AI data-center optical communications supply chain Report Interpretation

Overview

Based on CIOE 2026 forums and meetings across the optical-communications supply chain, Nomura argues that AI data-center optical-network demand remains strong. Its central view is that 1.6T transceivers will remain the near-term growth pillar, while shortages and more complex next-generation architectures raise barriers that favor leading suppliers.

Core views

Nomura attended CIOE 2026 in Shenzhen on 9-10 September and met companies across optical transceivers, components, electronic and optical chips, fiber, materials and equipment. It left the event more confident in the AI data-center optical-network demand outlook. The report argues that suppliers able to remain ahead of the technology curve should benefit from the upcycle, and identifies Zhongji InnoLight, Accelink, TFC, YOFC and Shenzhen T&S as preferred Buy-rated names. The immediate growth driver is the 1.6T transceiver, which Nomura says has already become the mainstream product in scale-out networks. Better automation could improve assembly efficiency and support faster capacity expansion at transceiver manufacturers. However, component availability remains the critical bottleneck. Nomura expects incumbent 1.6T makers, particularly InnoLight, to retain leadership because effective supply-chain management is important when constrained inputs limit industry output. The report expects the 1.6T cycle to last longer because 2.4T and 3.2T transceivers face substantial challenges in the 400G-per-lane era on both the optical and electronic sides. Rather than following a conventional silicon-photonics path, higher-speed products may require new materials such as thin-film lithium niobate (TFLN) or modulation approaches such as Coherent Lite. Nomura notes progress in TFLN modulators by Murata and unlisted Liobate, but presents these technologies as part of an unresolved upgrade path rather than an immediate replacement for 1.6T. NPO was a major conference focus. Suppliers displayed several approaches, including Hisilicon's VCSEL-based and internal-light-source CPO products, CIG's silicon-photonics CPO using an external light source, and Accelink's 6.4T NPO solution for scale-up networks. Nomura emphasizes that industry-wide NPO standards have not yet emerged. It believes early movers may earn higher content value and margins through customized solutions, but only if they secure orders from large global AI customers during the initial commercialization phase. Supply constraints remain central to the thesis. While 70mW CW lasers and 100G EML/PD chips are already in mass production and are supporting 800G transceiver shipment growth, 100mW-plus CW lasers and 200G EML/PD suppliers still need to improve yields to close the gap for 1.6T transceivers and NPO products. Nomura also identifies 3nm DSPs, mSAP-technology PCBs, multicore fiber and high-precision ferrules as potential new choke points. These constraints could cause lower-than-expected high-end transceiver shipments, but they may also allow domestic suppliers to gain global share. Nomura sees Chinese optical-communications companies as important participants in the global AI data-center supply chain. It notes that InnoLight, Eoptolink and TFC already hold leading positions in transceivers or components, while upstream materials, lasers and equipment retain room for domestic substitution. Examples include InP and SiGe substrates, high-end EML/CW laser and PD chips, multicore and hollow-core fibers, and automated transceiver coupling and testing equipment. OCS is presented as another potentially important network technology because it could be adopted across several network segments rather than remaining limited to Google's scale-out network. Cigna AI projects OCS revenue to exceed USD2bn in 2026 and USD8bn by 2030E, with scale-up OCS tied to GPUs potentially exceeding scale-out OCS applications by 2030E. The report cites liquid-crystal OCS from Coherent, MEMS-based OCS from Accelink and silicon-photonics OCS from Taclink, but stresses that architecture remains fragmented and that the full supply chain must advance commercialization. Finally, Nomura highlights NVIDIA's recently discussed scale-in network as a potential future driver for AI networking. NVIDIA describes the architecture as powered by BlueField-4 and connected through Spectrum-X Ethernet. Nomura believes scale-in networking could increase demand for AI-network access, security, data movement and infrastructure operations if AI factories become widely adopted in enterprise environments.

Analysis framework

Nomura combines CIOE 2026 company meetings and technology-forum observations with a supply-chain assessment. It traces demand from AI data-center networking through transceiver speed upgrades, component yields and shortages, then evaluates NPO, CPO, OCS and scale-in architectures as possible future demand and value-content drivers.

Methodology notes

• Industry AnalysisSupply-demand framework

  Supply-demand assessment of AI optical-network products and their critical components.

  The report links strong AI data-center demand with shortages in lasers, EML/PD chips, DSPs, PCBs and fiber inputs to explain both shipment constraints and supplier advantages.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI networking demand transmitted through transceivers, components, materials, fiber and equipment.

