AI networking scale-across and scale-in optical interconnect infrastructure: Scale-across AI networking could create substantial demand for coherent optics, hollow-core fiber and OCS

Overview

This industry report examines how AI infrastructure may evolve beyond single data-center clusters. Nomura argues that scale-across networks connecting multiple data centers into one computing pool, alongside NVIDIA's scale-in architecture, will raise demand for advanced optical networking hardware and related infrastructure.

Core views

Nomura's central thesis is that AI training models exceeding trillions of parameters and rapidly expanding inference workloads are exposing the power, cooling, land and topology limits of conventional scale-up and scale-out architectures within a single cluster or data center. Scale-across, proposed by NVIDIA in 2025, links geographically distributed data centers into a unified AI computing pool capable of running a single task. This differs from scale-up, which connects GPUs within a node or rack, and scale-out, which adds cabinets within one data center. The report believes scale-across becomes increasingly necessary when AI clusters exceed the physical limits of a single campus. The transition raises networking requirements beyond traditional data-center interconnect. Traditional DCI typically supports synchronization, backup and content distribution at 100G-400G, whereas distributed AI training demands higher bandwidth, lower latency, lower packet loss and high reliability. Cisco estimates that connecting 1mn XPUs through scale-across requires roughly 14 times the bandwidth of WAN/DCI networks. Nomura highlights 800G and 1.6T coherent transceivers, hollow-core fiber, OCS, and DCI/WAN/metro routing and switching as the cross-site backbone. Scale-across reach can extend beyond 500m, up to 20km for synchronous training and 100km or more for asynchronous use, compared with less than 4m typical scale-up links and 100m-2km scale-out links. The report identifies several enabling technologies. Coherent modules are more suitable than IM-DD modules for long-distance links because IM-DD transmission is below 2km while coherent modules can extend 80-120km. Coherent-lite, initially proposed by Google in 2021, uses simplified DSPs to offer lower cost and power consumption for 2-20km links; Nomura expects it to be used across scale-out, scale-across and DCI settings. Hollow-core fiber replaces glass transmission media with air, inert gas or vacuum, lowering latency and loss; FiberHome cites 31% lower transmission latency than traditional solid-core fiber. OCS offers flexible all-optical switching at high bandwidth, and Nomura believes the combination of OCS and coherent-lite can support scaling into the 2.4T/3.2T module era. Nomura also describes multi-rail photonic line systems as a way to accommodate scale-across capacity that could reach tens of petabits per second across hundreds of fiber pairs without an impractical increase in sites, power and space. Ciena estimates multi-rail can increase single-rack fiber-pair capacity by 32 times, supporting demand for coherent transceivers and high-density amplifiers. NVIDIA's Spectrum-XGS solution, meanwhile, combines software and hardware for scale-across, including ConnectX-8 and BlueField-3 SuperNICs supporting 800Gb/s, 1.6T optical modules and coherent transceivers for geographically dispersed data centers. Scale-in is the report's second major architectural theme. NVIDIA positions it as a north-south boundary layer linking enterprise data to AI compute, where DPUs and Ethernet offload user access, data handling, storage, isolation and security work from CPUs and GPUs. Nomura sees greater relevance in the Agentic AI era because multi-round agent interactions increase transfer and storage complexity and sectors such as finance and healthcare require encryption and isolation before data reaches compute clusters. NVIDIA's BlueField-4 is described as providing hardware-level Zero Trust isolation and multi-tenant storage offload, with 18x AI Factory service bandwidth and 2x storage-acceleration performance. The market evidence supports a rapidly expanding opportunity in Nomura's view. Ciena projects AI-WAN, including network, DCI and scale-across fabrics, to grow from USD7bn to USD14bn in 2026-29E, or about 30% CAGR, and estimates scale-across at USD8-10bn by 2029E. For an 80km span serving a 500,000-GPU cluster, Ciena estimates scale-across TAM of about USD500mn, or USD1,000 per GPU excluding amplifiers; this rises to USD1bn for 1,000km and USD2bn for 2,000km as higher-performance modems replace ZR/ZR+ pluggables. Cignal AI expects 800G ZR/ZR+ to grow at a 145% CAGR from 2025 to 2029E, while the OCS market is forecast to expand from about USD400mn in 2025 to more than USD2.5bn in 2029E, a roughly 58% four-year CAGR. Coherent raised its 2030 global OCS market guidance from USD2bn to USD4bn. Fiber demand is another key link in the argument. CRU estimates AI-workload fiber-optic-cable demand grew 77% year on year in 2025E and forecasts a 26% CAGR for 2025-29E. Nomura notes that a single GW data center may require several million to 10mn fiber-kilometers. North American CSPs are already deploying or testing hollow-core fiber: AWS had deployed it to 10 core data centers by 2026, Microsoft plans 15,000km across Azure's global network by end-2026, and Google and Meta have begun related tests. In China, the report says scale-across remains a relatively small share of the market and has not yet reached a large-scale demand inflection point, although China Mobile and ZTE have introduced relevant technologies and commercial verification. The competitive landscape spans optical DSPs, transceivers, OCS and fiber. Marvell and Acacia lead coherent 400G+ markets, followed by Ciena, Nokia and Huawei, while vertically integrated Ciena, Infinera/Nokia and Huawei develop DSPs for internal use. Google, Coherent and Lumentum are identified as current principal OCS participants. In hollow-core fiber, Nomura identifies YOFC, Hengtong Optoelectronics, Lightera and Relativity Networks as core suppliers. Nomura expects YOFC to benefit from hollow-core-fiber commercialization, and Zhongji InnoLight and Accelink from rising demand for high-speed coherent transceivers.

