Scale-up technology is driving the development of China's AI supernodes, benefiting domestic interconnect and AI accelerator supply chains

Overview

This report focuses on the development of China's AI server Scale-up networks and AI supernodes. Morgan Stanley believes that, constrained by chip-level factors such as wafer process technology, competition in China's AI computing market is shifting from the performance of individual accelerators to system-level solutions. At WAIC 2026, domestic vendors reduced pure chip launches and instead showcased more rack-level or multi-rack supernode solutions, typically connecting 64 or more accelerators through self-developed Scale-up interconnect technologies.

Core views

The core view is that Scale-up is becoming the main direction of differentiation in China's AI computing architecture, and the domestic ecosystem is catching up quickly. Short-distance interconnects within racks are still mainly electrical because costs are lower and integration is simpler; when Scale-up domains extend across racks, constraints from copper cable distance, signal loss, and power consumption intensify, increasing the importance of optical interconnects. As a result, domestic interconnect and optical engine solutions such as PCIe retimers, PCIe switches, CXL/PCIe AECs, NPO/CPO switching modules, silicon photonics PICs, and TIAs may see incremental demand.

Analysis framework

The report analyzes the topic through supply-chain technology decomposition, observations from WAIC 2026, comparisons of domestic and overseas Scale-up network architectures, and valuation and risk frameworks for covered companies. Technically, it distinguishes Scale-up from Scale-out: Scale-up is used for high-speed communication among accelerators within the same server or rack, enabling them to share memory and process the same workload like a large supercomputer; Scale-out connects multiple systems or supernodes across broader data center infrastructure.

Methodology notes

• Technical architecture comparisonScale-up vs Scale-out

  Scale-up connects GPUs or accelerators within the same server, rack, or cross-rack supernode; Scale-out connects multiple supernodes or systems to form large-scale clusters.

  This framework is used to determine the application positioning and investment implications of different interconnect technologies: Scale-up places greater emphasis on low latency, high bandwidth, and tight coupling, while Scale-out relies more on data center networks such as Ethernet, RoCE, UBoE, or InfiniBand.

• Supply chain mappingAI supernode interconnect supply chain

  Decomposes the beneficiary segments from CPUs, xPUs, switching chips, NICs, PCIe/CXL links, and optical engines to full-rack architectures.

  The report maps China's AI supernode expansion to supply chains for PCIe retimers, PCIe switches, optical interconnects, NPO/CPO, silicon photonics, and domestic AI GPUs.

• Valuation methodResidual income model

  Uses a residual income model for covered companies and discloses key assumptions such as cost of equity, payout ratio, medium-term growth rate, and terminal growth rate.

  For example, the base-case scenario for Montage H-shares uses an 8.4% cost of equity, a 30% payout ratio, a 19.3% medium-term growth rate, and a 4% terminal growth rate, and assumes an exchange rate of 1.15 HKD to 1 RMB.

Asset mapping & comparison

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

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

  A beneficiary in PCIe/CXL interconnects, referenced with an Overweight rating.

  Strengths

  Well positioned in China's data center semiconductor localization trend; PCIe 6.x/CXL3.x x16 AEC has completed interoperability testing, and larger supernodes may increase content for PCIe retimers and switches.

  Weaknesses

  Affected by cloud demand, DRAM interface technology migration, and the launch pace of new products.

  Comparison

  Compared with direct AI GPU vendors, Montage is more of a pick-and-shovel play on increased interconnect content.

  Risks

  Cloud demand weaker than expected, DRAM interface migration slower than expected, delays in new product launches.

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

  A beneficiary in domestic CPU/GPU and AI server ecosystems, referenced with an Overweight rating.

  Strengths

  May enhance AI server rack performance through Scale-up technology together with the Sugon ecosystem and further differentiate itself in China's CPU and GPU markets.

  Weaknesses

  Leading-edge process capacity, yields, and local GPU price competition may constrain growth.

  Comparison

  Compared with interconnect chip companies, Hygon has more direct exposure to domestic AI computing demand and CSP orders.

  Risks

  More intense price competition among local GPU vendors, weaker-than-expected AI demand in China, slower yield improvement and capacity build-out at local leading-node foundries.

• Cambricon Technology Corporation (688256.SS)

  A beneficiary in domestic AI accelerators, referenced with an Overweight rating.

  Strengths

  Expected to benefit from stronger AI demand, CSP order ramp-up, and accelerated localization.

  Weaknesses

  Capacity, yields, and customer concentration are the main constraints.

