Global AI infrastructure investment is entering a new supercycle. Nomura's latest research indicates that driven by multiple technology upgrade roadmaps and key component shortages, the growth momentum of the global AI network market will extend into 2026 and even 2027, creating structural opportunities in core sectors like optical modules, optical fibers and cables, and high-speed copper cables.
According to market intelligence, Nomura's Bing Duan team, in a report published on January 9, forecasts that shipments of 800G optical modules will increase from 20 million units in 2025 to 43 million units in 2026, while 1.6T optical module shipments are expected to surge from 2.5 million to 20 million units. The penetration rate of Silicon Photonics (SiPh) technology in the 800G/1.6T market is projected to reach 50-70%. Supply constraints for optical chips (laser chips/materials) are expected to persist, providing leading manufacturers with room for price increases and margin expansion.
Accelerating technology roadmaps are a key driver. Global AI giants, including NVIDIA, Google, Meta, and Amazon AWS, are actively advancing network architecture upgrades, exploring various technological paths such as Silicon Photonics (SiPh), Co-Packaged Optics (CPO), Optical Circuit Switching (OCS), Active Electrical Cables (AEC), and Hollow Core Fiber (HCF). NVIDIA plans to launch its Rubin platform in 2026, equipped with 1.6T network interfaces; Google's TPU v7 utilizes 800G optical modules and OCS technology; Meta has already deployed a customized network architecture at scale within its 18K GPU clusters.
Supply chain bottlenecks will strengthen the advantages of leading companies. According to Nomura estimates, the advanced optical chip capacity (CW+EML+VCSEL) of major global suppliers will grow by over 80% year-over-year in 2026, yet still lag behind demand by 5%-15%. Among key players, Zhongji Innolight maintains a leading global market share in the 800G/1.6T optical module market; TFC possesses core competitiveness in optical engines and FAU components; T&S Communications holds a solid position in the high-end connector market, including MPOs; and Yangtze Optical Fibre and Cable is accelerating its global penetration of AIDC-related optical fiber and cable products.
Technology upgrades are progressing on multiple fronts, charting an evolutionary path from pluggable modules towards CPO.
Nomura states that AI data centers need to scale compute and interconnect capabilities across different tiers, forming three critical layers: scale-up (intra-rack interconnect), scale-out (inter-rack/cluster interconnect), and scale-across (inter-data center interconnect).
In scale-up networks, copper cable interconnects still hold significant importance. NVIDIA's GB200 NVL72 utilizes NVLink 5.0 and PCIe 6.0 protocols, employing custom copper cable boxes to interconnect 72 compute nodes, involving up to 5,184 copper cable connections. However, as GPU cluster sizes continue to expand, the limited effective transmission distance of copper cables at high speeds (less than 10 meters) becomes a bottleneck, increasingly highlighting the importance of optical communication solutions.
In scale-out networks, pluggable optical modules remain the mainstream solution, but technological iteration is accelerating. Product innovations include Silicon Photonics (SiPh), LPO/LRO, and NPO/CPO, primarily focusing on higher performance, lower power consumption, and cost reduction. Nomura analysis suggests that the 1.6T upgrade and SiPh migration are key drivers for 2026, with SiPh optical modules expected to capture 50-60% of the 800G market and 60-70% of the 1.6T market share that year.
CPO technology is expected to accelerate its commercial adoption starting in 2026. By integrating the optical engine directly with the switch ASIC, CPO can significantly enhance bandwidth density and port density while reducing power consumption per bit. Broadcom has already launched a 102.4T CPO switch chip based on its Tomahawk 6 platform, and NVIDIA introduced its Quantum-X CPO switch based on the InfiniBand network at the 2025 GTC conference. Nomura anticipates that CPO adoption in scale-out networks will accelerate in the medium term, benefiting from potential bundling strategies by global AI giants.
An analysis of the network architecture roadmaps of major cloud service providers.
Nomura indicates that NVIDIA continues to lead the evolution of AI network technology. Its Rubin series platform will adopt NVLink 6.0 technology (3600GB/s), equipped with CX9 NICs (1.6Tbps) and Spectrum 6 CPO switches (102T). In scale-up networks, the Rubin Ultra NVL576 will use an orthogonal backplane (PCB) to replace copper backplanes; in scale-out networks, 1.6T InfiniBand/Ethernet switches will be deployed. The company may also disclose more details about Spectrum-X Ethernet version CPO products at its GTC event in March 2026.
Google's TPU v7 employs a 3D torus topology, using copper cables for intra-rack TPU interconnects, optical modules for inter-rack connections, and OCS technology for cluster-level connection scheduling. Meta utilizes a hybrid interconnect architecture: strengthening copper on the scale-up side, deploying high-speed optical modules on the scale-out side, and testing CPO switches to enhance reliability.
