 concept; buy the hardest-to-bypass bottleneck.
- This is a very valuable "AI second-layer infrastructure" report. It reminds the market: After GPUs, the next thing to be re-priced isn't a single component, but the entire AI connectivity stack.
However, it also cannot be simplistically read as "CPO will explode immediately." A more accurate reading is:
Focus on pluggable/LPO/NPO/PCB/ABF/testing in 2026;
Watch for CPO pilot orders in 2027;
Watch to see if CPO and optical I/O truly enter the core architecture of AI computing in 2028 and beyond.
