AI connectivity has become a new bottleneck, and the coexistence of CPOs and copper cables is favorable for leading supply-chain players.

Overview

This in-depth research report highlights that the bottleneck in AI data centers has shifted from raw computational power to data movement—specifically, connectivity. As AI cluster sizes continue to grow, electrical and optical interconnects have assumed critical strategic importance in both scale-up (in-node expansion) and scale-out (inter-cluster expansion) networks. The report provides a detailed analysis of the technological prospects and adoption timeline for Co-Packaged Optics (CPO), along with its implications for the supply chain, while underscoring the continued dominance of copper cables in the near term. Furthermore, it delves into the surge in demand for PCBs, HDI substrates, and ABF laminates driven by the AI wave, as well as the evolving supply dynamics of key upstream materials, concluding that technology leaders with deep manufacturing capabilities and robust ecosystem control will capture the lion’s share of value.

Core views

Connectivity Has Become the New Battleground: AI data center architectures are shifting from being “compute‑limited” to being “connectivity‑limited.” The movement of data within and between racks has emerged as the primary bottleneck, making electrical (copper) and optical interconnects critical frontiers for both performance and capital expenditure. NVIDIA’s roadmap confirms that copper and optical interconnects will coexist, each optimized for distinct distance and power‑consumption requirements. The Gradual Adoption of CPO: Co-Packaged Optics (CPO), by integrating optical engines close to the chip, delivers superior energy efficiency and signal integrity, yet faces challenges related to cost, yield, test complexity, and maintainability. Our research suggests that CPO will begin small‑scale deployments in scale‑out networks starting in the second half of 2026, primarily to gather market feedback and validate the supply chain. However, given the advantages of pluggable optical modules—particularly in terms of serviceability and ecosystem maturity—cloud service providers (CSPs) are likely to continue favoring pluggable solutions over the next two to three years. Full‑scale adoption may not materialize until after 2028. The Enduring Vitality of Copper Cables: In scale‑up scenarios—such as intra‑rack connectivity—copper cables are expected to remain mainstream through 2026–2028, thanks to their low power consumption, high reliability, and cost‑effectiveness. Co‑Packaged Copper (CPC) technology, which integrates high‑speed copper cable connectors directly into ASIC or switch packages, eliminates PCB trace losses, rendering 448 Gbps‑class copper links technically feasible and extending the lifecycle of copper cabling. This trend will benefit suppliers such as Luxshare Precision. Structural Growth in PCBs and Substrates: The demand from AI servers for higher signal speeds and densities is driving structural growth in multilayer printed circuit boards (MLPCBs), high‑density interconnect (HDI) boards, and ABF substrates. For instance, in GB200/GB300 rack configurations, the per‑GPU value of HDI‑plus‑PCB content has doubled from $100–$150 to $300. Supply constraints on key materials—such as low‑dielectric constant glass fiber (T‑glass), ABF films, and high‑frequency copper foils—are reinforcing the near‑term earnings momentum of related suppliers, including ASE Technology, Ibe Electronics, and Ajinomoto. Reconfiguration of the Supply Chain Value Chain: The transition to CPO will reshape the value chain, favoring technology leaders with deep manufacturing capabilities and strong ecosystem control. TSMC’s COUPE technology is poised to become the dominant platform for CPO, as it better integrates advanced logic nodes with advanced packaging (CoWoS). Meanwhile, traditional optical module manufacturers may see their roles structurally diminished in the CPO/NPO era, while suppliers of lasers (e.g., Lumentum), fiber array units (FAUs), and test equipment stand to capture new growth opportunities.

Analysis framework

The research report adopts an analytical framework that integrates the “technology evolution path” with a “supply chain value mapping.” First, by comparing the two AI cluster‑scaling approaches—Scale‑up and Scale‑out—it delineates the application scenarios and physical limitations of different interconnect technologies (copper vs. optical). Second, it conducts an in‑depth analysis of the strengths and weaknesses of CPO technology (energy efficiency versus maintainability and cost), and, drawing on the product roadmaps of industry leaders such as NVIDIA and Broadcom, projects its commercialization timeline. Finally, spanning the upstream and downstream segments of the “chip‑package‑board‑materials” value chain, it quantifies the incremental value added at each stage (e.g., the PCB value per GPU doubling) and identifies niche market leaders that, owing to technological barriers or supply bottlenecks, have gained pricing power.

