A recent 97-page deep report from Bernstein points out that copper and optical interconnects in AI data centers are not mutually exclusive but will coexist long-term in scale-up and scale-out scenarios. Although CPO (Co-Packaged Optics) technology has advantages in power consumption and cost, its widespread deployment still faces obstacles due to manufacturing and maintenance challenges. Large-scale adoption is unlikely before 2028, making optical interconnects like LPO/NPO the likely leaders during the transition. However, CPO is fundamentally reshaping the value chain, shifting profit centers from traditional optical module suppliers to chip design, advanced packaging, and system integrators.
A special note on Bernstein: Sanford C. Bernstein is a world-renowned investment research and asset management firm headquartered in the US. Founded in 1967 and now part of AllianceBernstein (AB), it is one of the largest and oldest independent sell-side research institutions. Below is a detailed breakdown of this report.
In mid-February, I detailed the underlying logic of bottleneck transmission in the AI computing power industry chain, noting that optical interconnects would be a main AI theme the market switches to in 2025-26.
I only truly began focusing on and researching the field of optical interconnects at the end of last year: https://x.com/qinbafrank/status/2015377625167089671?s=20
This Bernstein report focuses on three core aspects:
Why has connectivity replaced computing power as the new bottleneck? What is the rhythm of CPO realization? Why are PCB/ABF substrates a more realistic direction for performance realization in 2026? Let's break it down.
The real message of this report isn't "CPO is about to explode," but rather:
The bottleneck of AI data centers is migrating from GPU/HBM/CoWoS to "connectivity systems." The future investment theme isn't a solo win for CPO, but a collective upgrade of optical, electrical, copper, boards, packaging, and testing.
To put it more bluntly:
In the past, the market looked at AI through GPU computing power.
Now, the market is looking at how GPUs are connected.
In the future, it will look at whether computing power utilization can be released by connectivity systems.
This is the so-called "War for AI Data Center Connectivity" mentioned in the report title.
1. Why is "Connectivity" Becoming the New Bottleneck for AI Data Centers?
An AI cluster isn't just about stacking GPUs. The real problem is that 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 communication between GPUs can't keep up, actual utilization will drop.
Think of an AI cluster as a giant factory:
Why will connectivity replace computing power as the new bottleneck?
The root of this lies in how large models are trained. There are two parallel methods: Tensor Parallelism and Expert Parallelism. Both require frequent, large-scale data exchange between GPUs.
The amount of data exchanged during a single training session is astronomical. In the past, you just needed to stack GPUs; now, the more you stack, the higher the communication overhead. At a certain critical point, adding GPUs no longer speeds up training but worsens communication congestion. This is the connectivity bottleneck.
Bernstein provides a comparison: inside a standard NVIDIA GB30 rack, copper cables are used between GPUs because the distance is short, and copper is cheap and stable. However, fiber optics must be used between racks because copper signals suffer unacceptable attenuation beyond 2 meters. Optical modules at both ends of the fiber convert electrical signals to optical and back.
The problem is that a 1.6T optical module consumes about 30 watts, more than half of which is consumed by a chip called the DSP (Digital Signal Processor). With hundreds of optical modules in a rack, the power consumption of optical communication alone is hard to suppress.
Therefore, the real issue for current AI data centers isn't insufficient computing power, but power consumption hitting the ceiling. NVIDIA says its new generation of switches can save 70% power compared to traditional optical modules. For a 51.2T switch, this alone saves 500 watts, allowing for more GPUs.
NVIDIA is 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 and maintenance costs. NVIDIA claims its photonics switches can achieve 1.6Tb/s per port, a 3.5x energy efficiency improvement, 63x signal integrity improvement, and 10x network resilience.
The underlying logic of the Bernstein report is: the next stage of AI CAPEX isn't just buying more GPUs, but buying more "connectivity capability to make GPUs work effectively."
2. The Core Judgment: Not "Copper Out, Fiber In," but "Multi-Route Coexistence"
There is a simple saying in the market: copper retreats, light advances.
