Investment Insights: Division of Labor and Layout of China-US AI Industry Chain Under Jensen Huang's "Five-Layer Cake" Framework

Jensen Huang points out that energy is the cornerstone of artificial intelligence; without electricity, it is impossible to support the expansion of computing power and data centers. Currently, China has approximately twice the installed power capacity of the United States, and can quickly put new facilities into operation through centralized approval and efficient engineering construction [1]. This means:

• Midstream energy equipment and new energy (wind power, photovoltaics, energy storage, substations)
  will receive continuous orders amid the large-scale expansion of AI;
• Intelligent power dispatch for cities and industrial parks, edge power sources (e.g., front-end gas turbines, hydrogen microgrids)
  are worth attention, especially projects co-built with AI data centers;
• Trading mechanisms such as carbon neutrality and green power certificates
  can provide premium services, benefiting power operators and energy management software companies.

In contrast, the United States faces long construction cycles, fragmented approval processes, and power gaps (Morgan Stanley estimates a gap of 47GW from 2025 to 2028) [3], making

key investment areas, while creating long-term demand for .

The United States maintains a leading position in chip design and advanced manufacturing processes, especially with absolute advantages in Blackwell and its subsequent architectures. Due to national security concerns, it still restricts exports to China; however, the latest approval for H200 exports shows a

. High-end computing power is still in the hands of the United States, while retaining “soft control over China” to maintain pricing power for major models [2].

• US manufacturers (Nvidia, AMD, Xilinx, Intel, etc.)
  can still continuously lock in demand from Chinese scientific research institutions and cloud factories for exportable models like H200;
• Chinese local chips (Ascend, Cambricon, KunLun, etc.) and packaging testing
  are encouraged by policies and are accelerating iteration in mid-to-low-end and specific computing power scenarios;
• For
  investors
  : Focus on domestic leaders in segmented fields such as
  EDA, domestic lithography alternatives, packaging modules, AI acceleration cards
  with technical/customer stickiness, as well as
  midstream system integrators
  that cooperate with US enterprises and can obtain H200 supporting facilities.

China excels in the “fast construction” capability of data centers, cold chains, network upgrades, and AI computing platform construction. Driven by central policies and local capital investment,

become the core of attention. Especially when AI applications penetrate into vertical industries, supporting the energy layer are expected to form rigid demand [1].
The United States relies on more mature cloud service providers (AWS, Azure, Google Cloud) and their high-end self-developed architectures in this layer. Investment focus can be placed on and other fields.

Currently, the United States still leads in super-large models (e.g., GPT, Gemini, Claude) with a lead cycle of about half a year, but China is quickly narrowing the gap through open-source communities and independent training [1]. From an investment perspective:

• Model training and tuning service providers (e.g., those controlling model lifecycle management, fine-tuning platforms)
  are expected to undertake the demand of Chinese and US enterprises or governments in multi-model parallel deployment;
• Algorithm infrastructure (e.g., efficient training frameworks, distributed scheduling and monitoring)
  is of great significance to cloud vendors and large technology companies;
• AI security and interpretability tools
  will become configuration points under model governance/compliance pressure.

In Jensen Huang’s view, Chinese people’s acceptance of AI applications reaches 80%, driving rapid implementation in healthcare, finance, retail, and manufacturing [1]. Therefore:

• Domestic platform enterprises (large cloud + scenario-based AI)
  have first-mover practical advantages and can achieve scale through SaaS/NaaS services;
• Industry solutions (intelligent manufacturing, supply chain finance, health AI)
  are built synchronously with infrastructure, suitable for layout of
  vertical clouds and industry AI service providers
  ;
• The United States can continue to deepen in
  high-end enterprise applications, AI security compliance, industrial automation, AI + robots
  and other directions, and gain part of the Chinese market share through exporting H200.

1. China:
   • Focus on tracking
     enterprises integrating energy + infrastructure dual chains
     (e.g., platforms with power + data center operation capabilities);
   • Seize the growth dividends of
     AI chip ecology (packaging testing, domestic accelerators), model governance tools, and industry AI solutions
     ;
   • Pay attention to
     system integrators and cloud service providers
     directly related to H200 orders, while keeping an eye on domestic alternatives supported by potential policies.

2. United States:
   • Focus on
     high-end chip design/IP, EDA tools, Blackwell’s subsequent architectures and Rubin series
     , relying on other global markets under restricted exports;
   • Invest in
     energy infrastructure (green power, self-built power plants, energy storage), data center transformation, and AI software platforms
     , leveraging leadership in models and applications;
   • Use freely exportable H200 to enter the Chinese market and extend the business chain through services and operation and maintenance.

• Policy risks
  : Export control policies fluctuate with political cycles; closely monitor relevant legislation from the US Department of Commerce and Congress (e.g., “Security Chip Act”);
• Technological substitution
  : If China’s independent R&D progress exceeds expectations, it may quickly reduce demand for imported chips such as H200;
• Energy constraints
  : If the United States cannot quickly fill the energy gap in the short term, it may drag down the large-scale deployment of AI and affect the valuation of related equipment/service manufacturers.

Jensen Huang’s “Five-Layer Cake” is not only an industrial logic but also outlines the division of labor and opportunities in different tracks for Chinese and US investors:

At the same time, the complementarity and competition between the two countries in model applications also leave space for cross-border cooperation and differentiated layout. It is recommended to continuously track policy/export control dynamics and core infrastructure expansion progress, and activate the “deep investment research mode” of Jinling AI when necessary to obtain more detailed industry data and comparative analysis.

[1] “Trump Approves H200 Export: Nvidia Lobbying Works, Domestic Computing Power Chips Face Pressure Test?” TF Finance, Link: https://www.tfcaijing.com/touch/article/page/654f46594f4e5868766342374f4933715a61753379673d3d
[2] “Trump Announces: US Will Allow Nvidia AI Chips to Be Exported to China”, RFI, Link: https://www.rfi.fr/cn/中国/20251208-美国计划批准英伟达h200高端芯片出口中国
[3] “US Chip Policy ‘Swing’ Cannot Conceal the Essence of Blocking China”, China Economic Net/Technology Daily, Link: http://intl.ce.cn/sjjj/qy/202512/t20251215_2642984.shtml