China's AI GPUs Are Accelerating the Closing of the Gap with the US Through Market Expansion, Cost Advantages, and Supernode Infrastructure

Overview

The report focuses on growth in China's AI GPU market, domestic substitution, inference costs, and supernode infrastructure, concluding that China is narrowing its AI computing gap with the US through market scale, economics, and systems engineering capabilities.

Core views

The report first concludes that China's AI GPU market is expanding and that domestic AI accelerators are gaining market share, while the market capitalizations of related companies are growing and more IPOs are expected. Morgan Stanley forecasts that China's AI chip total addressable market will reach US$91 billion by 2030. Its demand breakdown covers Chinese cloud service providers, telecom operators, sovereign and state-owned enterprises, and overseas capital expenditure, implying that the growth thesis does not depend on a single customer category but instead reflects joint investment by multiple computing-capacity purchasers. The report also assigns an Attractive view to the Greater China Technology Semiconductors industry, indicating that it is expected to be more attractive than the relevant broad market benchmark over the next 12—18 months. For near-term demand, the report establishes a tracking framework comprising market prices, token usage, and capital expenditure. Prices of Nvidia 5090 GPUs in the Chinese market continue to rise; monthly token volumes processed by ByteDance's Volcano Engine and Doubao have surged, indicating strong AI usage demand; the report also tracks average token prices for China's mainstream large models and compares cloud computing capital expenditure trends in China and globally. These indicators provide a cross-sectional view of the strength of China's AI GPU demand from different perspectives, including hardware prices, actual model usage, inference pricing, and infrastructure investment. The report argues that comparisons of AI computing between China and the US should not be limited to the performance of individual chips but should also examine chips, systems, and infrastructure. Its nine-factor comparison framework shows that China still faces a perceived technology gap at the chip level, but stronger infrastructure and systems integration capabilities are narrowing the overall gap. As the Chinese and US AI chip value chains gradually form relatively independent computing ecosystems, China can offset some single-chip limitations through large-scale system architecture, interconnects, and cluster engineering. Inference economics provides another core piece of evidence supporting the competitiveness of domestic chips. Based on company materials and Morgan Stanley Research estimates, domestic chips have a lower total cost of ownership in Chinese large-model inference scenarios, while their cost per token is comparable to that of Nvidia processors. Because domestic products are priced significantly lower, they deliver stronger performance per unit cost. The report therefore emphasizes that purchasers ultimately compare not only peak computing power but also total acquisition and operating costs, actual inference output, and the cost associated with each token. System-level progress is concentrated in supernode solutions. The report mentions Huawei CloudMatrix 384 A3 and highlights Huawei's new Atlas 950 SuperPod, unveiled at the 2026 World Artificial Intelligence Conference. Atlas 950 expands a single supernode to more than 1,000 NPUs, and its rear view demonstrates the optical interconnect used by UBlink. The report also compares various supernode solutions at the conference and examines orthogonal midplane-free designs, showing that Chinese vendors are developing a thriving supernode ecosystem through larger cluster scales, optical interconnects, and system architecture innovation. Together with the cost advantages of domestic chips, these infrastructure capabilities constitute the primary path toward narrowing the China-US AI computing gap.

Analysis framework

The report first forecasts China's AI chip TAM and breaks down demand sources by cloud service providers, telecom operators, sovereign and state-owned enterprises, and overseas capital expenditure; it then tracks near-term demand using GPU prices, token usage volumes, model pricing, and cloud capital expenditure; next, it compares Chinese and US capabilities across nine factors spanning chips, systems, and infrastructure; finally, it validates the economics and engineering progress of domestic AI computing using total cost of ownership, cost per token, performance per unit cost, and supernode solutions presented at the 2026 World Artificial Intelligence Conference.

Methodology notes

• Industry/Sector Analysis Framework

  AI Chip TAM Forecast

  The report estimates the long-term size of China's serviceable AI chip market and breaks down demand by different computing-capacity purchasers, arriving at a TAM forecast of US$91 billion for 2030.