But this report offers a more nuanced view: copper and light are not simple substitutes but will coexist long-term under different distances, bandwidths, maintenance requirements, and cost structures. Bernstein believes copper and optical interconnects will develop separately in Scale-up and Scale-out scenarios. This judgment is crucial.
1. Scale-up: Intra-rack/Short-distance Interconnects, Copper is Still Strong
Scale-up refers to high-speed interconnects between GPUs, GPUs and switches, or within/near racks. The priorities here are:
Low latency, low cost, high reliability, maintainability, and short-distance transmission capability.
In this scenario, copper is not dying out anytime soon.
Jensen Huang has also clearly stated: NVIDIA will not use CPO for main connections between flagship GPUs for now because traditional copper connections are currently much more reliable than CPO; NVIDIA will first use CPO in two new network chips for top-of-rack switches.
This is very important. It shows that CPO is the direction, but not an immediate total replacement for copper.
In other words, at this stage, NVIDIA's logic is:
Switches can adopt CPO first, while GPU/XPU sides should be more cautious.
The reason is simple: GPUs are the most expensive and critical assets in the system. You cannot sacrifice reliability just to save energy with optical interconnects. In an AI training cluster, a frequently dropping link results in more than just hardware costs—it means interrupted training tasks, decreased GPU utilization, and increased scheduling complexity.
2. Scale-out: Inter-rack/Inter-cluster Interconnects, Optical has the Advantage
Scale-out involves larger GPU cluster expansion, usually involving east-west traffic between racks or within data centers over longer distances.
In this scenario, optical solutions have clear advantages:
Longer distance, higher bandwidth, lighter cables, lower power consumption, and better wiring density.
So the future isn't "copper being completely replaced by light," but rather:
The most valuable part of Bernstein's report is that it doesn't stop at "CPO concept stocks" but breaks AI connectivity into multiple technical routes.
3. CPO: The Direction is Important, but 2026 is Not the Year of Full Explosion
The most easily misunderstood part of this report is CPO.
Many see CPO and immediately conclude:
Optical modules will be replaced, CPO will explode instantly, and traditional optical module factories are finished.
This understanding is too crude.
Bernstein expects small-scale deployment of CPO in Scale-out networks to start in the second half of 2026, mainly to verify real performance and supply chain maturity. However, in the more critical Scale-up scenarios, CPO adoption may be delayed until after the second half of 2028, as the industry needs to verify the long-term reliability of CPO on the switch side before applying it to higher-value, zero-tolerance XPU systems.
This aligns with Jensen Huang's previous statements: CPO will first be used for network switching chips, not large-scale GPU main connections.
Therefore, the timeline should be understood like this:
LightCounting's view also supports "gradual evolution" rather than "overnight switching." It predicts traditional retimed pluggables will still dominate for the next 5 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 mentions that Yole believes large-scale CPO deployment may occur between 2028 and 2030, while LightCounting believes optical modules will still account for the majority of data center optical links this decade, though optical components will continue to move closer to the ASIC.
So my judgment is:
CPO is a medium-to-long-term direction, but the more certain revenue in 2026 won't necessarily be in pure CPO concept stocks, but in the upgrades required on the eve of CPO: light sources, testing, packaging, PCB, ABF, CCL, 1.6T optical modules, and LPO/NPO.
4. LPO/NPO: The "Transition Mainlines" Before the CPO Explosion
An important point in this report is that it doesn't simply divide technical routes into "traditional optical modules vs. CPO."
There are LPO and NPO in between.
1. What is LPO?
LPO stands for Linear Pluggable Optics. It can be roughly understood as: keeping 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, and still maintains some maintainability.
Disadvantages: Harder system debugging, tighter link budgets, and higher requirements for host-side SerDes and system engineering.
Public abstracts mention that by removing the DSP and handing signal processing to linear components, LPO can significantly reduce power consumption compared to traditional pluggable modules while retaining modular maintenance convenience. Bernstein even believes LPO shipments could exceed CPO by 2030.
2. What is NPO?
NPO stands for Near-Packaged Optics, which means placing the optical engine closer to the ASIC, but not completely sealed together like CPO.
