AI networking: scale-across and scale-in infrastructure: Scale-across AI networking could create a major new market for optical interconnects

Overview

This Global AI Trend Tracker examines how increasingly distributed AI training and inference may reshape networking architecture. Nomura sees scale-across networks linking multiple data centers, and scale-in infrastructure connecting enterprise data to AI compute, as new sources of demand for high-performance optical and networking equipment.

Core views

Nomura argues that conventional scale-up and scale-out architectures are increasingly constrained as AI models exceed trillions of parameters and inference workloads grow exponentially. Scale-up links GPUs within a node or rack, while scale-out links cabinets within one data center. Scale-across instead connects geographically distributed data centers into a unified computing pool capable of running a single AI task. In Nomura's view, this becomes necessary when a single campus reaches power, cooling, land and topology limits. The report distinguishes scale-across from traditional DCI. Traditional DCI primarily supports synchronization, backup and content distribution at 100G-400G, whereas distributed AI training requires materially greater bandwidth, lower latency and lower packet-loss rates. Cisco estimates that connecting 1mn XPUs through scale-across requires roughly 14 times the bandwidth of WAN/DCI networks. The relevant architecture uses 800G/1.6T coherent transceivers, hollow-core fiber, DCI/WAN/metro routers and switches to create a cross-site optical backbone. Nomura identifies coherent optical modules, optical circuit switches (OCS) and hollow-core fiber as the principal enabling technologies. IM-DD modules have transmission distances below 2km, while coherent modules can extend to 80-120km because coherent transmission manages both light amplitude and phase and compensates for long-distance signal degradation. Coherent-lite, first proposed by Google in 2021, uses a simplified DSP and offers 2-20km transmission at lower cost and power consumption; Nomura expects it to be relevant across scale-out, scale-across and DCI applications. The report also says OCS can enable flexible, high-bandwidth all-optical switching over long distances, while hollow-core fiber reduces latency because it transmits through air, inert gas or vacuum rather than glass. FiberHome cites 31% lower latency for its new-generation ultra-low-loss hollow-core fiber versus traditional solid-core fiber. The report describes multi-rail photonic line systems as another architectural response to the density, power and operational burden of connecting hundreds of fiber pairs. Such systems use multiple high-capacity rails with dedicated amplification, monitoring and control. Ciena expects multi-rail technology to increase single-rack fiber-pair capacity by 32 times, which Nomura believes would support demand for high-speed coherent transceivers and dense optical amplifiers. Nomura also notes that combining coherent-lite with OCS can recover link-budget margin consumed by OCS and potentially extend optical-switching architectures into the 2.4T/3.2T module era. Scale-in is presented as a separate pillar of AI networking, alongside context memory, scale-up, scale-out and scale-across. It shifts user-access, data-handling, storage, security-isolation and related north-south traffic to DPUs and Ethernet at the AI-factory boundary. Nomura expects its importance to rise in the Agentic AI era because AI agents increase data-transfer, storage and multi-round communication complexity; CPUs otherwise spend substantial resources on storage and security tasks; and sensitive sectors such as finance and healthcare need encryption and isolation before data reaches GPU/CPU clusters. NVIDIA's BlueField-4, DOCA and Spectrum-X Ethernet underpin this approach. The report says BlueField-4 offers hardware-level Zero Trust isolation, multi-tenant storage offloading, 18x AI Factory service bandwidth and 2x storage-acceleration performance. Market data supports Nomura's growth thesis. 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 alone at USD8-10bn by 2029E. For a 500,000-GPU cluster, Ciena estimated scale-across TAM of roughly USD500mn for an 80km span requiring more than 400 fiber pairs, or about USD1,000 per GPU excluding amplifiers; the estimate rises to USD1bn at 1,000km and USD2bn at 2,000km as higher-performance modems are needed. Cignal AI estimates the 800G ZR/ZR+ market will grow at 145% CAGR in 2025-29E, while the OCS market rises from about USD400mn in 2025 to more than USD2.5bn in 2029E, or roughly 58% CAGR. Coherent raised its 2030E global OCS market guidance from USD2bn to USD4bn. Nomura sees North America as the more immediate scale-across market because individual data-center power and land constraints are driving CSP deployment. AWS had deployed hollow-core fiber to 10 core data centers by 2026; Microsoft plans 15,000km across Azure's global network by end-2026, while Google and Meta have begun testing. China faces similar single-site power and heat-dissipation limits, but Nomura considers scale-across a relatively small part of the domestic market and says large-scale demand has not yet reached an inflection point. China Mobile's GSE-DCI demonstration delivered over 98% per-node training-compute efficiency over 100km, while ZTE said its scale-across solution improved end-to-end latency by about 33% and reduced supporting equipment, modules and fiber by about 40%. The report maps the industry chain from upstream optoelectronic chips, components, new fibers and transceivers, through networking-equipment makers, to downstream CSPs, IDC providers and telecom operators. It expects YOFC to benefit from hollow-core-fiber commercialization, and Zhongji InnoLight and Accelink from demand for high-speed coherent transceivers. Competition remains concentrated in several layers: Marvell and Acacia lead 400G+ coherent components; Ciena and Nokia are expected to begin 800ZRx shipments; Google, Coherent and Lumentum are identified as key OCS players; and Corning, YOFC and Zhongtian Technology held reported 2025 global fiber-cable shares of about 19.52%, 14.57% and 11.43%, respectively.

