Domestic substitution in China’s AI accelerators has entered the commercialization selection phase

Overview

This report focuses on China’s AI accelerator and AI GPU industry, discussing which domestic vendors are more likely to gain meaningful market share against the backdrop of growing inference demand, ongoing export controls, and accelerating domestic substitution. The report argues that competition in AI computing in China is no longer just about single-chip specifications; system-level design, advanced packaging, rack-level architecture, network interconnects, software adaptation, and customer collaboration are narrowing the effective gap with U.S. solutions.

Core views

The report’s core view is that China’s AI accelerator market is entering a more commercialized stage, with customers increasingly prioritizing deployable economics rather than simply pursuing peak computing power. Morgan Stanley believes inference use cases will become the main demand driver. Domestic chips have a 30-60% TCO advantage over NVIDIA products available in China and can achieve cost per token close to or better than A100/H20-class products in some inference configurations. From an investment perspective, the report does not recommend viewing the sector as a single policy theme, but instead distinguishing winners with shipment scale, ecosystem credibility, pricing discipline, and customer relationships.

Analysis framework

The report uses a two-dimensional framework of “economics × execution” to screen domestic AI chip vendors. The economics dimension includes TCO, cost per token, TPS, performance per dollar, and performance per watt; the execution dimension includes access to advanced process technology or capacity, software ecosystem maturity, depth of CSP customer relationships, credibility of product roadmaps, and supply chain visibility. The report also incorporates channel checks, customer pre-orders, NVIDIA supply tightness, GPU rental prices, token pricing, and expectations for new products at WAIC to assess near-term demand.

Methodology notes

• Investment screening frameworkEconomics × Execution

  Use quantified inference economics and qualitative commercial execution capability together to judge AI accelerator vendors’ probability of success.

  Economics covers metrics such as TCO, cost per token, TPS, and performance per watt; execution covers foundry capacity, software ecosystem, CSP relationships, and roadmap credibility. The report believes that only vendors strong on both dimensions are more likely to gain sustainable share.

• Performance estimation frameworkTPS inference performance model

  Measure end-to-end tokens per second using the DeepSeek R1 inference workload.

  Model inputs include effective computing power, memory bandwidth, interconnect bandwidth, chip utilization, model size, number of layers, number of activated MoE experts, input/output token length, and batch size, calibrated using NVIDIA’s disclosed 5,899 TPS result for H200 in February 2025.

• Cost frameworkTCO and cost per token

  Assess customers’ true deployment economics by combining chip procurement, power, and infrastructure costs with actual inference throughput.

  The report believes domestic accelerators have lower chip prices, power, and infrastructure costs in the China market, resulting in lifecycle TCO that is 30-60% lower than NVIDIA solutions available in China, though the gap narrows when converted to cost per token.

Asset mapping & comparison

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

• Cambricon Technology Corporation (688256.SS)

  Top pick; initiating coverage with Overweight

  Strengths

  Leading inference performance, with MLU590/MLU690 showing strong TPS in some scenarios; deep collaboration with customers such as ByteDance; joint optimization of hardware and software; well suited for large-scale cloud inference deployment.

  Weaknesses

  Its path is more ASIC/DSA-oriented, so ecosystem generality and migration flexibility may be weaker than GPGPU; it still faces pressure from price competition and ongoing product iteration.

  Comparison

  The report believes its current traction in cloud inference is the strongest among domestic players, second only to Huawei Ascend.

  Risks

  Slowing AI demand, price declines, customer concentration, weaker-than-expected software optimization, and changes in policy and export restrictions.

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

  Top pick; initiating coverage with Overweight

  Strengths

  A diversified foundry strategy provides higher supply visibility; TianGai-150 has received advance orders from leading CSPs; its GPGPU architecture offers relatively high CUDA compatibility; this helps customers migrate from the NVIDIA platform.

  Weaknesses

  Execution of shipments in 2H26 still needs to be validated; commercial scale-up and software ecosystem stability still require continued proof.

  Comparison

  The report believes Iluvatar’s advantages mainly lie in supply chain resilience and commercial optionality, especially for enterprise customers needing an alternative path to NVIDIA.

  Risks

  Pre-orders converting below expectations, delays in mass production, price competition, unstable software migration experience, and supply chain or compliance changes.

• MetaX Integrated Circuits (688802.SS)

  Initiating coverage with Equal-weight

  Strengths

  Credible domestic GPGPU positioning; a CUDA-like software stack and compatibility layer reduce migration friction; uses a more mature process node to improve yield and supply stability.

