報告_MS_AI網通_20260727

PDF 原檔：報告_MS_AI網通_20260727_original.pdf

圖片清單（已驗證 2026-07-28）

Step 2.5 眼見為憑索引，是 lib/ 嵌圖的唯一真相來源。只挑分類為「真資料圖」且已逐張 Read 者嵌入，圖說照抄「親眼所見內容」。

逐張 Read 驗證 7 張（_006／_008／_009／_010／_011／_012／_014），全為真資料圖。★注意：抽圖編號與 trimmed 文中 Exhibit 順序不一致——_010 是壁仞（文中 Exhibit 11）、_011 是燧原（文中 Exhibit 10），嵌入時以本清單為準。

原始內容

M July 27, 2026 09:00 PM GMT

Greater China Semiconductors | Asia Pacific

AI Networking: Scale-up Technology Enables China's AI Supernodes

Following our WAIC takeaways, we further examine China's emerging scale-up landscape, the role of optical interconnect, and implications for the domestic interconnect supply chain.

AI Supernodes enabled by scale-up are becoming the focus of China's AI computing landscape. Alongside our China WAIC takeaways, we provide an indepth follow-up study of China's scale-up technology development. As we wrote in China's AI Accelerators - Who's Poised to Win?, Chinese AI GPUs are strong in optical networking and server rack design at the system level, although they remain constrained by wafer process technology at the chip level ( Exhibit 6 ). These chiplevel constraints are shifting competition in China's AI computing market from standalone chip specs to system solutions. WAIC 2026 featured fewer chip launches and more supernode solutions. Leading domestic accelerator vendors displayed rack-scale or multi-rack supernodes with scale-up domains of 64 accelerators or more, mostly enabled by proprietary scale-up technologies.

China's scale-up landscape is rapidly broadening, led by proprietary fabrics but spanning multiple system architectures ( Exhibit 4 ) . A year ago, our US semis team published Semiconductors: Scaling the AI Opportunity: A Scale-Up Network Primer (29 Aug 2025); China is now catching up quickly in scale-up technology. Huawei remains the domestic scale leader, with Atlas 950 using UnifiedBus 2.0 to connect 1,024 NPUs in the demonstrated configuration and designed to scale to 8,192 NPUs. Biren's next-gen BR2xx NPO architecture targets up to 1,024 GPUs through BLink 2.0. Sugon's scaleX640 connects 640 accelerators in a single rack. Moore Threads'(not covered) MTT C256 supports 128 GPUs in one rack and 256 across two racks through MT-Link 2.0, while MetaX, Alibaba and Enflame support 64- to 128-accelerator domains through MetaXLink-E, ICN, and GCU-LARE, respectively.

Stock implications: We believe Montage (OW, covered by Daniel Yen) is set to benefit from higher PCIe interconnect content in larger, distributed supernodes. As systems scale across racks, denser and longer PCIe links among CPUs, xPUs, switches, NICs and peripherals should support demand for PCIe retimers and PCIe switch chips. We also expect China AI GPU vendors, such as Hygon (OW, covered by Daisy Dai) (along with its ecosystem partner Sugon), Cambricon and Iluvatar (both OW, covered by Charlie Chan), to introduce or leverage scale-up technology to improve their AI server rack performance.

(Continued below.)

Idea

Morgan Stanley does and seeks to do business with companies covered in Morgan Stanley Research. As a result, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of Morgan Stanley Research. Investors should consider Morgan Stanley Research as only a single factor in making their investment decision.

For analyst certification and other important disclosures, refer to the Disclosure Section, located at the end of this report.

+= Analysts employed by non-U.S. affiliates are not registered with FINRA, may not be associated persons of the member and may not be subject to FINRA restrictions on communications with a subject company, public appearances and trading securities held by a research analyst account.

Camlull o.

Technology

Purpose

Bandwidth

Latency

Ecosystem

Scalability

Uu Minlelwulniy ouale up leulmluluyleo

NVLink5

UALink 1.0

GPU-GPU; CPU-GPU

xPU-xPU

SUE

xPU-XPU

PCle Gen 6

Primarily CPU-XPU;

xPU-xPU

ICI

TPU interconnect

NeruonLink

Trainium interconnect

M

China AI Interconnect - Scale-up

Слила. Латкаки соа /лемак Сакіли Валалелкі

not standardized varies

Global AI server scale-up technology trend

What is scale-up networking?

Scale-up networking refers to high-speed communication between GPUs/accelerators within the same server or rack. Interconnecting accelerators allow them to function as one large supercomputer, accessing the same memory and processing the same workload. This differs from scale-out networking, which connects multiple systems across a broader data center infrastructure.