  Comparison

  Similar to Hygon, Cambricon benefits more directly from demand for domestic AI accelerators, but is also more exposed to capacity and order execution risks.

  Risks

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

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

  A beneficiary in domestic AI GPUs, referenced with an Overweight rating.

  Strengths

  If CSP orders are stronger than expected, software replaces CUDA faster, and domestic and overseas capacity expansion proceeds smoothly, growth potential would increase.

  Weaknesses

  Order ramp-up, software ecosystem, and execution of capacity expansion still need to be validated.

  Comparison

  Like Cambricon, it is in the domestic AI accelerator segment, but the report specifically notes that progress in software replacement of CUDA is one of its upside drivers.

  Risks

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

• Lightelligence

  Not a covered company; a participant in the NPO/CPO and silicon photonics interconnect ecosystem.

  Strengths

  Showcased a 51.2T CPO switching module, NPO switching module, and LightSphere X, with technologies that can serve both Scale-up and Scale-out networks.

  Weaknesses

  The report does not provide coverage ratings or financial valuation.

  Comparison

  Compared with AI accelerator companies, Lightelligence is more focused on optical engines, silicon photonics, and optical switching infrastructure.

  Risks

  Uncertainty remains around the pace of commercial deployment, system compatibility, customer adoption, and the competitive landscape.

Key data

• Huawei Atlas 950 Scale-up scaleDemonstrated configuration connecting 1,024 NPUs, with design scalable to 8,192 NPUsUses UnifiedBus 2.0 and can scale across racks within the same SuperPoD through optical interconnect.
• Biren BR2xx target scaleSupports up to 1,024 GPUsTargets larger Scale-up domains through BLink 2.0 and an NPO architecture.
• Sugon scaleX640Connects 640 accelerators in a single rackOne of China's large-scale Scale-up supernode solutions.
• Moore Threads MTT C256128 GPUs in a single rack, 256 GPUs in two racksUses MT-Link 2.0 and adopts an all-copper-cable tray design.
• MetaX, Alibaba, Enflame Scale-up domains64 to 128 acceleratorsSupported respectively through MetaXLink-E, ICN, and GCU-LARE.
• Montage PCIe product progressPCIe 6.x/CXL3.x x16 AEC has completed interoperability testingTarget applications include supernodes and cross-rack deployments.
• Lightelligence showcase51.2T CPO switching module and NPO switching moduleImproves bandwidth density and energy efficiency by shortening the high-speed electrical path between the switch ASIC and the optical engine.
• Enflame and Lightelligence solutionNPO-equipped GPU-server solution for 512 or more acceleratorsShows that domestic optical technologies are entering the accelerator optical I/O layer and the distributed optical switching layer.

Impact & implications

From an investment perspective, the report believes that AI supernode expansion will increase domestic interconnect content. Montage may benefit from denser and longer-distance PCIe link demand among CPUs, xPUs, switches, NICs, and peripherals in larger and more distributed supernodes; domestic AI GPU vendors such as Hygon and its ecosystem partners Sugon, Cambricon, and Iluvatar may improve AI server rack performance by introducing or leveraging Scale-up technology. Over the longer term, cross-rack Scale-up will shift incremental value from pure accelerators to switching chips, optical engines, silicon photonics, CPO/NPO modules, and system architecture capabilities.

Risks

• China AI demand weaker than expected.
• Cloud capex or CSP order ramp-up below expectations.
• Intensifying price competition among local GPU companies.
• Leading-node capacity build-out and yield improvement slower than expected.
• DRAM interface technology migration slower than expected.
• Delays in new product launches or technology iteration slower than expected.
• Customer concentration risk.
• US executive orders, export controls, or escalating sanctions may affect related securities trading, supply chains, and commercialization pace.

What to watch

• Whether actual domestic AI supernode deployment scale expands from 64 to 128 accelerator domains to 512, 640, 1,024, or even higher.
• Whether Huawei Atlas 950, Biren BR2xx, Sugon scaleX640, Enflame NPO solutions, and others enter large-scale deployment.
• Whether cross-rack Scale-up drives demand for optical interconnects, NPO/CPO, silicon photonics PICs, and optical engines.
• Order and customer adoption progress for Montage's PCIe 6.x/CXL3.x AEC, PCIe retimer, and PCIe switch chips.
• Whether CSP capex becomes a key driver of demand for domestic AI GPUs.
• The impact of US export controls and executive orders on covered companies, customer procurement, and investor trading restrictions.