The long-term competition between copper and optical interconnects within AI data centers is reaching a critical juncture, yet the narrative of a simple "copper retreat, optical advance" is far from the complete picture. A recent Bernstein white paper highlights that as AI cluster scale rapidly expands, connectivity is becoming the core bottleneck affecting system performance and cost, emerging as a new focal point for industry competition. For many years to come, copper and optical interconnects will not be a simple substitution relationship but will coexist long-term across different distances and application scenarios, evolving separately along "Scale-up" and "Scale-out" development paths. NVIDIA and Broadcom are driving CPO (Co-Packaged Optics) from concept towards commercialization. NVIDIA has explicitly stated that its CPO switches will see small-scale deployment in the second half of 2026, with AI cloud service providers like CoreWeave and Lambda among the first users. However, due to high manufacturing, packaging, and testing complexity, hyperscale cloud providers will still primarily rely on pluggable optical modules between 2026 and 2028. Concurrently, copper interconnects, leveraging advantages in cost, power consumption, and maturity, will maintain their dominant position in Scale-up scenarios for at least the next three years. The true significance of CPO lies in its potential to restructure value distribution across the supply chain. Bernstein estimates that the cost of a CPO optical engine and laser combination is approximately 10% higher than the average selling price of a 1.6T pluggable module, but the profit center of gravity will shift from traditional optical module manufacturers towards chip fabrication and advanced packaging. This technological upgrade will also drive structural growth in foundational materials sectors such as high-end PCBs, ABF substrates, and T-glass fiber. However, with the concentrated release of new production capacity, price competition, rising depreciation, and supply expansion are expected to gradually compress profit margins starting from late 2026. Investors should focus on companies demonstrating leading capabilities in technology, manufacturing, and supply chain control. Copper and Optical: Distinct Roles for Scale-up and Scale-out The expansion of AI infrastructure primarily unfolds along two paths: Scale-up and Scale-out. Scale-up involves increasing computing resources within a single system, such as deploying more AI accelerators within the same rack or node, to enhance the computational efficiency of an individual training job. Scale-out involves connecting more racks and servers to expand a data center into a larger-scale computing cluster, thereby increasing overall capacity and throughput. Large Language Model training heavily relies on techniques like Tensor Parallelism and Expert Parallelism, requiring frequent data exchange within tightly coupled Scale-up pods. This makes latency and bandwidth requirements in Scale-up scenarios far more demanding than in Scale-out. Currently, copper interconnects, benefiting from lower cost, lower power consumption, and technical maturity, remain the mainstream solution for intra-rack connections. In NVIDIA's GB300 NVL72 architecture, high-speed communication between Superchips and switch chips still primarily relies on copper cables. In contrast, optical interconnects hold a more pronounced advantage in long-distance, high-bandwidth transmission. As per-lane data rates increase to 224Gbps and beyond, optical modules can achieve low-loss transmission over distances of 10 meters or more, supporting terabit-level scaling. Consequently, they have become the core technology for inter-rack connections in Scale-out architectures. Data from LightCounting shows that global sales of optical transceivers and related products exceeded $23 billion in 2025, representing year-over-year growth of approximately 50%. Within this, the Ethernet optical transceiver market reached about $17 billion, growing 60% year-over-year. The firm forecasts that the Ethernet optical transceiver market will maintain a compound annual growth rate of around 59% from 2024 to 2026. From 2026 to 2030, as the market matures, the growth rate is expected to moderate to roughly 15%. This indicates that the future connectivity architecture for AI data centers is not one of "copper being replaced by optics," but rather one of clear division of labor across different tiers: copper interconnects will continue to dominate short-distance, high-density Scale-up scenarios, while optical interconnects will underpin long-distance, high-bandwidth Scale-out networks. Both technologies will develop in parallel for many years to come, jointly forming the core foundation for AI infrastructure expansion. CPO: Real-World Challenges from Concept to Implementation Co-Packaged Optics (CPO) integrates the optical engine directly onto the substrate alongside the XPU or switch chip, eliminating the DSP found in traditional optical modules and allowing data to be transmitted