Montage's PCIe 6.x/CXL3.x x16 AEC has completed interoperability testing, and targets supernode and inter-rack deployments. Chinese systems deploying more xPUs per unit of compute could further increase PCIe content. Beyond standardized PCIe connectivity, multi-rack scale-up should also accelerate demand for domestic optical interconnect.

Exhibit 1: Scale-up interface data rate per lane

Exhibit 2: Scale-up bandwidth per accelerator

Source: Company data, Morgan Stanley Research

Source: Company data, Morgan Stanley Research

Exhibit 3: US AI networking scale-up technologies

Source: Company data, Morgan Stanley Research

Juale uu vo. suale vul fletwulnily

M

Exhibit 4: AI supernode scale-up solutions

Source: Company data, Morgan Stanley Research

Exhibit 5: Scale-up vs. scale-out networking

Source: Broadcom 2025 OCP Summit Keynote, Morgan Stanley Research

China AI server scale-up development

In China, electrical interconnect remains the preferred option within racks, while optical becomes more important as scale-up domains extend across racks . Short electrical links offer lower cost and simpler integration within boards, trays and singlerack systems. Moore Threads' C256 uses an all-copper cable tray design, while the orthogonal zero-cable architectures by Enflame and MetaX also remain electrical. Optical interconnection becomes more attractive across racks, where copper reach, signal loss and power consumption become constraints. Huawei's Atlas 950 SuperPoD extends UnifiedBus across racks within the same SuperPoD using optical interconnect.

The transition toward optical connectivity may create demand for domestic switching and optical engine solutions designed for AI systems. At WAIC 2026, Lightelligence (not covered) showcased its new NPO/CPO switch solutions ( Exhibit 8 , Exhibit 9 ), which pair the switch ASIC from a leading domestic vendor, which we believe is Centec Communications (not covered), with Lightelligence's optical engine, including the SiPh PIC and selected EICs such as the TIA. The NPO/CPO solutions shorten the high-speed

M

electrical path between the switch ASIC and optical engine, improving bandwidth density and power efficiency, while NPO retains greater serviceability than CPO. The same architecture can serve both scale-up and scale-out networks, with scale-up becoming increasingly relevant as tightly coupled accelerator domains extend across racks.

More broadly, scale-up is emerging as the principal area of architectural differentiation, while scale-out remains the backbone of large-scale clusters . System solutions generally use a specialized fabric to connect accelerators within a supernode, and Ethernet-based fabrics such as RoCE or UBoE, or Infiniband to connect multiple supernodes across a cluster. NVIDIA combines NVLink with InfiniBand or Spectrum-X Ethernet; AMD Helios pairs UALink with Ethernet scale-out; Huawei uses UnifiedBus within the SuperPoD, and UBoE or RoCE for scale-out between SuperPoDs.

Exhibit 6: Relative strengths of AI industries in the US and China - optical networking almost on par

Source: Morgan Stanley Research

Exhibit 7: Domestic chips have lower TCO and comparable per token cost (AI LLM inference) vs. NVIDIA's processors for China

TCO of 10MW capacity (US$ mn)

Per token cost (US$ cent) (Right-axis) (Lower is better)

Source: Company data, Morgan Stanley Research estimates

Lightelligence)

M

Optical interconnect is progressing towards scale-up system deployment

Enflame and Lightelligence previously demonstrated an xPU-CPO prototype that places the optical engine beside the accelerator, enabling direct optical output from the xPU and reducing the electrical distance between the compute silicon and optics. At WAIC 2026, Enflame displayed an NPO-equipped GPU-server solution targeting scale-up domains of 512 accelerators or more, developed in partnership with Lightelligence ( Exhibit 10 ). Biren also unveiled a next-gen BR2xx-based NPO architecture designed to support a scale-up domain of up to 1,024 GPUs ( Exhibit 11 ). Also, LightSphere X (jointly developed by Lightelligence, Biren, ZTE, and other partners) incorporates Lightelligence's SiPh interconnect and dOCS technology, enabling the accelerator domain to extend across racks and allowing the topology to be reconfigured according to workload requirements ( Exhibit 12 ). These projects show domestic optical technology being integrated at both the accelerator optical-I/O layer and the distributed optical-switching layer, to support larger, multi-rack scale-up domains.