Analysis framework

Nomura first compares scale-up, scale-out, scale-across and scale-in network roles, then links data-center constraints to required network performance. It evaluates the relevant optical technologies and value chain, uses company and third-party market estimates to size demand, and compares North American and Chinese adoption progress and competitive positioning.

Methodology notes

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Scale-across AI-networking value-chain analysis

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

• Industry AnalysisSupply-demand framework

  TAM and adoption-demand analysis

  Nomura uses projected market sizes, growth rates, deployment requirements and technology adoption evidence to explain potential demand for AI-networking hardware.

Asset mapping & comparison

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

• YOFC (6869 HK)

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

  Strengths

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

  Weaknesses

  Scale-across hollow-core-fiber volumes are expected to remain relatively small over the next two to three years.

  Comparison

  Reported 2025 global fiber-cable market share of about 14.57%, behind Corning's 19.52%.

  Risks

  Weaker China telecom-operator demand, slower AI-networking expansion or overseas expansion, and slower penetration of high-end AI products.

• Zhongji InnoLight (300308 CH)

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

  Strengths

  Exposure to high-end optical-module demand associated with AI networking.

  Comparison

  Operates in a market where Marvell, Acacia, Ciena and Nokia are notable coherent-optics competitors.

  Risks

  Weaker datacom and telecom demand, competition in 400G and 800G modules, slower 800G/1.6T upgrades and export-related price pressure.

• Accelink (002281 CH)

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

  Strengths

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

  Comparison

  The company is among early domestic suppliers of 1.6T coherent optical transceivers.

  Risks

  Lower optical-component demand, slower optical-chipset R&D, pricing competition and margin dilution, and technology-sector sanctions.

Key data

• AI-WAN TAMUSD7bn to USD14bnCiena projection for 2026-29E, representing about 30% CAGR
• Scale-across TAMUSD8-10bnCiena management expectation for 2029E
• 800G ZR/ZR+ market CAGR145%Cignal AI estimate for 2025-29E
• OCS marketAbout USD400mn in 2025; above USD2.5bn in 2029ECignal AI forecast, implying about 58% CAGR over four years
• AI fiber-cable demand growth77% y-y in 2025E; 26% five-year CAGR in 2025-29ECRU estimate
• Scale-across bandwidth requirementAbout 14x WAN/DCICisco estimate for connecting 1mn XPUs

Impact & implications

Nomura's thesis is that geographic distribution of AI computing moves optical networking from a conventional DCI role toward a higher-performance AI infrastructure layer. The report identifies coherent transceivers, OCS, hollow-core fiber, multi-rail systems and DPU-led scale-in infrastructure as the principal beneficiaries of this transition.

Risks

• Demand for optical components, modules or fiber from datacom and telecom markets could be weaker than expected.
• Product development or upgrades in optical chipsets, 800G and 1.6T technologies could progress more slowly than expected.
• Intensified pricing competition could pressure margins and exports.
• Technology-sector sanctions could worsen.