China AI accelerators are entering an inference-driven cycle of domestic substitution

Overview

This report focuses on China’s AI accelerator industry within Greater China semiconductors. Its core judgment is that China’s AI GPU market is entering a more commercialized phase, increasingly driven by inference economics. Export controls, domestic substitution, expanding cloud inference demand, and supply-chain localization are jointly enlarging the serviceable market for domestic AI accelerators. Morgan Stanley believes China still lags the U.S. by about two generations in chip process technology, but can narrow the effective gap through multi-chip design, advanced packaging, rack-scale system architecture, optical networking, and hardware-software co-optimization.

Core views

The report’s core view is that competition in AI accelerators should not be seen simply as a single policy theme, but should distinguish vendors with shipment scale, ecosystem credibility, customer stickiness, and pricing discipline. Cambricon is strongest in the ASIC/DSA route, inference performance, and cloud-customer anchoring; Iluvatar has solid commercial optionality thanks to its diversified foundry strategy, supply visibility, and CUDA-compatible migration capability; MetaX is a credible domestic GPGPU participant, but its ecosystem maturity still needs validation and its valuation appeal is relatively weaker.

Analysis framework

The report uses an “economics × execution” framework to evaluate domestic AI chip vendors. The economics dimension includes TCO, cost per token, TPS, and performance per watt/per dollar; the execution dimension includes access to advanced process capacity, software ecosystem maturity, depth of CSP customer relationships, and product roadmap credibility. The report also incorporates channel checks, CSP orders and pre-orders, tightening NVIDIA supply in China, changes in GPU rental and token pricing, and near-term industry signals such as expectations for new products at WAIC.

Methodology notes

• vendor_selectionEconomics × Execution

  Select AI accelerator winners by jointly assessing quantified inference economics and qualitative commercial execution.

  Chip specifications alone are insufficient to determine share; vendors must be competitive simultaneously in TCO, cost per token, TPS, software migration, supply assurance, and CSP relationships.

• performance_benchmarkTPS inference performance framework

  Measure end-to-end token throughput using the DeepSeek R1 inference scenario.

  The model uses variables such as effective compute, memory bandwidth, interconnect bandwidth, chip utilization, model size, number of active MoE experts, input/output token length, and batch size, and calibrates against NVIDIA H200’s disclosed 5,899 TPS result in February 2025.

• cost_analysisTCO and cost-per-token analysis

  Compare chip procurement, power, infrastructure, and lifecycle usage costs.

  The report believes domestic AI accelerators can deliver 30-60% lower TCO than NVIDIA products procurable in China, and can achieve or exceed cost-per-token parity in some inference configurations.

Asset mapping & comparison

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

• Cambricon (688256.SS)

  Top domestic AI inference chip pick; initiated at Overweight.

  Strengths

  Strong cloud inference performance; products such as MLU590/MLU690 are competitive in some DeepSeek R1 scenarios; deep cooperation with customers such as ByteDance creates hardware-software synergy and validation through real deployments.

  Weaknesses

  The ASIC/DSA route is relatively dependent on customer scenarios and software adaptation; if pricing competition intensifies or customer deployment pace slows, earnings leverage may come under pressure.

  Comparison

  The report believes it has strong traction in current cloud inference, second only to Huawei Ascend.

  Risks

  Slowing AI demand, earlier pricing pressure, changes in policy or export environment, customer concentration, and risks to large-scale deployment conversion.

• Iluvatar CoreX Semiconductor (9903.HK)

  Top domestic AI GPU pick; initiated at Overweight.

  Strengths

  A diversified foundry strategy improves supply-chain resilience, while a TSMC-compliant production path enhances capacity visibility; TianGai-150 has received pre-orders from leading CSPs, and CUDA compatibility reduces migration friction.

  Weaknesses

  Shipments are expected to begin in 2H26, and commercial execution and software stability still require validation through actual deployments.

  Comparison

  Compared with peers relying only on domestic foundries or non-compliant overseas manufacturing, its supply visibility is stronger.

  Risks

  Delays in mass production and delivery, slower-than-expected customer migration, pricing competition, and restrictions in advanced supply chains.

• MetaX (688802.SS)

  Differentiated domestic GPGPU participant; initiated at Equal-weight.