• Competition and Strategy FrameworkValue chain analysis

  Comparison of China-US AI Chip Value Chains and Computing Ecosystems

  The report examines the Chinese and US AI computing ecosystems across chips, systems, and infrastructure, focusing on how China can use systems engineering and infrastructure to offset part of the chip technology gap.

• Competition and Strategy Framework

  Nine-Factor Comparison of China and the US in AI

  The report uses nine factors for a cross-country comparison, avoiding an assessment of overall AI computing capabilities based solely on the performance of individual chips.

• Industry/Sector Analysis Framework

  Analysis of Total Cost of Ownership for Inference and Cost per Token

  The report links chip acquisition and operating costs to actual inference output, comparing domestic chips with Nvidia processors using total cost of ownership, cost per token, and performance per unit cost.

• Cycle and Business Conditions Framework

  Near-Term AI GPU Demand Tracking Indicators

  The report tracks near-term demand changes using Nvidia 5090 prices, ByteDance's monthly token volumes, token prices for mainstream large models, and cloud capital expenditure trends.

Asset mapping & comparison

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

• China's Domestic AI GPU and Accelerator Industry

  The report believes it will benefit from growth in China's AI chip TAM, rising domestic market share, and increased investment by multiple categories of computing-capacity purchasers.

  Strengths

  Lower total cost of ownership, cost per token comparable to Nvidia processors, and stronger performance per unit cost due to significantly lower pricing.

  Weaknesses

  The report's comparison framework still acknowledges a perceived technology gap for China at the chip level.

  Comparison

  Compared with Nvidia processors for the Chinese market, domestic chips have greater advantages in inference economics and performance per unit cost.

• Huawei Atlas 950 SuperPod

  A representative supernode solution illustrating how China is using systems and infrastructure capabilities to narrow the AI computing gap.

  Strengths

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

  Comparison

  The report compares it with other supernode solutions presented at the 2026 World Artificial Intelligence Conference.

• Nvidia Processors for the Chinese Market

  The report uses them as the primary benchmark for comparing domestic AI chips in terms of price, total cost of ownership, and cost per token.

  Weaknesses

  Prices of the 5090 in the Chinese market continue to rise; the report estimates that domestic chips have a lower total cost of ownership.

  Comparison

  Domestic chips have a comparable cost per token but achieve stronger performance per unit cost through lower pricing.

Key data

• China AI Chip TAMUS$91bnMorgan Stanley's forecast for 2030.
• Greater China Technology Semiconductors Industry ViewAttractiveAn industry assessment indicating greater attractiveness relative to the relevant broad market benchmark over the next 12—18 months.
• Atlas 950 SuperPod ScaleMore than 1,000 NPUsHuawei's new supernode solution presented at the 2026 World Artificial Intelligence Conference.
• Total Cost of Ownership for Inference Using Domestic ChipsLower than Nvidia processorsFor Chinese AI large-model inference scenarios, based on company materials and Morgan Stanley Research estimates.
• Cost per Token for Domestic ChipsComparable to Nvidia processorsThe comparison result used by the report to measure actual inference economics.
• Performance per Unit Cost of Domestic ChipsStrongerThe report attributes this to significantly lower product pricing.
• Nvidia 5090 Prices in the Chinese MarketContinuing to riseA near-term AI GPU demand tracking indicator presented in the report.
• ByteDance's Monthly Token Processing VolumeSurgingThe report views token trends at Volcano Engine and Doubao as signals of strong AI demand in China.

Impact & implications

The report believes that expansion of China's AI chip market and diversified investment in computing capacity will provide greater commercial opportunities for domestic accelerators. Even if a technology gap remains at the individual-chip level, lower inference costs, stronger performance per unit cost, and advances in thousand-accelerator supernodes, optical interconnects, and system architecture can enhance the overall competitiveness of domestic computing ecosystems and help narrow the gap between China and the US.

What to watch

• Monitor whether Nvidia 5090 prices in the Chinese market continue to rise.
• Monitor changes in monthly token processing volumes at ByteDance's Volcano Engine and Doubao.
• Monitor average token prices for China's mainstream AI large models.
• Compare cloud computing capital expenditure trends in China and globally.