Its value lies in the compromise:
This suggests that the next few years likely won't be a "one step to CPO" move, but rather:
Traditional pluggable → LPO/NPO → CPO → Optical I/O / optical fabric
This is why you can't just look at CPO in 2026. The companies that can truly realize performance are likely those that can supply across multiple stages.
In summary, the CPO story won't be realized in 2026. CPO will only ship in small batches in the second half of 2026 for scale-out scenarios, meaning large-scale deployment between racks will wait until 2028.
Why so slow? Bernstein gives three reasons:
First, cloud service providers are reluctant to switch. If a traditional optical module fails, maintenance just plugs in a new one in minutes. A CPO is soldered into the switch; if an optical engine fails, the entire switch must return to the factory. Downtime and maintenance costs are huge issues for Amazon, Google, and Microsoft. Moreover, optical module failure rates aren't low—the industry standard is one failure every 100,000 hours, meaning 9 out of 10,000 modules fail annually, and that's just hardware failures.
By integrating the optical engine into the chip, CPO reliability must improve by several orders of magnitude to reassure cloud providers. Bernstein explicitly stated that after communicating with Innolight, a Chinese optical module factory, they were told that no cloud provider client plans to deploy CPO on a large scale in 2026-2027. This is a heavy statement that the market might not have heard yet.
Second, transition solutions have emerged. CPO isn't the only choice. There are two intermediate technologies: LPO and NPO. LPO removes the power-hungry DSP chip, replacing it with simpler components. This cut reduces power consumption to 1/3 of traditional modules while keeping the pluggable 800G form factor, which is already in mass production.
NPO places the optical engine on the PCB next to the switch chip, but it's still detachable. NVIDIA's current CPO products are strictly speaking NPO. These two transition solutions can last 2 to 3 years. So cloud providers have every reason to use LPO first and wait for CPO to truly mature.
Third, in scale-up scenarios, copper isn't dead. Connections between GPUs are called scale-up. Copper's cost and reliability advantages currently have no substitutes.
Bernstein clearly states that from 2026 to 2028, scale-up will still be dominated by copper. Luxshare is a beneficiary here, competing directly with Amphenol for NVIDIA GB300 copper connectors. There is also a transition technology called CPC (Co-Packaged Copper), which further extends the life cycle of copper cables.
LightCounting predicts that by 2029, copper will still hold nearly half the share of the 1.6T connection market.
5. The Biggest Impact of CPO: Not Simple Cost Reduction, but Profit Pool Redistribution
The industrial significance of CPO isn't just energy saving or replacing optical modules.
It truly changes where profit is generated.
In the traditional pluggable era, the value chain was roughly:
DSP / Optical Chip / TOSA/ROSA / Module Packaging / Optical Module Factory / Switch Factory / Cloud Provider.
In the CPO era, it becomes:
Switch ASIC / Optical Engine / External Laser Source / FAU / Advanced Packaging / Wafer Fabrication / Testing / System Integration.
Bernstein performed a cost breakdown for the NVIDIA Quantum-X800 CPO switch: it features four switch ASICs, each integrated with 18 optical engines and 18 external light source modules. A single Quantum-X800 CPO switch is estimated to cost about $570,000. The abstract also notes that under the CPO architecture, the DSP is eliminated, the optical engine is co-packaged with the switch chip, and the value center shifts to chip design, advanced packaging, and wafer fabrication.
This is why the report favors these directions:
Relatively speaking, traditional optical module factories will face a problem: if value shifts from module packaging to ASICs, packaging, optical engines, and system integration, their profit pools may be restructured.
But this doesn't mean traditional factories lose value immediately. Because from 2026 to 2028, there will still be massive 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 still be the main growth engines in 2026.
So the correct understanding is:
CPO will change the profit distribution of the optical module industry chain, but it will not eliminate pluggable optical modules in 2026.
6. Why Does the Report Emphasize PCB, ABF, and CCL as More Realistic Directions for 2026?
This is what I believe deserves your most attention.
CPO has great imaginative space, but its realization cycle is later. In contrast, upgrades for PCB, ABF, and CCL are closer to current orders.
The reason is: even if CPO isn't commercially available on a large scale, AI servers and switches are already upgrading.