Methodology notes

• Industry/Industrial Analysis FrameworkDownstream, midstream, and upstream transmission within the industrial chain

  Starting from the AI chip computing power demand, we trace the implications to the interconnect bandwidth bottleneck and further analyze how this cascades into shifts in demand and value allocation across PCBs, substrates, optical modules, and upstream materials.

  The research report illustrates how the expansion of AI computing power cascades through the hardware infrastructure: increased compute capacity drives a surge in data traffic, which in turn exposes bottlenecks in interconnect bandwidth, necessitating more advanced interconnection technologies (CPO and high-speed copper cables), thereby boosting both the volume and prices of high‑tier PCBs, HDIs, ABF substrates, and specialty materials (T‑glass and ABF films). This top‑down transmission analysis helps investors identify the segments with the greatest upside potential along the value chain.

• Competition and Strategic FrameworkMoat / competitive advantage

  It emphasizes that technology leaders have erected competitive barriers through manufacturing depth, ecosystem control, and access to scarce materials.

  The research report notes that the high manufacturing complexity and testing challenges of CPO favor technology leaders such as TSMC, while the concentrated supply structure of T‑glass and ABF substrates—with Ajinomoto virtually monopolizing the ABF substrate market—endows these upstream material suppliers with robust pricing power and earnings stability. This underscores a competitive‑advantage analysis rooted in technological barriers and resource scarcity.

• Valuation MethodPE/PEG valuation

  We value the covered companies and set target prices using the price-to-earnings (P/E) ratio and the PEG ratio.

  In valuing companies such as Unimicron Technology and I-PEX, the research report explicitly employed target P/E multiples—e.g., 32x for Unimicron’s 2027E EPS and 30x for I-PEX’s Q5–Q8 EPS—to derive target prices, a common relative valuation approach in growth‑stock analysis.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• NVIDIA (NVDA)

  A leader in AI chips and network architectures, driving the development of CPO and copper-cable standards.

  Strengths

  Boasting a comprehensive AI software-and-hardware ecosystem, the Quantum and Spectrum series switches are leading the way in CPO deployments.

  Comparison

  While Broadcom has co-led the early CPO market, NVIDIA favors an NPO architecture to enhance maintainability.

• Broadcom (AVGO)

  A networking chip giant and a pioneer in CPO switches.

  Strengths

  The Tomahawk series of switches has been upgraded to its third-generation CPO and maintains close collaboration with CSPs such as Meta.

  Comparison

  Compared with NVIDIA, Broadcom’s CPO solution leans more toward true co-packaging (non‑detachable).

• TSMC (TSM)

  A leader in CPO manufacturing platforms, offering COUPE and CoWoS technologies.

  Strengths

  The seamless integration of advanced logic processes and cutting-edge packaging, coupled with robust ecosystem control.

  Weaknesses

  CPO’s direct revenue contribution is relatively modest in the early stages.

  Comparison

  Compared with GlobalFoundries and Tower, TSMC enjoys a distinct advantage in integrating EIC/PIC and CoWoS technologies.

• Unimicron (3711.TT)

  A leading supplier of ABF substrates and HDI, benefiting from demand for AI servers.

  Strengths

  It holds a leading share in the ASIC substrate market, and Taiwan’s HDI capacity utilization is improving.

  Weaknesses

  Restoring the gross margin to above 30% will be challenging, as the rise in ASP is primarily driven by raw material costs.

  Comparison

  Compared with Ibiden, Unimicron holds a larger market share in the ASIC segment, while Ibiden is the leader in NVIDIA GPU substrates.

  Risks

  Intensifying competition is putting downward pressure on gross margins.

• Ibiden (4062.JP)

  NVIDIA’s core supplier of ABF substrates for high-end GPUs

  Strengths

  The substrate value of the Rubin platform has doubled, and EMIB-T technology provides additional upside potential.