  Weaknesses

  Ecosystem maturity and stability still lag global leaders; peak performance may be constrained by mature process technology; valuation appeal is weaker than Cambricon and Iluvatar.

  Comparison

  The report recognizes MetaX’s differentiation in a scalable GPGPU path, but believes its risk-reward is less compelling than the Overweight names.

  Risks

  Insufficient validation of large-scale commercial deployment, slow progress in ecosystem compatibility, widening performance gaps, valuation digestion pressure, and price competition.

• NVIDIA China-accessible products

  Primary competitive benchmark

  Strengths

  Its software ecosystem, CUDA, compilers, libraries, developer base, and high-end performance still have significant advantages.

  Weaknesses

  Products available in China are affected by export restrictions; tight supply, elevated prices, and higher local deployment costs increase the appeal of domestic substitution.

  Comparison

  The report compares only within the range of products reasonably available in China and explicitly does not include cutting-edge platforms such as GB300 in this domestic substitution conclusion.

  Risks

  If NVIDIA supply available to China improves or prices decline, the pace of domestic substitution may slow.

Key data

• China AI chip TAMUS$67bn by 2030The report’s framework estimates that China’s total addressable market for AI chips could reach this scale by 2030.
• Domestic self-sufficiency rate86% by 2030The report expects the self-sufficiency rate of domestic AI accelerators to rise to 86% by 2030.
• TCO advantage of domestic AI chips30-60% lower TCOCompared with NVIDIA solutions currently available in China, domestic accelerators have advantages in procurement, power, and infrastructure costs.
• NVIDIA H200 calibration result5,899 TPSThe report uses NVIDIA’s DeepSeek R1 inference result disclosed in February 2025 to calibrate the TPS model.
• Performance of leading domestic chips relative to H2050-150% outperformance in selected scenariosIn some DeepSeek R1 inference scenarios assumed in the report, leading domestic accelerators such as Huawei Ascend 950PR/DT and Cambricon MLU690 can outperform NVIDIA H20.
• Cambricon rating and target priceOverweight; Rmb1,588Initiating coverage; the report is positive on its inference performance, CSP customer stickiness, and cloud deployment economics.
• Iluvatar rating and target priceOverweight; HK$600Initiating coverage; the report is positive on its supply chain resilience, order visibility, and software-compatible migration path.
• MetaX rating and target priceEqual-weight; Rmb758Initiating coverage; the report recognizes its GPGPU positioning and CUDA-like compatibility, but believes its valuation appeal is relatively limited.

Impact & implications

The investment implication of the report is that China’s AI accelerator industry is shifting from a policy narrative to commercial execution, and the beneficiaries will be vendors that can simultaneously prove themselves in inference economics, customer migration costs, supply availability, and software stability. Procurement standards among cloud service providers and leading LLM developers are more tilted toward cost per token, software maturity, and depth of strategic cooperation; sovereign AI, telecom, SOE, and government-related demand places greater emphasis on supply security, domestic controllability, and policy alignment.

Risks

• Growth in AI inference and training demand comes in below expectations.
• Price competition emerges earlier than expected, putting pressure on industry margins and valuations.
• Changes in policy, export controls, or supply chain compliance affect product manufacturing and customer procurement.
• Progress in domestic software ecosystems, framework compatibility, and cluster-level optimization is slower than expected.
• CSP customer orders or pre-orders fail to convert smoothly into revenue.
• NVIDIA or other global leaders regain a wider advantage in available products, pricing, or software ecosystems.

What to watch

• New domestic AI accelerator products at Shanghai WAIC in July 2026, especially Iluvatar’s next-generation products.
• The start of TianGai-150 shipments in 2H26, customer acceptance, and revenue recognition for Iluvatar.
• Cambricon’s ongoing share in large-scale cloud inference deployments, customer expansion, and TPS/cost-per-token performance.
• MetaX’s CUDA-like software stack, PyTorch compatibility, compiler adaptation, and validation of large-scale commercial deployment.
• NVIDIA GPU supply in China, spot prices of RTX 5090, GPU rental prices, and token price trends.
• CSP capital expenditure, growth in token volumes for applications such as ByteDance/Doubao, and the pace of commercialization of mainstream LLMs in China.
• Industry price discounts, order competition, and signs of potential consolidation.