Rubin, Rubin Ultra, GB300, cloud provider ASICs, and next-gen switch ASICs are all increasing:
Single-board rates, packaging area, power density, signal integrity requirements, heat dissipation requirements, and low-loss material requirements.
This is the most counter-intuitive but easily overlooked point in this research report. The ones truly making money in 2026 are the "old money" tracks: PCB, HDI, ABF, and substrates.
Why counter-intuitive? Because these tracks are too traditional. PCB is a decades-old industry with a global market of $85 billion in 2025. It doesn't sound sexy. Everyone is staring at CPO, optical modules, and NVIDIA, and no one wants to spend time researching printed circuit boards. But Bernstein's data tells us this track has already quietly taken off in 2025.
Bernstein provided a set of numbers: Victory Giant Technology, which makes HDI (High-Density Interconnect) boards, saw 2025 revenue grow 63% year-on-year. WUS Printed Circuit, which supplies MPCB for NVIDIA GB300, saw 45% growth. Gold Circuit Electronics, supplying AWS Trainium, grew 40%, and Shengyi Electronics, another AWS supplier, grew 40%. These are real performance results that have already happened, not expectations. Why is this track rising? There are three dimensions:
First, the PCB content in AI servers has doubled. In the past, an NVIDIA H100 server with 8 GPUs had a total HDI and PCB value of about $100-$150 per GPU. For the GB200 NVL72 rack, this figure jumps to $300 per GPU. What does this mean? For every GPU sold, PCB manufacturers earn twice as much.
And it's not over. The upcoming Vera Rubin platform will adopt a new structure called a midplane, replacing parts originally connected by copper with multi-layer PCBs. This midplane is a 44-layer board using high-end M8 grade copper-clad laminate. The next-gen Rubin Ultra might use 78-layer M9 grade boards. Layer counts double, materials upgrade, and value doubles again.
Second, upstream materials are a bottleneck. ABF substrates require a key material called T-glass (low thermal expansion coefficient glass fiber) to prevent AI chips from deforming the substrate at high temperatures and causing solder joint failure.
Currently, only one company in the world can achieve top-tier T-glass specifications: Nittobo, with a CTE value of 2.8%. Other manufacturers can't reach this level. Nittobo's new capacity won't come online until the end of 2026, with formal shipments in 2027, meaning T-glass will be in continuous shortage throughout 2026.
What does a T-glass shortage mean? It means ABF substrate manufacturers can rightfully raise prices. Unimicron has already renegotiated prices with customers. Bernstein's model predicts that ABF substrate ASP will rise 5%-7% quarter-on-quarter in 2026, with an annual cumulative increase potentially exceeding 20%.
Third, the invisible monopolist of ABF film. ABF film is a core material for ABF substrates, invented by Ajinomoto, the Japanese food company known for MSG. In the 90s, while researching MSG, 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 world's ABF film has come from Ajinomoto.
According to Bernstein's data, Ajinomoto's ABF business has a 60% gross margin, with a 32% growth rate in FY2026, expected to accelerate to 45% in FY2027. This company's ABF business has been unshakable for 30 years.
So, the certainty for 2026 isn't an "overnight CPO explosion," but rather:
High-speed PCB upgrades; ABF substrate upgrades; CCL upgrades to lower-loss materials; upgrades in copper foil, glass fiber cloth, and low Dk/Df materials; and upgrades in testing and verification.
Therefore, a more realistic strategy for 2026 is to grasp three types of certainty: optical demand brought by 1.6T and LPO/NPO transitions, PCB/ABF/CCL upgrades brought by Rubin/ASIC, and the testing/FAU/light source/advanced packaging that must be invested in before CPO trial production.
Capital markets often make a mistake: they like to buy the most distant concepts, but the ones that produce results first are often the "infrastructure that must be built before the long-term concept."
CPO is like a future high-speed rail station. But before the station is fully operational, the ones making money are those building the roads, laying the tracks, supplying power, signaling systems, and testing equipment.
