AI Data Center Connectivity Battle: Is Copper Withdrawing and Optical Advancing Real? What Changes Are There in Supply Chain Profit Distribution in the CPO Era?

AI Data Center Connectivity Battle: Is Copper Withdrawing and Optical Advancing Real? What Changes Are There in Supply Chain Profit Distribution in the CPO Era?

The long-standing rivalry between copper and optical technologies within AI data centers is reaching a critical turning point, but the narrative of “copper retreats, optics advances” is far from the whole truth.

Bernstein’s latest white paper points out that as AI cluster sizes rapidly expand, connectivity capabilities are becoming the core bottleneck affecting system performance and costs, and are a new focal point in industry competition. For years to come, copper interconnects and optical interconnects will not simply replace each other, but will coexist over the long term in different distances and application scenarios, evolving separately along “Scale-up” and “Scale-out” paths.

Nvidia and Broadcom are pushing CPO from concept to commercialization. Nvidia has stated its CPO switches will see limited deployment in the second half of 2026, with AI cloud service providers such as CoreWeave and Lambda as the first users. However, due to the complexity in manufacturing, packaging, and testing, from 2026 to 2028, hyperscale cloud service providers will continue to rely mainly on pluggable optical modules. Meanwhile, copper interconnects, owing to their cost, power consumption, and maturity, will continue to dominate Scale-up scenarios for at least three more years.

The real significance of CPO lies in redefining the value distribution in the industry chain. Bernstein estimates that the cost of CPO optical engines and lasers is about 10% higher than the average price of 1.6T pluggable modules, but the profit center will shift from traditional optical module manufacturers to chip fabrication and advanced packaging.

This upgrade will also drive structural growth in basic material sectors, such as high-end PCB, ABF substrates, and T-glass fiber. However, as new production capacity is released in bulk from the end of 2026, increased price competition, depreciation, and supply expansion will gradually compress profit margins. Investors should focus on companies with leading capabilities in technology, manufacturing, and supply chain control.

Copper and optics each have their role—scale-up and scale-out paths are distinct

The expansion of AI infrastructure mainly follows two paths: Scale-up and Scale-out.

Scale-up involves continually increasing computing resources within a single system—for example, deploying more AI accelerators in the same cabinet or node to boost the efficiency of individual training tasks. Scale-out connects more cabinets and servers, expanding the data center into larger compute clusters to improve overall capacity and throughput.

Training large language models heavily relies on Tensor Parallelism and Expert Parallelism, requiring frequent data exchange within tightly coupled Scale-up Pods. Thus, Scale-up scenarios demand far higher latency and bandwidth than Scale-out. Currently, copper interconnects, due to their low cost, low power consumption, and technological maturity, remain the mainstream solution for intra-cabinet connectivity. In Nvidia’s GB300 NVL72 architecture, high-speed communication between Superchip and switch chips is still mainly reliant on copper cabling.

By contrast, optical interconnects excel in long-distance, high-bandwidth transmission. As single-channel rates rise to 224Gbps and above, optical modules can deliver low-loss transmission over 10 meters or more, and support terabit-level scalability, making them core technology for inter-cabinet connections in Scale-out architectures.

According to LightCounting data, in 2025, global sales of optical transceivers and related products will exceed $23 billion, up about 50% year-over-year; among these, the Ethernet optical transceiver market is about $17 billion, up 60%. The agency expects the Ethernet optical transceiver market to maintain a compound annual growth rate of about 59% from 2024 to 2026; from 2026 to 2030, as the market matures, growth will slow to about 15%.

This means that, the future AI data center connectivity architecture is not about “copper being replaced by optics,” but about a clear division of labor at different levels: copper interconnects continue to dominate short-distance, high-density Scale-up scenarios, while optical interconnects support long-distance, high-bandwidth Scale-out networks. The two technologies will develop in parallel for years to come, forming the core foundation for AI infrastructure expansion.

