China is training its AI on America's, and Washington has noticed

Reviewed byNidhi Govil

8 Sources

Share

Chinese AI companies have made distilling American models routine industrial practice, a Reuters review finds. The technique is legal, and it lands just as Washington tightens chip curbs again.

Chinese AI companies have turned a quiet training shortcut into mainstream industrial practice, according to a Reuters review of more than 300 Chinese papers, patents, company filings and government records

8

1

. The technique, called distillation, means using the output of leading American models to train smaller domestic ones. It is not illegal. It also lands the day after Washington said it would tighten China's access to advanced chips from firms like Nvidia

8

1

. Reuters' assessment is that broad distillation could undercut the effect of the newly announced curbs

8

1

.

Source: Analytics Insight

Source: Analytics Insight

How the workaround works

Distillation is straightforward in principle. A team asks a powerful US model hard questions, keeps the answers, and uses them as training material for a smaller local model. For a developer priced out of hardware, the appeal is simple: one expensive query to an American model does repeated work. The team keeps the examples to train a local system around, as a former Alibaba engineer explained to Reuters

8

5

. The hardware alternative has gotten expensive. An eight-GPU Nvidia H200 server runs to roughly $900,000 on China's gray market

4

7

.

Source: ET

Source: ET

The reviewed documents show the practice at every stage of the last two years. In November 2024, researchers at Shanghai Jiao Tong University took about 5,000 answers from OpenAI's o1, translated them with GPT-4o mini, and assembled 10,750 bilingual examples to retrain Alibaba's Qwen2.5-72B

8

1

. In an April 2025 paper, Alibaba researchers described drawing on GPT-4 and GPT-4o to produce training data in languages beyond Chinese. The responses were then refined and scored before training smaller Qwen models

8

2

. This February, researchers from the same university and from Meituan, the food-delivery company, described having GPT-4o draft plans for how AI agents coordinate on programming problems

8

4

. Companies have built software to automate the process and employ specialist workers to prepare the training material

8

1

.

Beijing has encouraged the technique: a national AI and manufacturing policy package from eight government agencies in January backs model training, fine-tuning and distillation for industrial use

8

1

. Shanghai and Shenzhen have backed ways to cut computing requirements

8

1

. Those documents encourage distillation; they stop short of telling companies to distil American models specifically.

A dispute that went public

US Commerce Department officials raised the practice with Chinese regulators at meetings in Chengdu in July, warning that collecting output from leading American models could amount to intellectual-property theft

8

1

. Chinese officials called distillation a standard technique in AI development, said they were willing to act where conduct amounted to theft of intellectual property, and made no concrete commitments

8

1

. In September, US authorities publicly accused Chinese companies of aggressive, targeted distillation of American models on what officials call "industrial-scale distillation"

8

1

.

How much credit the technique deserves is itself contested. A US official familiar with those talks said Washington believes distillation is the main reason Chinese models have narrowed the gap with American ones

8

1

. Two people familiar with Chinese developers' operations dispute that, pointing instead to engineering, algorithms, data and model design

8

1

.

The technique has real limits: it cannot copy a leading model's full range of abilities, and it does not remove the need for advanced computing power

8

1

. It may breach contracts, though. OpenAI's terms restrict using its output to build competing models, with some exceptions, and Anthropic likewise requires permission before its models train rivals

8

1

. Chinese developers themselves see the ambiguity. In recent months they have asked the industry regulator, the Ministry of Industry and Information Technology, for guidance on where ordinary use of foreign models ends and organized collection for training begins. That is according to one person familiar with the requests

8

3

. Even the researchers behind the 2024 project cautioned that leaning too hard on distillation could choke original innovation, leaving developers waiting on whoever builds the more advanced model

8

1

.

The report lands two days before the Asia-Pacific Economic Cooperation summit in Shenzhen. Donald Trump and Xi Jinping are expected to meet there on November 18 and 19, with distillation now among the frictions between them

8

3

.

Source: Reuters

Source: Reuters

© 2026 TheOutpost.AI All rights reserved