  Strengths

  Its CUDA-like software stack and compatibility layer provide a relatively clear migration path, with ongoing progress in compiler adaptation, PyTorch compatibility, and runtime optimization; a mature-node manufacturing strategy may bring better yields and supply stability.

  Weaknesses

  Ecosystem maturity and stability still lag global leaders, large-scale commercial deployment remains to be validated, and valuation appeal is weaker than peers.

  Comparison

  It is credible on the GPGPU route, but the report rates it below Cambricon and Iluvatar.

  Risks

  Software ecosystem progress falling short of expectations, insufficient orders from major customers, mature-node limits on peak performance, and valuation compression.

• China domestic AI accelerator industry

  Benefiting from growth in AI inference demand, export controls, and domestic substitution.

  Strengths

  China has relative advantages in server systems, software optimization, AI data centers, power costs, and policy support; system-level innovation can partially offset process disadvantages.

  Weaknesses

  Key links such as front-end wafer fabrication and HBM/LPDDR5 remain relatively behind, and the industry may enter price competition earlier.

  Comparison

  There is still an absolute technology gap versus leading U.S. platforms, but within China’s procurable supply universe it has economic competitiveness.

  Risks

  Slowing demand, product homogenization, industry consolidation, export controls, and supply-chain bottlenecks.

• NVIDIA products procurable in China

  The main performance and cost benchmark for domestic vendors.

  Strengths

  Software ecosystem, CUDA developer base, compilers, and communication libraries remain core moats, with absolute leadership at the cutting edge.

  Weaknesses

  Products procurable in China are affected by export restrictions, with tightening supply and higher TCO.

  Comparison

  The report believes leading domestic accelerators can achieve cost-per-token parity or partially outperform within the universe of NVIDIA products procurable in China, but this does not mean surpassing NVIDIA’s most advanced global systems.

  Risks

  If NVIDIA improves product supply to China or lowers prices, the pace of domestic substitution may be affected.

Key data

• China AI chip TAM in 2030US$67bnChina AI chip serviceable market forecast under the report’s framework.
• Domestic self-sufficiency rate in 203086%The report expects China’s domestic AI accelerator self-sufficiency to rise to this level.
• Domestic AI chip TCO advantage30-60% lowerRelative to NVIDIA solutions currently procurable in China.
• NVIDIA H200 calibration benchmark5,899 TPSFebruary 2025 DeepSeek R1 inference result, used to calibrate the TPS model.
• Cambricon rating and target priceOverweight; Rmb1,588Initiation of coverage.
• Iluvatar rating and target priceOverweight; HK$600Initiation of coverage.
• MetaX rating and target priceEqual-weight; Rmb758Initiation of coverage.
• WAIC timingJuly 2026, ShanghaiThe report expects to see next-generation China AI accelerator products, especially from Iluvatar.

Impact & implications

If the report’s judgment proves correct, the investment thesis for China AI accelerators will shift from policy-driven valuation expansion to fundamentals validation around order conversion, software migration, customer stickiness, and inference economics. Procurement decisions by cloud service providers and large-model developers will depend more on deployable cost per token and software adaptation costs, rather than simply peak computing power. Vendors with stronger domestic supply-chain resilience, real CSP deployments, and the ability to control pricing pressure are more likely to gain sustained share.

Risks

• AI demand growth is slower than expected.
• Pricing competition begins earlier than expected and compresses revenue and margins.
• Policy, export controls, or supply-chain restrictions create uncertainty.
• Domestic vendors’ software ecosystems and large-scale deployment stability fall short of expectations.
• CSP orders, pre-orders, or customer migration pace come in below expectations.
• Industry homogenization accelerates consolidation, marginalizing weaker vendors.

What to watch

• Progress on next-generation domestic AI accelerator launches at WAIC in Shanghai in July 2026, especially Iluvatar’s products.
• Changes in NVIDIA GPU supply, spot prices, and procurable products in China.
• Mainstream large-model token prices, GPU rental prices, and the strength of inference demand.
• China CSP capex and allocation of orders for domestic AI chips.
• Execution by Cambricon, Iluvatar, and MetaX in software adaptation, customer deployment, shipment scale, and pricing discipline.
• Whether domestic AI chip vendors expand from inference into some training workloads.