China’s AI semiconductor competition is upgrading from chip specs to SuperPod system capabilities

Overview

This report summarizes Morgan Stanley’s key observations from WAIC Shanghai 2026 and its China technology field research on July 16–17. The report believes the domestic AI computing market has moved beyond competition based on single-accelerator specifications into a new stage centered on SuperPod, scale-up interconnects, memory sharing, workload scheduling, fault protection, and full-system utilization. Although China’s AI GPUs remain constrained at the wafer process level, the country has relative advantages in optical networking, server rack design, advanced packaging, and system architecture, which are expected to support the expansion of domestic AI computing clusters.

Core views

Core views include: first, WAIC 2026 saw fewer major chip launches but significantly more SuperPod solutions, with nearly all domestic vendors showcasing 64-card or 128-card scale-up systems; second, Huawei Atlas 950 expands SuperPod scale from CloudMatrix 384’s 384 Ascend 910C chips to 1,024 next-generation Ascend 950DT NPUs, becoming a key hardware foundation for domestic large-model training; third, the separation of Prefill and Decode is emerging as a new theme in AI inference, helping operators optimize hardware and scheduling separately for compute-intensive prefill and bandwidth/latency-sensitive decode; fourth, domestic AI GPUs, advanced-node foundries, equipment, advanced packaging, and data-center-related analog/optoelectronic chips all benefit from AI demand and import substitution.

Analysis framework

The report uses conference observations, management discussions, product specification comparisons, supply chain mapping, and stock rating linkages to assess investment implications across AI accelerators, SuperPod interconnects, advanced packaging, foundry capacity, semiconductor equipment, CIS, and analog chips.

Methodology notes

• Supply chain researchAI semiconductor supply chain mapping

  Transmission from AI accelerators to system-level enabling segments

  The report maps domestic AI computing demand across AI GPU/NPU, scale-up interconnects, advanced-node foundries, HBM and memory, semiconductor equipment, advanced packaging, and optoelectronic analog components to judge the degree of benefit for each company.

• Competitiveness analysisSystem-level scaling capability assessment

  SuperPod and scale-up architecture

  The report compares more than single-chip performance, emphasizing the importance of interconnect topology, memory pooling, optical/copper links, fault protection, communication efficiency, and cluster utilization for large-model training and inference.

• Event studyWAIC conference takeaway tracking

  Combining product launches with management discussions

  The report combines on-site WAIC demonstrations and China technology field research meetings to extract marginal changes for companies or segments such as Huawei, Iluvatar, Oriental Computing, AMEC, Hua Hong, JCET, and OmniVision.

Asset mapping & comparison

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

• Cambricon Technology Corporation (688256.SS)

  Core beneficiary in domestic AI computing, rated OW

  Strengths

  Benefits from domestic AI GPU demand, order visibility, and import substitution.

  Weaknesses

  The report excerpt does not provide specific financial forecasts or target price.

  Comparison

  Compared with MetaX, the report prefers Cambricon; MetaX is rated EW.

  Risks

  Uncertainty around demand timing, supply chain constraints, product performance, and customer deployment progress.

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

  Key beneficiary in domestic AI accelerators, rated OW

  Strengths

  Tiangai 300 covers decode and post-training, with some metrics said by management to exceed H100; shipments are expected at about 100,000 cards this year and not lower next year.

  Weaknesses

  Tiangai 300 still trails Blackwell; progress in large training clusters is relatively cautious.

  Comparison

  Tiangai 400 is planned to tape out in 2027 and benchmark against the Blackwell generation.

  Risks

  U.S. policy, domestic manufacturing progress, HBM costs, customer validation, and supply chain resources remain key risks.

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

  Preferred domestic AI computing name, rated OW

  Strengths

  The report lists it as one of the preferred OW names in domestic AI computing.

  Weaknesses

  The excerpt lacks specific product specifications and operating data.

  Comparison

  Included in the preferred list together with Cambricon and Iluvatar.

  Risks

  Order conversion, supply chain, and changes in the competitive landscape.

• MetaX Integrated Circuits (688802.SS)

  Domestic AI computing-related name, rated EW

  Strengths

  Participates in the domestic AI computing supply chain.

  Weaknesses

  The report explicitly prefers Cambricon, Iluvatar, and Hygon within domestic AI computing, while maintaining EW on MetaX.

  Comparison

  Less attractive relative to the OW names in the preferred list.

  Risks

  Product competitiveness, order visibility, and the pace of customer deployment.

• SMIC (0981.HK)

  Beneficiary of advanced-node capacity expansion, rated OW

  Strengths

  Demand for advanced-node capacity from domestic AI semiconductors is rising, and the report lists SMIC among its preferred key enabling names.

  Weaknesses

  Advanced processes and capacity expansion are affected by equipment, yields, and external restrictions.

  Comparison

  Superior within foundries to Hua Hong, which is rated only EW.

  Risks

  Export controls, capital expenditure efficiency, customer demand volatility, and advanced-node ramp-up.

• Hua Hong Semiconductor Ltd (1347.HK)

  Beneficiary of specialty-process and AI PMIC demand, rated EW

  Strengths

  8-inch fabs are running at about 110% utilization, prices have risen about 10–15% since the end of last year, and AI PMIC demand supports specialty wafer pricing.

  Weaknesses

  Power semiconductor demand is relatively weak, and depreciation on new 12-inch capacity is rising.

  Comparison

  The report acknowledges demand and price recovery, but the rating remains EW, below SMIC’s OW.

  Risks

  Weakness in power products, 9A/9B capacity ramp-up, HLMC Fab 5 integration, and depreciation pressure.

