China’s AI Strategy Turns American Models Into Training Resources

China’s efforts to develop independent artificial intelligence capabilities are increasingly relying on an unexpected resource: the technology developed by its American competitors. As restrictions on advanced semiconductor exports continue to constrain access to computing hardware, Chinese researchers and technology companies are finding ways to extract value from leading American AI models without necessarily developing equally powerful systems from the ground up.

The approach, known as model distillation, is becoming an important part of China’s broader artificial intelligence strategy. It involves using a more advanced AI model to generate answers, solve problems, write computer code or demonstrate particular capabilities. The resulting material is then used to train a smaller model, allowing developers to reproduce selected functions at a lower computational cost.

An examination of more than 300 Chinese academic papers, patents, company disclosures, recruitment notices and government documents has revealed how this approach is moving beyond isolated research projects into wider industrial applications. The development raises questions about the effectiveness of American technology restrictions, the economics of AI development and the increasingly complicated relationship between technological competition and intellectual property rights.

Semiconductor Restrictions Are Changing the Economics of AI

The United States has increasingly used export controls to restrict China’s access to advanced semiconductor technology. The measures are intended to slow China’s ability to develop powerful computing infrastructure, particularly the systems required to train and operate sophisticated artificial intelligence models.

Advanced graphics processing units are central to this process. Training large AI models requires enormous computing capacity, and access to high-performance chips can influence how quickly companies develop new systems and how much money they need to invest.

For Chinese developers, restrictions create a problem that cannot be solved simply by spending more money. Even companies with substantial financial resources may face limitations in obtaining the most advanced processors and associated technology.

This has encouraged a search for alternative approaches that make more efficient use of available computing resources. Instead of attempting to reproduce every capability of a large American model, developers can use its outputs to improve smaller domestic systems.

The distinction is important. Model distillation does not eliminate the need for advanced chips, nor does it automatically give a smaller model all the capabilities of its larger counterpart. However, it can reduce the amount of computing required to develop particular functions, making the process economically attractive.

China’s approach therefore reflects a broader adjustment in its AI strategy. Rather than treating access to advanced hardware as the only measure of technological progress, developers are also concentrating on extracting greater value from existing models, data and engineering expertise.

How American AI Knowledge Is Being Reused

The distillation process works through a relationship between a more capable teacher model and a smaller student model. The teacher generates responses to selected questions or demonstrates how particular tasks should be completed. Developers then collect, organise and evaluate those responses before using them as training material.

The advantage is that a company does not necessarily need to make the expensive teacher model perform every task repeatedly. It can retain selected examples and use them to train a domestic system capable of handling similar problems.

Chinese research institutions have already explored this approach in several projects. A November 2024 research paper from Shanghai Jiao Tong University described the use of approximately 5,000 answers generated by OpenAI’s o1 model. Researchers translated some of the material and assembled 10,750 bilingual training examples to improve Alibaba’s Qwen model.

Another project described by Alibaba researchers in 2025 involved using GPT-4 and GPT-4o to generate training material in languages other than Chinese. The responses were subsequently refined and evaluated before being used to train smaller versions of the company’s domestic models.

Research published in February 2026 by Shanghai Jiao Tong University researchers and Chinese food delivery company Meituan also examined how GPT-4o could generate plans for coordinating multiple AI agents working on programming problems.

These examples indicate that distillation is not restricted to copying simple answers. It can also be used to transfer selected problem-solving approaches, programming capabilities and methods of coordinating different AI systems.

However, there is a limitation. A smaller model trained using another system’s outputs does not automatically inherit the original model’s full reasoning capabilities or underlying architecture. Its performance depends on the quality of the training examples, the selection of tasks and the engineering methods used during development.

Washington’s Concerns Go Beyond Technology Competition

The growing use of American model outputs has become a sensitive issue in relations between Washington and Beijing. The dispute is not simply about whether Chinese companies are using foreign AI products. It concerns the distinction between legitimate access to a commercial service and the systematic extraction of its capabilities to develop a competing product.

