China's AI Development: A Seven-Layer Cake Model

The "Proposal on Formulating the 15th Five-Year Plan for National Economic and Social Development" (hereinafter referred to as the 15th Five-Year Plan) emphasizes AI strategy as the most important part of its technology mainline. To achieve a pioneering position in this industrial revolution, China should consider comprehensive development through a seven-layer cake model and implement it in the planning.

From China's current AI development speed, it has the strength to maintain a top-three position globally. Some large models' scores have approached or surpassed GPT and Gemini. NVIDIA CEO Jensen Huang even mentioned in a private conversation that China will win because sanctions are the best national mobilization order, with one million researchers working day and night to catch up.

Even with external optimism, we need to clearly see our current strengths and weaknesses. To ensure sufficient future momentum, we need comprehensive construction of upstream and downstream nodes from infrastructure to applications, and careful operation of the commercial ecosystem.

The Seven Layers

Therefore, the seven-layer cake model consists of: Energy, Chips, Infrastructure, Data, Models, Applications, and Trust Governance. To become a world-leading AI power, we must pass through seven gates to form industrial advantages.

Layer 1: Energy - The Physical Limit of Computing Power

China's electricity is relatively cheap. The United States builds power plants by state, with power grids and approvals being bottlenecks, while China has unified scheduling. Compared with developed countries, electricity prices are cheap - residential electricity is only one-tenth of Japan's, which greatly supports AI development.

However, this is still not enough. Although China has the world's largest power system, coal power accounts for a high proportion, and there is pressure to reduce emissions.

A founder who built AI data centers for Alibaba, ByteDance, and Deepseek said that recently, the biggest worry is where to place AI data factories - they need to be close to energy sources, cheap, and have liquid cooling systems. He even considered placing them in the sea.

PUE, power supply capacity, and expansion cycles are transforming from engineering parameters to business parameters, after all, the first principle of AI is thermodynamics.

Layer 2: Chips - The Foundation of Computing

The United States leads significantly, with top GPUs and integrated software ecosystems.

In the book "Chip War," author Chris Miller mentions that chip technology is the most critical technology competition globally, with leading manufacturers in each process accumulating decades of expertise to reach their positions.

A computing center founder stated that Chinese manufacturers lag behind NVIDIA by several years overall, with training computing power unusable and inference barely manageable with domestic chips.

China's shortcomings also include processes, packaging, and ecosystem migration costs, not just chips.

Chips also depend on ecosystems and usage efficiency. A large model company preparing for IPO spent most of its financing on purchasing chips. However, an employee proposed a method for efficiently utilizing GPUs, running several times more efficient tokens with the same hardware, and was competitively recruited by multiple manufacturers at sky-high salaries. This shows the market lacks better chip utilizers.

A former Huawei researcher said that NVIDIA's key is the CUDA system, which is difficult to learn in a short time. The Beijing Academy of Artificial Intelligence proposed the Flagopen heterogeneous computing power usage solution.

Layer 3: Infrastructure - From Chaos to Certainty

This layer transforms chaos into certainty and stability.

An entrepreneur invested in AI data factories, with a plug-and-play container model. While building an AI data center in the United States takes three years, his company only needs three months, greatly improving speed and efficiency.

Server cluster design, network design, database IO, and scheduling and operation systems are guarantees that this layer's computing power can run.

The entrepreneur has opened factories in Malaysia and is preparing to invest in building future AI data center factories in Hangzhou. This approach will greatly improve infrastructure efficiency.

China needs to address efficiency issues of long-distance networks, cross-regional scheduling, and east-west load matching.

Layer 4: Data - The Raw Material Warehouse

This is the raw material warehouse and industry ticket.

OpenAI co-founder Ilya Sutskever mentioned in last month's podcast interview that training data for large language models is about to run out. So what data will be used to train tens of thousands of new intelligent agents?

China has national Personal Information Protection Law and Data Security Law, achieving relatively unified national standards.

The United States does not have a single federal comprehensive privacy security law, and state laws vary greatly.

The real barriers lie in the collection, cleaning, annotation, governance, compliance, and iteration of data in various industries. These data pipelines can support algorithm updates at the bottom layer.

Layer 5: Models - The Cake Base

Models have high scores, but online instability, bugs that create new bugs when fixed - how to better train and fine-tune remains a research focus.

The Stanford AI Index 2025 report points out that China leads in AI papers and patents. For example, in 2023, China accounted for 69.7% of total AI patent grants, with the United States ranking second at 14.2%. As of April 2025, according to CCTV data, China's AI patent applications reached 1.5764 million, accounting for 38.58% of global applications, ranking first globally.

The report points out that in 2024, there were 40 notable models from US institutions and 15 from China, but the gap between the two countries on benchmarks such as MMLU and HumanEval is converging.

China's capital market is also accelerating support for large model companies, with Zhipu and MiniMax both preparing for IPOs.

Layer 6: Applications - Market Landing

How applications land in the market to provide value to society is most critical.

In 2024, US private AI investment was about $109.1 billion, while China was about $9.3 billion - a huge gap.

China's market has high population density and higher acceptance of AI, which is a better foundation for market development. According to a South China Morning Post survey, 80% of respondents in China are optimistic about AI, but only 39% in the United States.

Layer 7: Trust and Governance - Risk Management

This is the top-level designer's risk management for the future.

Ilya Sutskever, the main inventor of ChatGPT, after resigning, founded a safe artificial intelligence application company and raised $3 billion, showing that safe and trustworthy AI cannot be ignored.

China has "Interim Measures for the Management of Generative Artificial Intelligence Services," adopting a relatively inclusive attitude toward innovative things, which is correct.

The US National Institute of Standards and Technology (NIST) released the "Artificial Intelligence Risk Management Framework" AI RMF 1.0, positioned for voluntary industry adoption, but Trump is pushing for a unified national regulatory framework.

The scope of AI use, responsibility when problems occur, and thresholds for large-scale implementation are here. Data leakage, unauthorized access, model drift, compliance capabilities, convenient responsibility design, and traceability all need security and trust rules for protection.

Conclusion

The real winners are often not those strongest in one layer, but those who can "mesh" the seven layers together.

When formulating AI strategy, China should perhaps ask less "whether to develop our own large models" and more three realistic questions: Where does electricity come from? Where does data come from? Where does trust come from? The first two determine how fast you can run, the last determines how far you can run.

Competition in the AI era is increasingly like an "hardware-engineering-compliance-product" integrated endurance race, not a sprint on a PPT.

From the seven-layer AI cake model and combined with realistic development, in the next five years, China may make major breakthroughs in chips, but may need to do more foundational work in electricity and data, and also need better construction and guidance of the business environment for trust and security governance.