7. The Industry Chain Beneficiary Sequence in This Report
If we divide the AI connectivity industry chain into four layers:
Layer 1: Strongest Platform Winners
These companies don't just sell a part; 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 already includes TSMC, Coherent, Corning, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo Electric, TFC Communication, etc., in its ecosystem.
This shows NVIDIA is doing one thing: not just selling GPUs, but bringing the network architecture of AI factories under its platform control.
TSMC: The Invisible Hub of This Entire Story
Its COUPE platform combines electronic and photonic chips using hybrid bonding technology. All major clients—NVIDIA, Broadcom, AI labs—are migrating toward TSMC. This company doesn't make much from CPO itself, but CPO 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 provider custom chip ecosystem.
In October 2025, Broadcom announced Tomahawk 6 Davisson, its third-generation CPO Ethernet switch with 102.4Tbps switching capacity, stating it is already shipping. Broadcom claims that by integrating TSMC COUPE optical engines and advanced multi-chip packaging, it reduces optical interconnect power consumption by 70% while supporting scale-up for 512 XPUs and 100,000+ XPUs in a two-tier network.
This indicates that TSMC and Broadcom are critical companies alongside NVIDIA in the AI network and CPO value chain.
Layer 2: Optical and High-Speed Interconnects with Strong Certainty
This includes: 1.6T optical modules, LPO/NPO, silicon photonics, lasers, external light sources, FAU, and optical connectors.
Representative directions include Coherent, Lumentum, Fabrinet, Innolight, Eoptolink, SENKO, Corning, Sumitomo, etc. NVIDIA's official ecosystem list includes several optical, packaging, and connection-related companies.
The focus here isn't "who is most like CPO," but:
Who can simultaneously capture demand for 800G/1.6T, LPO/NPO, CPO trial production, external light sources, and FAU.
Companies that can span multiple stages have a higher win rate than single-concept companies.
Layer 3: PCB, ABF, CCL, Materials
This is where I believe 2026 is most likely to be undervalued.
Public reports mention that the original report covered or mentioned companies like Chroma, Luxshare, Unimicron, NVIDIA, Broadcom, TSMC, and Ibiden.
Among these, substrate/PCB chain companies like Unimicron and Ibiden are very noteworthy because as AI server complexity increases, PCB and packaging substrates are no longer just followers but performance constraints themselves.
Layer 4: Testing Equipment, Yield, Reliability
The biggest difficulty for CPO isn't the PPT, but mass production.
Mass production must solve: optical-electrical coupling yield; external laser source stability; reliability in high-temperature environments; packaging stress; field maintenance; testing time; consistency; and repair modes after failure.
Therefore, testing equipment and reliability verification could 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.
8. Investment Implications: Don't Buy the "Most Conceptual," Buy the "Hardest to Bypass"
The biggest takeaway from this report for investment is:
AI connectivity is not a single-point technical revolution, but a bottleneck migration. Invest in common bottlenecks, not single routes.
What is a common bottleneck?
Something that cannot be bypassed regardless of whether the final solution is CPO, LPO, NPO, or continued upgrades of traditional pluggables. For example:
Conversely, single-route risks are higher.
For example, if you only buy "pure CPO concepts," the risk is: CPO mass production is delayed, orders aren't realized, and valuations are slashed first.
If you only buy traditional optical modules, the risk is: CPO/NPO/LPO restructure the value chain, and long-term profit pools are taken by platform factories and chip/packaging factories.
If you only buy PCB/materials, the risk is: customers expand capacity too quickly, supply is released centrally, and gross margins reverse.
So a better combination is:
Buy certainty in 2026, buy order elasticity in 2027, and buy architectural options after 2028.
9. Personal Evaluation of the Report's Rationality
What is very reasonable
First, expanding the AI bottleneck from GPUs to connectivity systems is a very correct direction. Product releases from NVIDIA and Broadcom are verifying this.
Second, opposing the simple narrative of "copper out, fiber in" is a crucial judgment. Reuters' reporting on Jensen Huang has clearly shown that copper still has reliability advantages in core GPU/XPU connections in the short term.
Third, the view that CPO is the direction but scale-up requires reliability verification is also reasonable. Industry judgments from LightCounting and Yole/EDN lean toward "gradual migration rather than immediate total replacement."