CPO: Real-world challenges from concept to implementation

Co-Packaged Optics (CPO) integrates optical engines directly onto the board containing the XPU or switch chip, eliminating the DSP found in traditional optical modules, and allowing data transmission via low-power SerDes. This architecture significantly shortens the electrical signal path and is seen as a key direction for next-generation high-speed interconnects.

Nvidia says its CPO switch can deliver about 3.5x energy efficiency improvement, 63x signal integrity enhancement, and 10x network resilience improvement over traditional pluggable optical modules. Broadcom notes that CPO can reduce per-bit optical costs by about 40%.

However, CPO still faces many practical challenges before widespread adoption. Manufacturing yield, testing complexity, fiber coupling accuracy, and concerns about maintainability and supplier concentration among cloud service providers are all important hurdles. Because optical devices are packaged inside the switch, failures usually require replacing the entire switch or sending it back for repair, leading to longer downtime than traditional solutions. In contrast, pluggable optical modules can be quickly swapped on-site by data center operators, with minimal business impact.

Given these constraints, Bernstein expects CPO will start limited deployment in Scale-out networks in the second half of 2026, mainly to validate actual performance and test supply chain maturity. Early adopters will likely include AI cloud service providers like CoreWeave and Lambda.

For more critical Scale-up scenarios, CPO’s entry may be postponed until after the latter half of 2028. This is because the industry needs thorough long-term reliability testing on switches before applying the technology to more valuable and fault-tolerant XPU systems.

LightCounting expects CPO's true large-scale shipments will not occur until after 2028. Before then, Linear Pluggable Optics (LPO) could be a more practical transitional option. LPO eliminates DSPs and lets linear components handle signal processing, reducing power consumption by about two-thirds compared to traditional pluggable modules, while retaining modular design for convenient maintenance.

Bernstein believes LPO shipments will likely exceed CPO's before 2030. This means that over the next few years, the mainstream direction for optical interconnects in data centers won't be a straight jump to co-packaged optics, but a gradual evolution among pluggable, LPO, and CPO architectures.

CPO redefines value distribution: profits shift from module vendors to chips and packaging

The cost structure breakdown for Nvidia’s Quantum-X800 CPO Switch points to a clear conclusion: co-packaged optics technology is profoundly rewriting the rules of value distribution in the industry chain.

According to estimates, this switch is equipped with four switch ASICs, each surrounded by 18 optical engines, for a total of 18 external light source modules. Each light source module contains eight CW (continuous wave) lasers, supplying stable laser input to all optical engines. Under this architecture, the total cost of a single Quantum-X800 CPO Switch is about $570,000.

In terms of pricing structure, the optical engine and laser combo in CPO has an average selling price (ASP) at least 10% higher than the 1.6T pluggable optical module. This comparison accounts for the traditional optical module vendor’s gross margin of about 40%, and CPO system vendors’ gross margin of about 50%. In other words, CPO not only does not reduce overall value, but also creates greater value per unit through higher integration.

More importantly, CPO shifts the profit center in the industry chain. For traditional pluggable optical modules, value is mainly concentrated with module vendors; DSPs and other electrical chips are key cost components. In CPO architecture, DSPs are eliminated, and optical engines are co-packaged directly with the switch chip, shifting the value focus to chip design, advanced packaging, and wafer fabrication.

This means that Nvidia, Broadcom, TSMC, and various OSAT (Outsourced Semiconductor Assembly and Test) vendors will be core beneficiaries in the CPO era. At the same time, upstream key component suppliers will also share in the growth, including optical device companies like Lumentum and Coherent, and test equipment vendors like Chroma ATE.

By contrast, traditional optical module vendors’ roles will be structurally weakened in CPO and future Near-Packaged Optics (NPO) architectures. As packaging and system integration become core competitive strengths, industry profits will no longer be concentrated in transceiver assembly, but will shift to companies with chip design, advanced manufacturing, and system integration capabilities.

Risk disclaimerThe market carries risks, and investments should be made with caution. This article does not constitute personal investment advice and does not take into account individual users' special investment objectives, financial circumstances, or needs. Users should consider whether any opinions, views, or conclusions in this article suit their particular circumstances. If you invest accordingly, you bear responsibility.