• Advanced Micro-Fabrication Equipment Inc (AMEC, 688012.SS)

  Beneficiary of domestic semiconductor equipment substitution, rated OW

  Strengths

  Three-year demand visibility from memory customers is strong, advanced logic investment is expected to remain robust over the next five years, and products are expanding into ALD, CMP, inspection, and metrology.

  Weaknesses

  Some 2027 demand has not yet turned into formal orders, and overseas component lead times are lengthening.

  Comparison

  Together with NAURA, ACM Research, and ASMPTc, it is among the preferred equipment names.

  Risks

  Delays in customer capex, the pace of import substitution, delivery, and uncertainty around commercialization of new platforms.

• JCET Group Co Ltd (600584.SS)

  Beneficiary of advanced packaging expansion, rated EW

  Strengths

  Announced a CNY7.8bn advanced packaging and testing investment in Lingang, Shanghai, and management believes China AI GPU ramp-up is still in the early stage.

  Weaknesses

  OSAT pricing follows the market, and substrate tightness may affect new capacity.

  Comparison

  Part of the beneficiary chain for advanced AI chip packaging, but rated EW.

  Risks

  Utilization of new capacity, substrate supply, returns on capex, and customer qualification timing.

• OmniVision Integrated Circuits Group Inc (603501.SS)

  CIS and analog/optoelectronic conversion-related name, rated EW

  Strengths

  Action cameras, machine vision, AR/VR, medical imaging, and analog products for data center optical modules provide growth opportunities.

  Weaknesses

  Smartphone CIS demand is weak, with insufficient Android shipments, model launches, and specification upgrades.

  Comparison

  The path for analog products in data center optical modules is relatively clear, while near-term smartphone growth contribution is limited.

  Risks

  Most products are still under development or in early qualification, and customer adoption and scalable profitability need time to be validated.

Key data

• Atlas 950 scale1,024 NPUsHuawei Atlas 950 consists of 16 compute racks, each with 64 NPUs, and is equipped with 4 UnifiedBus interconnect racks.
• Atlas 950 compute powerAbout 1 EFLOPS FP8, 2 EFLOPS FP4Each 64-NPU compute rack delivers up to 64 PFLOPS FP8 and 128 PFLOPS FP4.
• Atlas 950 memoryAbout 96TB HBM, with up to 256TB total addressable memoryHBM and external DRAM memory pooling improve available memory for large-model training.
• Oriental Computing DF10006.4TB/s memory bandwidth, 520TFLOPS BF16Uses a domestic 14+nm-class process and wafer-level hybrid bonding to vertically stack DRAM and logic dies.
• Expected China AI GPU TAMUS$91bn in 2030The report’s chart expects China’s total AI GPU market size to grow to US$91bn by 2030.
• Expected China AI chip self-sufficiency rate70% by 2030eThe report’s chart expects China’s AI chip self-sufficiency rate to reach 70% by 2030.
• Iluvatar shipmentsAbout 100,000 cards this year, next year not lower than this yearThe company prioritizes scalable inference deployment rather than aggressively entering large training clusters.
• Hua Hong 8-inch line utilizationAbout 110%Its three 8-inch fabs are running at around 110% utilization, and wafer prices have risen about 10–15% since the end of last year.
• JCET advanced packaging investmentCNY7.8bn, about US$1.1bnFor a new advanced packaging and testing plant in Lingang, Shanghai, reinforcing AI computing and high-performance chip capacity.

Impact & implications

In terms of investment implications, the report is more constructive on companies that directly benefit from domestic AI computing demand, order visibility, and system-level scaling capabilities. Among AI accelerators, it prefers Cambricon, Iluvatar, and Hygon; expansion in advanced-node capacity benefits SMIC; AI PMIC and specialty-process demand support Hua Hong pricing, though the rating remains EW; domestic investment in memory and logic manufacturing continues to benefit NAURA, AMEC, ACM Research, and ASMPTc; advanced packaging expansion benefits OSAT players such as JCET, though supply, substrate tightness, and new capacity ramp-up need monitoring.

Risks

• U.S. executive orders, export controls, and related compliance restrictions may affect trading in certain securities, supply chains, and access to technology.
• Domestic AI GPUs remain constrained by process nodes, HBM, advanced packaging, capacity, and external policy.
• As SuperPod scales up, communication efficiency, fault protection, software scheduling, and system utilization become key risks to training stability.
• Some companies’ orders and 2027 demand have not yet fully converted into formal orders, creating demand realization risk.
• Changes in wafer, HBM, substrate, and overseas component costs or lead times may compress margins or affect delivery.
• New capacity ramp-up, rising depreciation, and customer qualification progress may affect profit realization for foundry, equipment, and packaging companies.

What to watch

• Actual deliveries, customer deployment, and large-model training stability of Huawei Atlas 950 and the Ascend 950DT platform.
• Interconnect efficiency, software stack maturity, and commercial orders for 64-card, 128-card, and larger SuperPod solutions from domestic vendors.
• Core customer feedback for Iluvatar Tiangai 300, the 100,000-card shipment target, and progress on the 2027 Tiangai 400 tape-out.
• Whether China CSP capex and large-model token demand continue to drive domestic AI GPU demand.
• SMIC advanced-node capacity expansion, Hua Hong 9A/9B ramp-up, and integration of the HLMC Fab 5 acquisition.
• The pace at which equipment companies such as AMEC convert memory, advanced logic, and localization awards from preliminary wins into purchase orders and shipments.
• JCET’s Lingang advanced packaging plant in Shanghai, JME 2.5D packaging capacity, and substrate supply conditions.