Skip to content
Featured on Startup Fame

China trains its AI models using US outputs as Washington cracks down on chip exports

Chinese developers are feeding outputs from leading US AI models into their own systems to create lighter, cheaper alternatives—a method called distillation. The practice has spread from labs to factories across China, just as Washington moves to cut off access to advanced chips. US officials call it industrial-scale IP theft; Beijing calls it routine AI training.

Editor, Lazyfounder

Published 5 min read
China trains its AI models using US outputs as Washington cracks down on chip exports
Image: ET Tech (The Economic Times) via source

KEY HIGHLIGHTS

  • Chinese researchers use answers and code from OpenAI and Anthropic models to train smaller AI systems through a technique called distillation.
  • The US Commerce Department warned in July that some of this activity may amount to intellectual-property theft.
  • A January policy document from eight Chinese agencies explicitly encourages distillation for industrial AI deployment.
  • Alibaba co-founded a military-linked AI lab, and a Singapore company it worked with reportedly bought over $4 billion worth of Nvidia chips—some allegedly smuggled into China.
  • OpenAI and Anthropic both prohibit using their outputs to train competing models without permission.
  • Washington is tightening rules on renting overseas data-center chips to Chinese firms and plans further curbs on Nvidia exports.

Distillation: how China clones US AI without the chips

Take a large US-built model like OpenAI’s GPT-4 or Anthropic’s Claude. Ask it to solve math problems, write software, or draft legal briefs. Record the answers, translate them into Chinese, and use that dataset to teach a smaller, cheaper model. That’s distillation—an old academic trick that has now become industrial policy in China.

The payoff is clear: a startup can mimic the capabilities of a cutting-edge US model without buying thousands of expensive Nvidia chips. When US export controls slashed China’s access to those chips, Beijing doubled down on distillation to keep its AI ambitions alive.

From lab benches to factory floors—Beijing’s push

What started in university labs has gone mainstream. Last January, eight Chinese ministries jointly issued an “AI+Manufacturing” roadmap that explicitly tells companies to use distillation to trim training costs and computing requirements. Shanghai and Shenzhen have since launched local grants for startups that adopt the technique.

Chinese teams are already putting it to work. Researchers at Shanghai Jiao Tong University took 5,000 answers from OpenAI’s o1 model, translated them, and turned them into a bilingual training set for Alibaba’s Qwen2.5-72B. A separate team inside Meituan used GPT-4o to generate step-by-step plans for teams of AI agents, then distilled those plans into a leaner model tuned for factory automation.

Washington accuses Beijing of large-scale IP theft

In September, US officials publicly accused Chinese companies of running “industrial-scale distillation” campaigns. A month earlier, Commerce Department officials had warned their Chinese counterparts that repeatedly harvesting answers from US models could cross into intellectual-property theft.

China rejects the charge. Officials describe distillation as a standard, globally accepted training method and say they will act only when clear evidence of IP theft emerges. Meanwhile, Chinese AI developers have formally asked Beijing’s industry regulator for guidance on where ordinary use ends and organised theft begins.

Chips, smuggling, and a billion-dollar shadow network

Even as China turns to distillation, it still wants the chips. A single Singapore-based company, Megaspeed, bought more than $4 billion worth of Nvidia hardware in just a few years. Megaspeed’s game: sign “customer commitments” with Chinese cloud providers like Alibaba, use those letters to raise capital, then stuff the chips into data centres in Malaysia—where some reportedly vanished, only to reappear in China.

One of Megaspeed’s co-founders, Alice Huang, is described by US officials as a Chinese intelligence asset. During investor pitches in Shanghai, she boasted of her ties to Jensen Huang, Nvidia’s former CEO. Nvidia says its sales team merely introduced Huang to Alibaba in 2022; the company was never told the chips might end up in military-linked labs.

A military lab, lawsuits, and new US export rules

Alibaba co-founded an AI lab in Hangzhou with Zhejiang Lab, a research centre that US sources say is deeply embedded in China’s military AI programs. The lab’s teams work on offensive cyber tools and war-game simulations. It is unclear whether any of the Nvidia chips allegedly smuggled from Malaysia ever reached Zhejiang Lab, but the centre’s own website advertises access to export-controlled Nvidia hardware.

When the Pentagon listed Alibaba as a company supporting China’s military, the company sued. In October, a US judge granted a preliminary injunction blocking the listing. Days later, the Commerce Department unveiled a fresh round of proposed restrictions on Nvidia and others, this time targeting chips rented in overseas data centres. Malaysia has opened its own probe into Megaspeed’s subsidiary, Speedmatrix; YTL, the Malaysian conglomerate that hosted the chips, saw its share price drop almost 20 percent in two weeks.

Cheaper AI today, fewer breakthroughs tomorrow

Distillation narrows the gap—Chinese models can now write code or explain physics nearly as well as their US teachers—but it does so by standing on those teachers’ shoulders. Some researchers fear that over-reliance on foreign outputs will keep China stuck in catch-up mode, unable to take the bold leaps that create entirely new architectures or algorithms.

Both OpenAI and Anthropic have updated their terms of service: neither allows developers to use model outputs to train a competing system unless there is an explicit exception. Enforcement is difficult, however, especially when translation, re-formatting, or scraping via third-party APIs obscures the origin of the training data.

Key takeaways

  • Distillation gives China a low-cost way to match US AI benchmarks without buying high-end chips, blunting the immediate impact of export controls.
  • US accusations of industrial-scale IP theft and China’s military links to cloud providers are hardening the divide, making cross-border collaboration riskier for startups.
  • As Washington closes off both chip sales and overseas cloud access, Chinese labs will likely double down on distillation—at the cost of original research breakthroughs.

Sources

  1. ET Tech (The Economic Times) · 2026-11-17
    China taps US AI models to train homegrown rivals as chip curbs bite
  2. The Next Web · 2026-11-19
    Megaspeed’s backer Alibaba co-founded a military-linked lab, NYT reports

This story is an original summary drafted with AI by Lazyfounder from the reporting listed above and checked by automated validation. Facts are attributed to their original publishers; sections marked as analysis are Lazyfounder's. Where a source is in another language, facts were machine-translated and quotations are reported, not reproduced. Read the original coverage via the links, and see our AI policy and corrections policy.

About the author

Editor, Lazyfounder

Tarun Mottlia edits LazyFounders, covering Indian startups, funding rounds, AI and product launches. Every story on the site is AI-assisted and checked against its cited sources before publication.

More stories by Tarun Mottlia

Get the LazyFounder Brief

Startup, funding and AI news in a five-minute read. Join the early-access list.

Lazy Founder - Powered by Blogy.in

Contact us

Have a story tip, correction or partnership idea?

Write to us at tarun.kumar@blogy.in or talk to the founder directly. We read every message.

Contact us