Source: Company data, Morgan Stanley Research

Exhibit 8: 51.2T CPO switch module (displayed by Lightelligence)

Exhibit 9: NPO switch module (displayed by Lightelligence)

Source: Company data, Morgan Stanley Research

Source: Company data, Morgan Stanley Research

CAInUI 7.

Wru owultuuule uloplayeu by Llytemyellue)

cxmnbll 1e. Liymopnleten (Llymemyenue)

((Enflame

M

B.HiCable Tray

15583%P9

RAX.

GPU 38-16

Exhibit 10: Enflame's NPO optical interconnect solution for scale-up

GPUB A-NI

GPUB 4-3

GPUB 4-3

Exhibit 11: Biren's NPO optical interconnect solution for scale-

up

1B15 MSWERTENEPS

10 1E64/32H128#

Source: Company data

Source: Company data

CPUB A-1

Source: Company data

Exhibit 12: LightSphere X (Lightelligence)

Source: Company data, Morgan Stanley Research

NPOW3IS

" T. T

CAnnot To.

ulmlese voro vaves will be a ney uetallu univel lot utmlese hi orus lel glolmea Met de aibne -mllae enl

XPU Partner/Supplier

Scale-Up Supplier

• Scale-Out Scale-Up

60,000

M

XPU

Next Gen Scale- up Network?

NVLINK Fusion nVIDIA

14, Mellampx

$17bn

UALink

US scale-up networking ecosystem and TAM forecast

20,000

| MARVELL

alchip

10,000

I BROADCOM

I MARVELL

Trainium2

$4bn

S6bn

2024

MTIA v2

Maia ustom NeruonLink

PCle

2025

2026

amazon

Exhibit 13: Chinese CSPs' capex will be a key demand driver for Chinese AI GPUs

L BROADCOM

@ BROADCOM

/Multi-vendor

2027

Source: Company data, Morgan Stanley Research

Exhibit 14: US semi : Scale-up TAM estimate ($mn)

Source: Company data, Morgan Stanley Research (E) estimates

M

This report references U.S. Executive Order 14032 and/or entities or securities that are designated thereunder. U.S. persons may be prohibited from buying certain securities of entities named in this report. Readers are solely responsible for ensuring that their investment activities are carried out in compliance with applicable laws.

This report references export controls and/or entities that may be subject to export control restrictions. Readers are solely responsible for ensuring that their investment or trade activities are carried out in compliance with applicable laws.

This report references U.S. Executive Order 14105 and/or entities that may be in scope of such order. U.S. persons may be prohibited from engaging in certain transactions or otherwise require certain other transactions be notified to the U.S. Department of Treasury. Readers are solely responsible for ensuring that their investment or trade activities are carried out in compliance with applicable laws.

M

Valuation Methodology and Risks

Montage Technology Co Ltd (6809.HK)

Base case, residual income model. Key assumptions:

• n 8.4% CoE (1.2 beta, 3.0% risk-free rate, and 4.5% risk premium)
• n 30% payout ratio
• n 19.3% medium-term growth rate
• n 4% terminal growth rate

These assumptions reflect cloud capex growth, DRAM interface technology migration, and Montage's strong position in China's datacenter semi localization. We then apply an exchange rate of 1.15 HKD:1 RMB, assuming no H-share discount vs the A-share.

Risks to Upside

• n Faster-than-expected phase-out of US peers
• n Faster-than-expected spec migration

Risks to Downside

• n Weaker-than-expected cloud demand
• n Slower-than-expected DRAM interface technology migration
• n Delay in new product launches

Montage Technology Co Ltd (688008.SS)

Base case, residual income model. Key assumptions:

• n 8.4% CoE (1.2 beta, 3.0% risk-free rate, and 4.5% risk premium)
• n 30% payout ratio
• n 19.3% medium-term growth rate
• n 4% terminal growth rate

We believe these assumptions are justified, given the cloud capex growth, DRAM interface technology migration, and Montage's strong position in China's datacenter semi localization.

Risks to Upside

• n Faster-than-expected phase-out of US peers
• n Faster-than-expected spec migration

Risks to Downside

• n Weaker-than-expected cloud demand
• n Slower-than-expected DRAM interface technology migration
• n Delay in new product launches

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

We assume an 8.0% cost of equity (beta 1.05, risk-free rate 2.0% and risk premium 5.8%), a payout ratio of 50%, a medium-term growth rate of 25.0%, and a terminal growth rate of 5.0%, all of which are in line with other China AI chip companies under our coverage.

