China's AI GPUs are accelerating the narrowing of the gap with the US through market expansion, cost advantages, and system-level innovation

Overview

The report examines how China's AI GPU industry is narrowing the gap with the US in terms of market size, inference costs, system architecture, and infrastructure. Its core view is that Chinese AI demand continues to expand rapidly, and even though domestic solutions are still catching up at the chip level, they can build stronger overall competitiveness through lower prices, system-level scaling, and local infrastructure.

Core views

The report begins with market size and industry capitalization: the market capitalization of companies related to China's AI GPU industry is growing, more IPOs are expected in the future, and the market share of domestic AI accelerators in 2026 has also become an important dimension to watch. Morgan Stanley expects China's total addressable market for AI chips to increase to US$91bn by 2030. Its demand framework covers Chinese cloud service providers, telecom operators, government entities and state-owned enterprises, as well as overseas capital expenditure, indicating that the growth outlook does not rely on a single customer group but is based on computing power procurement demand across multiple customer categories. The report then uses a set of high-frequency indicators to track recent demand strength. The continued rise in NVIDIA 5090 prices in the Chinese market and the surge in monthly token volumes processed by ByteDance's Volcano Engine and Doubao are both used to indicate that demand for AI computing power remains robust. The report also examines the average token prices of China's mainstream large models and China's cloud capital expenditure trend relative to the global market to assess whether demand growth can translate into sustained chip procurement and infrastructure investment. In its comparison between China and the US, the report does not focus solely on individual chips. Instead, it compares nine factors across three levels: chips, systems, and infrastructure, and characterizes the current shift as a gradual decoupling of AI computing ecosystems. Its assessment is that a market-perceived gap remains at the chip technology level, but local infrastructure and system integration capabilities are narrowing this gap. Therefore, evaluating the competitiveness of domestic AI computing cannot rely solely on comparing peak chip performance; it also requires examining cluster scale, interconnect methods, and the operating efficiency of the overall computing system. Inference economics is an important basis for the report's view supporting domestic substitution. Based on company information and Morgan Stanley Research estimates, domestic chips have a lower total cost of ownership when used for Chinese large-model inference, meaning lower costs after accounting for computing hardware and operating expenditures; their cost per token is comparable to NVIDIA processors. Because domestic chips are priced significantly lower, they offer stronger performance per unit of cost. This means that even if differences in individual chip capabilities remain, domestic solutions can still achieve competitive real-world inference costs through pricing and system configuration. System-level scaling further supports this logic. The report discusses Huawei CloudMatrix 384 A3 and focuses on Huawei's Atlas 950 SuperPod showcased at the 2026 World Artificial Intelligence Conference. Atlas 950 expands the SuperPod to more than 1,000 NPUs, and its rear view demonstrates the optical interconnect used by UBlink. The report also compares various SuperPod solutions and midplane-free orthogonal designs presented at WAIC 2026, concluding that China is developing an active supernode ecosystem. The research focus extends from individual chip differences to large-scale clusters, optical interconnects, and system architecture, indicating that infrastructure innovation can provide a path to offsetting the chip gap. Considering market expansion, demand indicators, cost comparisons, and system capabilities, the report assigns an Attractive view to the Greater China technology semiconductor industry over the next 12—18 months. The main implication is that China's AI computing industry is developing a local value chain that is relatively independent of the US ecosystem, and the commercial competitiveness of domestic accelerators will depend jointly on demand growth, pricing advantages, inference costs, and system scaling capabilities rather than on any single chip specification.

Analysis framework

The report sequentially estimates market size, breaks down customer demand, and tracks recent demand indicators before comparing the Chinese and US AI computing ecosystems across the three levels of chips, systems, and infrastructure. It then evaluates the economics of domestic chips using TCO, cost per token, and performance per unit of cost, tests the system-level scaling path through Huawei Atlas 950 and the SuperPod solutions showcased at WAIC 2026, and finally forms its industry view.

Methodology notes

• Industry/Sector Analysis FrameworkSupply-demand framework

  AI chip TAM and customer demand breakdown

  The report estimates China's total addressable AI chip market in 2030 and analyzes sources of demand across Chinese cloud service providers, telecom operators, government entities and state-owned enterprises, as well as overseas capital expenditure, to assess the breadth of market growth.

