Scale-up technology is pushing China's AI supernodes into a stage of system-level competition

Overview

Based on observations at WAIC, this report focuses on the development of scale-up networking in China's AI servers, the role of optical interconnect in multi-rack supernodes, and the investment implications for the domestic interconnect supply chain and AI GPU vendors. The core view is that because China's AI chips remain constrained at the process-node level, competition is shifting from single-chip specifications to system-level solutions, with supernodes and interconnect architectures becoming key differentiators.

Core views

China's scale-up ecosystem is expanding rapidly and is currently led by each vendor's proprietary interconnect technologies across multiple system architectures. Short-distance connections within a rack still favor electrical interconnect because of lower cost and simpler integration; when the scale-up domain extends across racks, copper cable distance, signal loss, and power consumption become constraints, increasing the importance of optical interconnect. The report believes that domestic optical engine, switch chip, PCIe retimer, PCIe switch, and AI GPU vendors may all benefit from larger-scale supernode deployments.

Analysis framework

The report uses an industry chain and system architecture analytical framework, breaking down China's AI servers into several layers including chips, racks, cross-rack interconnects, electro-optical conversion, and cluster networks, and comparing them with U.S. AI networking technology roadmaps. The stock impact section assesses beneficiaries based on increased interconnect content, domestic substitution, CSP capital expenditure, technology migration, and changes in AI demand.

Methodology notes

• Technical architectureScale-up networking

  High-speed communication among GPUs/accelerators within the same server or rack, enabling multiple accelerators to share memory and process the same workload like a large supercomputer.

  The report distinguishes scale-up from scale-out: scale-up connects accelerators inside a supernode and is a key point of system differentiation; scale-out connects multiple systems or supernodes and is the backbone network for large-scale clusters.

• Industry chain mappingInterconnect content analysis

  As AI systems expand from a single rack to multiple racks, denser and longer-distance PCIe/CXL and optical interconnects are required among CPUs, xPUs, switch chips, NICs, and peripherals.

  This framework is used to explain the potential incremental demand for PCIe interconnect suppliers such as Montage, as well as domestic vendors of optical interconnect, NPO/CPO, and silicon photonics solutions.

• Valuation methodResidual income model

  The report uses a residual income model for some covered companies and provides key assumptions such as cost of equity, dividend payout ratio, medium-term growth rate, and terminal growth rate.

  Examples include Montage's 8.4% cost of equity, 30% payout ratio, 19.3% medium-term growth rate, and 4% terminal growth rate; other AI chip companies also disclose assumptions such as cost of equity, payout ratio, and growth rates.

Asset mapping & comparison

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

• Montage Technology Co Ltd (6809.HK / 688008.SS)

  Beneficiary of increased PCIe/CXL interconnect content

  Strengths

  PCIe 6.x/CXL3.x x16 AEC has completed interoperability testing, suitable for supernodes and cross-rack deployment; larger distributed supernodes will increase demand for PCIe retimers and PCIe switches.

  Weaknesses

  Demand still depends on cloud capital expenditure, DRAM interface technology migration, and the pace of new product launches.

  Comparison

  Compared with pure AI GPU vendors, Montage is more of an interconnect infrastructure beneficiary, with upside driven by higher PCIe link density and distance per unit of computing power.

  Risks

  Weaker-than-expected cloud demand, slower DRAM interface migration, delays in new product launches.

• Hygon Information Technology Co., Ltd. (688041.SS)

  Beneficiary of domestic CPU/GPU and supernode ecosystem

  Strengths

  The report believes Hygon and ecosystem partner Sugon can improve AI server rack performance through scale-up technology and gain share in China's CPU and GPU markets.

  Weaknesses

  Local GPU competition and advanced process capacity may still limit the pace of expansion.

  Comparison

  Compared with interconnect chip suppliers, Hygon is more directly exposed to China's AI demand, the GPU competitive landscape, and local advanced process supply.

  Risks

  Intensified local GPU price competition, weaker-than-expected AI demand in China, and slower-than-expected yield and capacity ramp-up in local advanced process nodes.

• Cambricon Technology Corporation (688256.SS)

  Beneficiary of domestic AI GPU and AI accelerators

  Strengths

  The report lists Cambricon as one of the domestic AI GPU vendors that may introduce or utilize scale-up technology to improve AI server rack performance.

  Weaknesses

  Business elasticity depends heavily on CSP orders, capacity and yield constraints, and customer concentration.

  Comparison

  Compared with interconnect suppliers such as Montage, Cambricon is closer to underlying AI accelerator demand and is more affected by AI demand and accelerated domestic substitution.