M

Risks to Upside

• n Hygon further differentiates itself from local CPU and GPU competitors and gains significant market share in China's CPU and GPU markets
• n China AI demand is stronger than expected
• n Faster ramp-up of leading node capacity

Risks to Downside

• n Intensified pricing competition among local GPU companies
• n China AI demand is weaker than expected
• n Slower yield improvement and capacity buildout at local leading node foundries

Cambricon Technology Corporation (688256.SS)

Key valuation assumptions underpinning our model include: an 8.4% cost of equity (derived from a beta of 1.06, risk-free rate of 2.0%, and equity risk premium of 6.0%), a long-term payout ratio of 40% (was 57%), a medium-term growth rate of 16%, and a terminal growth rate of 6.0%.

Risks to Upside

• n Stronger-than-expected AI demand
• n CSP order ramp-up
• n Accelerating localization

Risks to Downside

• n Capacity and yield constraints
• n Customer concentration risk
• n Slower technology iteration

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

Key valuation assumptions include:

• n An 8.3% cost of equity (derived from a beta of 1.05, risk-free rate of 2.0%, and equity risk premium of 6.0%)
• n A long-term payout ratio of 34% (was 33%)
• n A medium-term growth rate of 16%
• n A perpetual terminal growth rate of 6%

Risks to Upside

• n Stronger-than-expected CSP orders
• n Faster CUDA replacement with Iluvatar's software
• n Expansion of overseas and domestic capacity

Risks to Downside

• n Order ramp below expectations
• n Escalation of sanctions
• n Intensifying competition

Idea

M

Disclosure Section

The information and opinions in Morgan Stanley Research were prepared or are disseminated by Morgan Stanley Asia Limited (which accepts the responsibility for its contents) and/or Morgan Stanley Asia (Singapore) Pte. (Registration number 199206298Z) and/or Morgan Stanley Asia (Singapore) Securities Pte Ltd (Registration number 200008434H), regulated by the Monetary Authority of Singapore (which accepts legal responsibility for its contents and should be contacted with respect to any matters arising from, or in connection with, Morgan Stanley Research), and/or Morgan Stanley Taiwan Limited and/or Morgan Stanley & Co International plc, Seoul Branch, and/or Morgan Stanley Australia Limited (A.B.N. 67 003 734 576, holder of Australian financial services license No. 233742, which accepts responsibility for its contents), and/or Morgan Stanley Wealth Management Australia Pty Ltd (A.B.N. 19 009 145 555, holder of Australian financial services license No. 240813, which accepts responsibility for its contents), and/or Morgan Stanley India Company Private Limited having Corporate Identification No (CIN) U22990MH1998PTC115305, regulated by the Securities and Exchange Board of India ('SEBI') and holder of licenses as a Research Analyst (SEBI Registration No. INH000001105); Stock Broker (SEBI Stock Broker Registration No. INZ000244438), Merchant Banker (SEBI Registration No. INM000011203), and depository participant with National Securities Depository Limited (SEBI Registration No. IN-DP-NSDL-567-2021) having registered office at Altimus, Level 39 & 40, Pandurang Budhkar Marg, Worli, Mumbai 400018, India; Telephone no. +91-22-61181000; Compliance Officer Details: Mr. Tejarshi Hardas, Tel. No.: +91-22-61181000 or Email: tejarshi.hardas@morganstanley.com; Grievance officer details: Mr. Tejarshi Hardas, Tel. No.: +91-22-61181000 or Email: msic-compliance@morganstanley.com which accepts the responsibility for its contents and should be contacted with respect to any matters arising from, or in connection with, Morgan Stanley Research, and their affiliates (collectively, "Morgan Stanley"). Morgan Stanley India Company Private Limited (MSICPL) may use AI tools in providing research services. All recommendations contained herein are made by the duly qualified research analysts.

For important disclosures, stock price charts and equity rating histories regarding companies that are the subject of this report, please see the Morgan Stanley Research Disclosure Website at www.morganstanley.com/eqr/disclosures/webapp/generalresearch, or contact your investment representative or Morgan Stanley Research at 1585 Broadway, (Attention: Research Management), New York, NY, 10036 USA.

For valuation methodology and risks associated with any recommendation, rating or price target referenced in this research report, please contact the Client Support Team as follows: US/Canada +1 800 303-2495; Hong Kong +852 2848-5999; Latin America +1 718 754-5444 (U.S.); London +44 (0)20-7425-8169; Singapore +65 6834-6860; Sydney +61 (0)2-9770-1505; Tokyo +81 (0)3-6836-9000. Alternatively you may contact your investment representative or Morgan Stanley Research at 1585 Broadway, (Attention: Research Management), New York, NY 10036 USA.