  Weaknesses

  Non-AI businesses weighed on overall growth.

  Comparison

  It holds a leading position in the NVIDIA GPU substrate market, with high technological barriers.

  Risks

  Weak demand for Intel chips

• Ajinomoto (2802.JP)

  ABF film is supplied by a near-monopolistic vendor.

  Strengths

  95% market share, high profit margin (54%), and strong pricing power.

  Comparison

  As an upstream materials supplier, it stands to benefit from the surge in AI demand even earlier than downstream substrate manufacturers.

• Luxshare (002475.SZ)

  GB300 Copper Cable Connects with a New Supplier

  Strengths

  Entering NVIDIA’s supply chain, benefiting from CPC technology to extend the lifespan of copper cables.

  Comparison

  Competing with Amphenol, but potentially offering advantages in cost efficiency and response speed.

• Chroma ATE (2360.TW)

  CPO photonics test equipment supplier

  Strengths

  It has accumulated extensive experience over the long term in laser diode and photonic component testing.

  Comparison

  Compared with traditional electrical test equipment vendors, Chroma boasts greater expertise in the optoelectronic hybrid testing domain.

  Risks

  CPO production ramp-up has fallen short of expectations.

Key data

• Ethernet Optical Transceiver Market Size (2025)~US$17 billionYear-on-year growth of 60%, accounting for approximately 74% of the total optical transceiver market.
• Total Market Size of Optical Transceivers (2025)>US$23 billionA year-on-year increase of 50%
• AI Server HDI + PCB Single-GPU Value~$300Compared to the Hopper server, the price has doubled (from the original $100–$150).
• Estimated cost premium for CPO switching~10%Compared with pluggable optical modules, after factoring in system gross margin, the upfront cost of CPO is at least 10% higher.
• NVIDIA GPU substrate market growth rate127% (FY26/3)Driven by the migration to the Rubin platform, the CAGR over the subsequent two years is expected to be 44%.
• Ajinomoto ABF Business Profit Margin54%The group’s most profitable segment accounts for nearly one-third of operating profit.

Impact & implications

For semiconductor and hardware manufacturers, the advancement of AI‑related connectivity translates into higher average selling prices (ASPs) and more complex manufacturing processes. TSMC will consolidate its dominant position in advanced packaging through COUPE and CoWoS technologies; although CPO’s direct revenue contribution will be limited in the early stages, it will nonetheless boost both front‑end wafer fabrication and packaging operations. For PCB and substrate suppliers such as Unimicron and Ibiden, the increased share of high‑end products driven by AI will enhance their product mix, though they should remain vigilant about margin pressures stemming from intensifying competition after 2026. Upstream materials providers—like Ajinomoto and Nitto Denko—benefit from strong pricing power due to supply constraints, making them relatively certain winners in the near term. As for optical module manufacturers, while CPO may erode the standalone module value over the long run, plug‑and‑play modules—including transitional LPO/NPO solutions—will continue to dominate in the medium term, with iterative upgrades to high‑speed modules (1.6T/3.2T) sustaining revenue growth.

Risks

• The adoption of CPO technology has been slower than expected, constrained by yield issues, testing complexity, and maintainability challenges.
• Slower growth in AI capital expenditures is weighing on demand across the broader hardware supply chain.
• Prolonged supply bottlenecks for critical raw materials, such as T-glass and InP substrates, are constraining the ramp-up of production capacity.
• Intensified competition in the PCB and substrate industries is expected to constrain gross margin expansion after 2026.
• Geopolitical risks are undermining supply chain stability.

What to watch

• The mass-production timeline for NVIDIA’s Rubin platform, as well as the final design choices for its midplane and backplane.
• Feedback from the small-scale deployment of CPO in 2H26, along with signs of large-scale rollout in 2027–2028.
• The progress of capacity expansion and price fluctuations for key materials such as T-glass.
• Changes in the procurement preferences of major CSPs (such as Meta, Google, and Microsoft) for CPOs and pluggable optical modules.
• The penetration rate of Intel’s EMIB‑T technology in AI ASIC packaging.