  Risks

  Capacity and yield constraints, customer concentration risk, slower-than-expected technology iteration.

• Iluvatar CoreX Semiconductor Co., Ltd. (9903.HK)

  Beneficiary of domestic AI GPU and CUDA substitution

  Strengths

  The report believes Iluvatar can improve AI server rack performance through scale-up technology; potential catalysts include stronger CSP orders and faster software ecosystem substitution.

  Weaknesses

  It still faces order ramp-up, sanctions, and competitive pressure.

  Comparison

  Compared with other AI GPU vendors, the report specifically notes that faster CUDA substitution and overseas and domestic capacity expansion could form a bull-case scenario.

  Risks

  Order ramp-up below expectations, escalating sanctions, intensified competition.

• Domestic optical interconnect and NPO/CPO ecosystem

  Infrastructure beneficiary of multi-rack scale-up expansion

  Strengths

  Cross-rack scale-up increases the attractiveness of optical interconnect. Lightelligence showcased solutions such as 51.2T CPO switch modules, NPO switch modules, and LightSphere X, indicating that domestic optical technology is entering AI system deployment.

  Weaknesses

  Some related companies are not covered, and commercialization pace, maintainability, power consumption, and cost still need validation.

  Comparison

  Electrical interconnect still has cost and integration advantages within a single rack; optical interconnect is more suitable for cross-rack, larger-scale, and higher-bandwidth-density scenarios.

  Risks

  Slower-than-expected deployment progress, insufficient reliability of optoelectronic integration, and uncertain adoption pace by system vendors.

Key data

• Report date2026-07-27The report cover time is July 27, 2026 09:00 PM GMT.
• Industry viewAttractiveThe covered industry is Greater China Technology Semiconductors.
• Huawei Atlas 950Demonstrated configuration of 1,024 NPUs, designed to scale to 8,192 NPUsUses UnifiedBus 2.0 and is one of the leading domestic large-scale scale-up solutions.
• Biren BR2xxMaximum target of 1,024 GPUsAdvancing a large-scale scale-up domain through BLink 2.0 and NPO architecture.
• Sugon scaleX640640 accelerators in a single rackReflects the expansion of domestic supernode solutions toward high-density rack-level deployment.
• Moore Threads MTT C256128 GPUs in a single rack, 256 GPUs across two racksAchieves scale-up through MT-Link 2.0 and an all-copper-cable tray design.
• Other domestic scale-up domains64 to 128 acceleratorsMetaX, Alibaba, and Enflame support this scale range through MetaXLink-E, ICN, and GCU-LARE, respectively.
• Montage interconnect productsPCIe 6.x/CXL3.x x16 AEC has completed interoperability testingTarget applications include supernodes and cross-rack deployments.
• Enflame and Lightelligence solutionNPO GPU servers target 512 or more acceleratorsShows that domestic optical interconnect technology is entering the accelerator optical I/O and distributed optical switching layers.

Impact & implications

The investment implications are concentrated along two main themes: first, larger supernode scale increases content demand for PCIe/CXL, retimers, switches, and optical interconnects; second, domestic AI GPU vendors can use system-level scale-up capabilities to offset single-chip process constraints, improving AI server rack performance and competitiveness in domestic substitution. If China's CSP capital expenditure continues to grow, demand elasticity for related interconnect and accelerator supply chains may strengthen.

Risks

• China AI demand or CSP capital expenditure may be weaker than expected.
• Intensifying price competition among local AI GPU vendors may affect profitability.
• Progress in advanced process capacity, yield improvement, and supply chain localization may be slower than expected.
• Technology migration in DRAM interfaces, PCIe/CXL, or optical interconnect may be slower than expected.
• New product launches or supernode deployments may be delayed.
• High customer concentration may result in order ramp-up below expectations.
• Export controls, U.S. executive orders, or escalating sanctions may affect certain entities and securities trading.

What to watch

• The actual mass production and delivery pace of domestic AI supernodes after WAIC.
• Whether the scale-up domain size of vendors such as Huawei, Biren, Sugon, Enflame, MetaX, and Alibaba continues to expand.
• The adoption progress of optical interconnect, NPO/CPO, silicon photonics PICs, and optical engines in cross-rack scale-up.
• Customer adoption and content increases for Montage's PCIe 6.x/CXL3.x related products.
• Changes in China CSP capital expenditure and domestic AI GPU orders.
• Local advanced process capacity, yields, and supply stability.
• Changes in regulation, export controls, and restricted securities trading rules.