Fourth, emphasizing that "front-end links" like PCB/ABF/CCL, testing, and light sources are more likely to be realized in 2026 is very helpful for investment. Capital markets tend to over-trade the most distant stories while underestimating the links that are actually receiving orders in the near term.
What to watch out for
First, public summaries might "investorize" or "sensationalize" Bernstein's views. For example, the phrase "The real AI battlefield isn't in chips, but in connectivity" is catchy, but strictly speaking, GPU/HBM/CoWoS are still core bottlenecks; it's just that the marginal importance of connectivity is rising, not that chips are unimportant.
Second, the direction of CPO value transfer is correct, but the speed might be overestimated by the market. CPO must solve problems in manufacturing, packaging, field maintenance, failure replacement, and reliability; it's not a technology that scales immediately after a press conference.
Third, the transition value of LPO/NPO is great, but their system debugging difficulty is not low. LPO isn't just a "low-power version of pluggable"; it shifts a lot of complexity to the host side and system-level debugging.
Fourth, while the PCB/ABF/CCL line has strong certainty, one must also be wary of expansion cycles. Once the material and substrate industries see high prosperity, it's easy to expand capacity, and if the customer platform rhythm slows down later, gross margins will suffer.
10. Tracking Timeline for the Next 2-3 Years
2026: Don't just look at CPO, look at three certainties
The focus in 2026 isn't a CPO explosion, but:
- Whether 1.6T pluggable optical modules scale up;
- Whether LPO/NPO gain more cloud provider/switch platform certifications;
- Whether PCB/ABF/CCL continue to rise in price or expand capacity;
- Whether CPO-related testing equipment, FAU, and external light sources start having actual orders.
If these happen, it means the report's logic is entering the realization phase.
2027: Watch CPO pilots move from "prototypes" to "customer deployment"
Key indicators:
- Real customer deployment of NVIDIA Quantum-X / Spectrum-X Photonics;
- Customer expansion of Broadcom Davisson/Tomahawk CPO;
- Adoption by CoreWeave, Lambda, Meta, Google, Microsoft, Amazon, etc.;
- Revenue recognition for CPO external light sources, FAU, and testing equipment.
2028 and Beyond: See if CPO enters Scale-up
The most critical turning point:
- Whether CPO moves from the switch side to near the XPU/GPU;
- Whether optical I/O enters high-end ASIC/GPU packaging;
- Whether OCS/optical fabric starts changing data center network topologies.
If it reaches this step, CPO is no longer just an optical module replacement but a change in AI computing architecture.
11. Investment Framework Based on This Report: Four Asset Classes, Four Logics
If using this report to guide investments in US, HK, or A-shares, I would divide them into four categories.
My most recognized strategy is:
Buy platform winners for core holdings, buy optical and PCB certainty for elasticity holdings, and buy long-term CPO directions with a small proportion for option holdings.
I do not recommend betting all funds on "pure CPO concept stocks" right away.
12. Five Core Takeaways from This Report
- The bottleneck of AI data centers is shifting from "calculating fast" to "connecting fast, connecting stably, and connecting with power efficiency."
- Light will not immediately eliminate copper, nor will copper hold all scenarios forever; different distances and system levels will choose different solutions.
- CPO is the direction, but the more realistic revenue in 2026 is in 1.6T, LPO/NPO, light sources, testing, PCB, ABF, and CCL.
- The true impact of CPO isn't making optical modules cheaper, but shifting the profit pool from traditional module packaging to chips, packaging, optical engines, light sources, testing, and system platforms.
- When investing in AI connectivity, don't buy the hottest concepts; buy the bottlenecks that are hardest to bypass.
This is a very valuable "AI second-layer infrastructure" report. It reminds the market: after GPUs, the next thing to be repriced isn't a single part, but the entire AI connectivity stack.
But it shouldn't be simply read as "CPO will explode immediately." A more accurate reading is:
2026: Look at pluggable/LPO/NPO/PCB/ABF/testing;
2027: Look at CPO pilot orders;
2028 and beyond: See if CPO and optical I/O truly enter the AI computing core architecture.