AWS's Trainium3 + NeuronLink employs a hybrid interconnect of PCB + backplane + AEC, and uses its proprietary EFA protocol to replace RoCE v2/InfiniBand, aiming to improve cluster scale and congestion control.
The optical communication market is experiencing structural growth opportunities. According to data from market research firm LightCounting, the Ethernet optical module market revenue surged 93% year-over-year in 2024, with projected growth of 48% and 35% for 2025 and 2026, respectively. Nomura emphasizes that this growth is constrained by EML laser chip capacity.
Supply chain bottlenecks are reshaping the competitive landscape. The primary cost components of optical modules include TOSA (optical transmitter assembly, containing the laser), ROSA (optical receiver assembly, containing the photodetector), and electronic chips (e.g., DSP), collectively accounting for 56-61% of the total cost of 800G/1.6T modules. The high-end chip market is still dominated by US and Japanese companies like Broadcom, Coherent/Finisar, Cisco/Acacia, Mitsubishi Electric, and Sumitomo Electric.
Chinese companies continue to make breakthroughs in the midstream and downstream segments. As the world's largest data center optical module supplier, Nomura expects Zhongji Innolight to maintain 25-30%/35-40% market share in the 800G/1.6T markets, respectively. TFC, as a supplier of high-end optical engines, will benefit from the 1.6T upgrade and is expected to secure orders for its FAU (Fiber Array Unit) products in the CPO market.
The optical fiber and cable market shows a diverging trend. Demand in the domestic traditional telecom market faces pressure, but demand from AI data centers is robust. According to CRU data, demand for optical cables in AI applications grew 138% year-over-year in 2024 and is projected to grow another 77% in 2025. Yangtze Optical Fibre and Cable's AI data center business is accelerating its global market expansion.
The CPO market is on the cusp of commercialization. Nomura expects the penetration rate of CPO switches to increase from 3% in 2026 to 20% by 2030, with the market size growing from $1.6 billion in 2026 to $13.1 billion by 2030.
The technology roadmap is becoming clearer. NVIDIA's Quantum X800 CPO switch incorporates four 28.8T switch ASIC chips, each surrounded by 6 sub-components, with each sub-component containing three 1.6T CPO optical engines. The entire switch requires 144 FAUs, 72 1.6T optical engines, and 18 External Laser Sources (ELS), with Nomura estimating a total cost of approximately $80,000-$90,000.
Based on the partner list announced at NVIDIA GTC 2025, the CPO switch supply chain includes: optical engine and FAU suppliers; EIC+PIC packaging; optical engine packaging; CW laser chip suppliers; MPO and fiber suppliers; and switch assembly. Nomura states that TFC, leveraging its position as an FAU supplier in the pluggable optical module market, is poised to secure CPO orders based on customer relationships and technological积累.
Market expectations that optical communications will entirely replace copper cables are misguided. Nomura emphasizes that copper cables will continue to play a vital role in scale-up and some scale-out networks, offering advantages in speed and efficiency crucial for handling massive AI tasks. Despite transmission distance limitations, copper cables hold a cost advantage over optical modules, and AEC can extend the transmission distance to 10 meters.
Market size expectations are optimistic. According to LightCounting, high-speed copper cable sales are expected to double over the next five years, reaching $2.8 billion by 2028. NVIDIA pioneered the use of DAC in its Blackwell NVL72 system and is likely to continue using copper cables in its 2026 Rubin platform. Major AI customers like Amazon AWS, Meta, and Microsoft also use copper cables in their AIDC projects.
The supply chain landscape shows consolidation trends. Key players in the copper cable supply chain include Amphenol, Credo, as well as Cheng Uei Precision and WOLONG Electric Group, which have entered the supply chains of global AI customers.
The Ethernet switch market maintains strong growth. According to IDC data, global Ethernet switch market revenue reached $14.7 billion in Q3 2025, up 35.2% year-over-year, with the data center segment growing 62.0%. 800G switch revenue grew 91.6% quarter-over-quarter, accounting for 18.3% of data center market revenue.
Nomura indicates that the ODM direct sales model continues to expand. In Q3 2025, ODM direct sales revenue grew 152.4% year-over-year, accounting for 30.2% of data center market revenue. This reflects the growing preference of global cloud service providers for "white box" switch products to reduce costs and increase flexibility, a trend that may negatively impact the margins of branded switch vendors with proprietary IP and software.
CPO and OCS technologies are developing in parallel. Broadcom continues to promote CPO adoption based on its Tomahawk 6 platform, and NVIDIA's Quantum-X and Spectrum-X switches may also launch CPO versions. Meanwhile, Google's OCS technology is gaining attention for its strong performance in large-scale TPU training clusters. The Open Compute Project Foundation has established an OCS sub-project.
Comments