Analysis framework

Nomura first distinguishes scale-up, scale-out, scale-across and scale-in network architectures, then links AI-cluster constraints to required networking performance. It evaluates the technology stack, market-size estimates, deployment examples, value chain and competitive positions to identify potential beneficiaries.

Methodology notes

• Industry AnalysisSupply-demand framework

  AI networking supply-demand analysis

  The report links rising distributed AI-compute requirements to demand for optical modules, fiber, switches, OCS and related network equipment.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Scale-across value-chain analysis

  The report traces demand from CSPs, IDC providers and telecom operators through networking equipment and optical systems to chips, components, transceivers and fiber suppliers.

• Industry AnalysisVolume-price decomposition

  TAM and capacity-based market sizing

  The report uses GPU-cluster size, route distance, bandwidth needs, fiber pairs and market forecasts to illustrate the opportunity for scale-across equipment.

Asset mapping & comparison

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

• YOFC (6869 HK)

  Nomura expects the company to benefit from hollow-core-fiber commercialization for scale-across networking.

  Strengths

  Mass production of hollow-core fiber; reported deliveries to three telecom operators.

  Comparison

  YOFC ranked second globally in fiber-optic-cable market share in 2025 at about 14.57%, behind Corning at about 19.52%, according to CRU.

  Risks

  Weaker China telecom-fiber demand, slower AI-networking expansion or overseas expansion, and slower penetration of hollow-core fiber could impede its target-price achievement.

• Zhongji InnoLight (300308 CH)

  Nomura expects the company to benefit from increasing demand for high-speed coherent optical transceivers.

  Risks

  Weaker high-end optical-module demand, competition in 400G/800G modules, slower 800G/1.6T upgrades, or price competition could impede target-price achievement.

• Accelink (002281 CH)

  Nomura expects the company to benefit from demand for high-speed coherent optical transceivers.

  Strengths

  Uses Marvell DSP in a 1.6T coherent-lite transceiver for interconnection distances of up to 20km.

  Risks

  Lower demand for datacom and telecom optical components, slower optical-chipset R&D, pricing competition and margin dilution, or worsening technology-sector sanctions could impede target-price achievement.

• NVIDIA (NVDA US)

  NVIDIA introduced the scale-across and scale-in concepts and provides the BlueField and Spectrum-X Ethernet platforms discussed in the report.

  Strengths

  Its architecture positions DPUs and Ethernet as core components of AI-factory networking.

Key data

• AI-WAN TAMUSD7bn to USD14bnCiena projection for 2026-29E, representing about 30% CAGR.
• Scale-across market TAMUSD8-10bnCiena management expectation for 2029E.
• Scale-across bandwidth requirementAbout 14x WAN/DCICisco estimate for connecting 1mn XPUs.
• Coherent-transceiver reach80-120kmVersus less than 2km for IM-DD optical transceivers.
• 800G ZR/ZR+ market CAGR145%Cignal AI estimate for 2025-29E.
• OCS market sizeAbout USD400mn in 2025 to over USD2.5bn in 2029ECignal AI forecast, implying roughly 58% four-year CAGR.
• AI-workload fiber demand growth77% y-y in 2025E; 26% CAGR in 2025-29ECRU estimates.
• Hollow-core fiber latency31% lower than traditional solid-core fiberFiberHome figure for its new-generation ultra-low-loss fiber.

Impact & implications

Nomura believes distributed AI infrastructure shifts network spending toward long-reach, high-bandwidth and low-latency optical systems. The report highlights opportunities for coherent transceiver, OCS and hollow-core-fiber suppliers, while noting that China remains earlier in the adoption curve than the overseas market.

What to watch

• Commercial deployment of scale-across networks by North American CSPs and the pace of hollow-core-fiber rollout.
• Adoption of 800G and 1.6T coherent optics, coherent-lite and OCS in AI data-center interconnection.
• Whether large-scale scale-across demand in China reaches an inflection point.