directly via lower-power SerDes. This architecture significantly shortens the electrical signal path and is viewed as a key direction for next-generation high-speed interconnects. NVIDIA states that its CPO switch can achieve approximately 3.5x improvement in energy efficiency, a 63x improvement in signal integrity, and a 10x enhancement in network resilience compared to traditional pluggable optical modules. Broadcom notes that with CPO adoption, the per-bit optical cost is expected to decline by about 40%. However, CPO still faces numerous real-world challenges on the path to widespread adoption. Manufacturing yield, testing complexity, fiber coupling precision, and cloud service providers' concerns regarding maintainability and supplier concentration are significant obstacles. Since optical components are packaged within the switch, a fault typically necessitates replacing the entire switch or returning it for factory repair, leading to significantly longer downtime compared to traditional solutions. In contrast, pluggable optical modules can be quickly replaced on-site by data center operations personnel, minimizing business impact. Given these constraints, Bernstein anticipates that small-scale deployment of CPO in Scale-out networks will commence in the second half of 2026, primarily to validate real-world performance and test supply chain maturity. Early adopters are expected to include AI cloud service providers like CoreWeave and Lambda. In the more critical Scale-up scenarios, CPO adoption may be further delayed until after the second half of 2028. The reason is that the industry needs first to thoroughly validate long-term reliability in switches before applying this technology to higher-value, more fault-intolerant XPU systems. LightCounting expects that truly large-scale shipments of CPO will not occur until after 2028. Before then, Linear Pluggable Optics (LPO) may serve as a more practical transitional solution. LPO eliminates the DSP, offloading signal processing to linear components, which can reduce power consumption by about two-thirds compared to traditional pluggable modules while retaining the maintenance convenience of a modular design. Bernstein believes that by 2030, LPO shipment volumes could surpass those of CPO. This suggests that the mainstream direction for data center optical interconnects in the coming years is not an immediate leap to co-packaging, but rather a gradual evolution among pluggable, LPO, and CPO architectures. CPO Reshapes Value Allocation: Profits Shift from Module Makers to Chips & Packaging A breakdown of the cost structure for NVIDIA's Quantum-X800 CPO Switch clearly points to one conclusion: Co-Packaged Optics technology is profoundly rewriting the rules of value distribution across the industry chain. Based on estimates, this switch is configured with 4 switch ASICs. Each ASIC is surrounded by 18 integrated optical engines, with the entire system equipped with 18 external light source modules. Each light source module contains 8 Continuous Wave (CW) lasers, providing stable laser input to all optical engines. Under this architecture, the total estimated cost for a single Quantum-X800 CPO switch is approximately $570,000. From a pricing structure perspective, the combined optical engine and laser components in CPO carry an Average Selling Price (ASP) at least 10% higher than that of a 1.6T pluggable optical module. This comparison already accounts for the approximately 40% gross margin typical for traditional optical module manufacturers and the roughly 50% gross margin for CPO system manufacturers. In other words, CPO does not reduce overall value; instead, it creates greater per-unit value through higher integration. More importantly, CPO alters where profits are captured within the supply chain. The value in traditional pluggable optical modules is largely concentrated with the module manufacturers, with DSPs and other electrical chips constituting a significant portion of the cost. Under the CPO architecture, the DSP is eliminated, and the optical engine is co-packaged directly with the switch chip. Consequently, the value center of gravity shifts towards chip design, advanced packaging, and wafer manufacturing. This implies that companies like NVIDIA, Broadcom, TSMC, and various OSAT (Outsourced Semiconductor Assembly and Test) providers will become core beneficiaries in the CPO era. Simultaneously, upstream suppliers of key components are also poised to share in the growth, including optical device companies like Lumentum and Coherent, as well as test equipment manufacturers like Chroma ATE. In contrast, the role of traditional optical module manufacturers will be structurally diminished in CPO and future NPO (Near-Packaged Optics) architectures. As packaging and system integration become core competitive advantages, industry profits will no longer be primarily concentrated in the transceiver assembly segment but will instead gravitate towards companies that command chip design, advanced manufacturing, and system integration capabilities.