Using an AI model to generate information is not inherently illegal. Distillation is also a recognised technique in the global AI industry and is used by developers in several countries.

The controversy emerges when the process involves large-scale, organised collection of outputs in ways that may violate contractual restrictions or intellectual property protections.

American AI companies have introduced contractual conditions restricting the use of their model outputs to develop competing systems, although the precise restrictions and exceptions differ between providers.

US officials have previously raised concerns that some Chinese companies may be collecting American model outputs on an industrial scale. During discussions with Chinese regulators in July, American officials reportedly warned that certain activities could cross the boundary into intellectual property theft.

Chinese officials defended distillation as a standard development technique and indicated that they were open to action where actual intellectual property violations occurred. However, the discussions did not produce concrete commitments.

The difficulty for Washington is that distinguishing legitimate technical research from organised commercial extraction can be complicated. A company may use an American model for ordinary research and subsequently incorporate the resulting information into a domestic product. Establishing whether that activity violates contractual or legal restrictions requires examination of how the information was obtained and used.

Beijing Sees Smaller AI Models as an Industrial Advantage

China’s interest in distillation also reflects its ambition to spread artificial intelligence across manufacturing and other major sectors of its economy.

Large AI models can be expensive to train and operate. Their computational requirements can make widespread deployment difficult, particularly for manufacturers and smaller businesses that cannot afford extensive data centre infrastructure.

Smaller models offer a different economic proposition. They can be designed for specific industrial tasks and may require considerably fewer resources to operate. This makes them potentially useful in manufacturing, logistics, software development and business operations.

Chinese government policy has increasingly encouraged the development of AI applications across industry. A national AI and manufacturing policy package published by eight government agencies in January included support for model training, fine-tuning and distillation for industrial purposes.

Shanghai and Shenzhen have also encouraged methods intended to reduce computing requirements.

This suggests that China’s objective is not necessarily to create one domestic model that surpasses every American competitor. A more practical strategy may involve developing numerous specialised systems that can be deployed across different industries at manageable costs.

For an economy with a large manufacturing base, this could offer considerable commercial advantages. AI systems that improve production planning, equipment management and quality control may deliver measurable economic benefits even without matching the most advanced general-purpose models.

Dependence on American Innovation Remains a Weakness

Despite its advantages, the strategy also exposes a potential weakness in China’s AI ambitions. Relying heavily on the outputs of foreign models may help domestic developers close specific performance gaps, but it does not necessarily encourage original breakthroughs.

Researchers involved in earlier distillation work have themselves warned that excessive dependence on the technique could weaken independent innovation. If domestic developers increasingly depend on American systems to generate difficult answers and problem-solving examples, the long-term result could be a technology industry that becomes skilled at adapting foreign advances rather than producing its own.

There is also a question about sustainability. Access to foreign AI models can change because of contractual restrictions, commercial decisions or government intervention. If American providers restrict access to their services, Chinese developers may find it harder to continue using those systems as training resources.

The strategy therefore offers an important short-term advantage but does not remove China’s need to strengthen its own semiconductor industry, research capabilities and model architecture development.

The larger significance of the development is that the competition between the United States and China is no longer confined to chip manufacturing and access to computing infrastructure. It is increasingly about who controls the knowledge, training material and technical capabilities that determine the next generation of AI systems.

Washington may restrict the hardware available to Chinese developers, but controlling the movement of information generated by commercial AI systems is a considerably more complicated challenge. Beijing, meanwhile, has an incentive to use every available method to reduce the cost of developing domestic technology.

The resulting competition could redefine how technological leadership is measured. Access to advanced chips will remain important, but the ability to develop useful AI systems with fewer resources may become equally significant.

(Adapted from TBSNews.net)



Categories: Economy & Finance, Geopolitics, Regulations & Legal, Strategy

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.