  Nomura assesses how technology upgrades and bottlenecks affect different positions in the optical-communications supply chain, including potential domestic substitution upstream.

Asset mapping & comparison

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

• Zhongji InnoLight (300308 CH / 3308 HK)

  Preferred high-end optical-transceiver and NPO-solution provider.

  Strengths

  Nomura views incumbent 1.6T transceiver makers as advantaged by supply-chain management; InnoLight is specifically cited as a leading incumbent.

  Comparison

  Named alongside Accelink as a preferred optical-networking supplier.

  Risks

  Component shortages may limit high-end product shipments.

• Accelink (002281 CH)

  Preferred high-end optical-transceiver and NPO supplier.

  Strengths

  Showcased a 6.4T NPO solution for scale-up networks.

  Weaknesses

  NPO standards and commercialization remain unsettled.

  Comparison

  Competes across optical transceiver, NPO and OCS technology paths.

  Risks

  Early NPO value capture depends on securing orders from large global AI customers.

• TFC (300394 CH)

  Preferred NPO/CPO component maker.

  Strengths

  Already holds a leading position in the transceiver/component sector, according to Nomura.

  Comparison

  Part of the domestic component group positioned for AI data-center optical-network demand.

  Risks

  High-end component constraints could restrict product shipments.

• YOFC (6869 HK)

  Preferred high-end AI data-center fiber and fiber-connector supplier.

  Strengths

  Named as a producer of multicore and hollow-core fiber.

  Comparison

  Participates in upstream fiber opportunities alongside other domestic fiber suppliers.

  Risks

  Multicore fibers and high-precision ferrules are potential supply-chain choke points.

• Shenzhen T&S (300570 CH)

  Preferred high-end AI data-center fiber and fiber-connector supplier.

  Strengths

  Included among Nomura's preferred beneficiaries of the optical-network upcycle.

  Comparison

  Named alongside YOFC in high-end AI data-center fiber and connector exposure.

• Coherent (COHR US)

  Example of a liquid-crystal OCS solution provider.

  Strengths

  Displayed a liquid-crystal OCS approach at CIOE 2026.

  Weaknesses

  OCS architecture remains fragmented.

  Comparison

  Its liquid-crystal approach differs from MEMS-based OCS from Accelink and silicon-photonics OCS from Taclink.

  Risks

  OCS commercialization requires further supply-chain progress.

• NVIDIA (NVDA US)

  Architecture driver for the scale-in AI-network concept.

  Strengths

  Its scale-in network is described as using BlueField-4 and Spectrum-X Ethernet.

  Comparison

  Scale-in is presented as an additional AI-network pillar beyond scale-up, scale-out and scale-across approaches.

  Risks

  The demand effect depends on broad enterprise adoption of AI factories.

Key data

• CIOE 2026 attendance9-10 September 2026Nomura attended forums and met dozens of companies in Shenzhen.
• Mainstream transceiver speed1.6TIdentified as the mainstream product in scale-out networks.
• Next-generation transceiver speeds2.4T / 3.2TFace significant technology challenges in the 400G-per-lane era.
• CW laser and EML/PD status70mW CW laser and 100G EML/PDAlready in mass production and supporting 800G transceiver shipment growth.
• OCS market revenue forecastOver USD2bn in 2026; over USD8bn by 2030ECited from Cigna AI's presentation.

Impact & implications

Nomura's conclusions favor suppliers with established technology, supply-chain control and customer relationships as 1.6T demand expands under component constraints. The report also identifies domestic substitution in upstream materials, lasers and equipment, plus NPO/CPO, OCS and scale-in networking, as areas with potential strategic importance as AI-network architectures evolve.

Risks

• Yield improvement delays for 100mW-plus CW lasers and 200G EML/PD chips could leave a supply-demand gap for 1.6T transceivers and NPO products.
• Shortages in high-end components, including 3nm DSPs, mSAP PCBs, multicore fibers and high-precision ferrules, could lead to lower-than-expected high-end transceiver shipments.
• NPO and OCS architectures remain fragmented, which may delay broader commercialization.

What to watch

• Progress in domestic substitution across optical materials, lasers and equipment.
• NPO and CPO penetration and the emergence of industry standards.
• OCS commercialization and adoption across scale-up and scale-out networks.
• Development of scale-in networks and enterprise adoption of AI factories.
• Yield improvement and supply availability for high-end lasers, EML/PD chips, DSPs, PCBs and multicore fiber.