• Competition and Strategy FrameworkValue chain analysis

  Nine-factor comparison across chips, systems, and infrastructure

  The report breaks down China-US AI competition into three levels—chips, systems, and infrastructure—to avoid measuring overall computing capabilities solely by individual chip performance and uses this framework to analyze the increasing independence of the local AI computing value chain.

• Industry/Sector Analysis FrameworkSubstitution Effect Analysis

  Comparison of inference economics between domestic chips and NVIDIA processors

  The report compares TCO, cost per token, and performance per unit of cost to determine whether domestic chips can form viable substitutes through lower pricing and system configuration.

• Cycle and Business Conditions Framework

  Recent AI GPU market tracking indicators

  The report tracks NVIDIA 5090 prices in China, monthly token volumes on ByteDance platforms, average token prices of mainstream large models, and cloud capital expenditure in China and overseas to observe recent changes in computing power demand and industry conditions.

Asset mapping & comparison

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

• China's Domestic AI GPU and AI Accelerator Industry

  The report expects the industry's TAM, domestic market share, market capitalization, and number of listed companies to continue growing.

  Strengths

  Lower inference TCO, cost per token comparable to NVIDIA processors, and stronger performance per unit of cost due to significantly lower pricing.

  Weaknesses

  A market-perceived gap with the US remains at the chip technology level.

  Comparison

  Compared with US solutions, China relies more on pricing, system scaling, and infrastructure capabilities to narrow the overall gap.

• Huawei Atlas 950 SuperPod

  A representative case of domestic AI system-level scaling, used to illustrate how infrastructure and cluster design can offset the individual chip gap.

  Strengths

  Scalable to more than 1,000 NPUs and demonstrates the optical interconnect used by UBlink.

  Comparison

  The report places it within an industry comparison of various SuperPod solutions and midplane-free orthogonal designs presented at WAIC 2026.

• NVIDIA Processors in the Chinese Market

  A benchmark for comparing domestic AI chips in terms of performance, TCO, cost per token, and performance per unit of cost.

  Weaknesses

  The report believes that NVIDIA processors have higher TCO than domestic chips in Chinese inference scenarios and are disadvantaged in performance per unit of cost because domestic solutions are priced significantly lower.

  Comparison

  Domestic chips have comparable cost per token, while the price of NVIDIA 5090 in China continues to rise.

Key data

• China AI Chip TAMUS$91bnMorgan Stanley's forecast for 2030.
• Atlas 950 SuperPod ScaleMore than 1,000 NPUsThe system-level scaling solution showcased by Huawei at WAIC 2026.
• Domestic Chip Inference TCOLower than NVIDIA processorsComparison for Chinese AI large-model inference scenarios.
• Domestic Chip Cost per TokenComparable to NVIDIA processorsBased on company information and Morgan Stanley Research estimates.
• NVIDIA 5090 Price Trend in ChinaContinuing to riseThe report uses this as an indicator for tracking recent AI computing demand.
• ByteDance-Related Monthly Token VolumeSurgingChanges in processing volumes at Volcano Engine and Doubao are used to indicate the strength of Chinese AI demand.
• Industry ViewAttractiveApplicable to the Greater China technology semiconductor industry, with a view horizon of the next 12—18 months.

Impact & implications

The report believes that competition in China's AI GPU industry is shifting from individual chip performance toward comprehensive competition across pricing, inference costs, cluster scale, interconnects, and infrastructure. Expanding demand and improving economics of domestic solutions should help increase the share of local accelerators and support market capitalization growth and more IPOs among related companies; meanwhile, system-level innovations such as SuperPods may accelerate the divergence of China's AI computing ecosystem from the US supply ecosystem.

What to watch

• Track the market share and revenue trends of domestic Chinese AI accelerators.
• Monitor changes in NVIDIA 5090 prices in the Chinese market.
• Track the monthly number of tokens processed by ByteDance's Volcano Engine and Doubao.
• Monitor the average token prices of China's mainstream large models.
• Compare capital expenditure trends among Chinese and global cloud service providers.
• Monitor progress in cluster scale, optical interconnects, and system architecture for Atlas 950 and other SuperPod solutions.