NVIDIA CEO Jensen Huang states that AI model distillation is competition rather than theft, contradicting U.S. government export-control stance.

Nvidia CEO Jensen Huang has characterized model distillation—the technique of training AI models on outputs from competing models—as a form of competition rather than intellectual property theft. This position directly contradicts the U.S. government's stance, which has classified Chinese AI companies' unauthorized extraction of American model capabilities as IP theft and raised the possibility of sanctions.

Speaking on CNBC's "Squawk Box" on September 28th, Huang responded to a question about whether model distillation constitutes "robbery" by stating simply, "That's called competition." He elaborated that "anyone can test other people's products as much as they want," noting that Nvidia's own products are routinely reverse-engineered by competitors seeking to understand their design. "I wish nobody learned from our products and we could just cruise along enjoying that benefit, but honestly, competition makes everything better," Huang said. He suggested that companies concerned about others learning from their products could identify their customers and restrict access.

Model distillation is a technique in which one AI model learns from the outputs of another to improve its performance and capabilities efficiently. While widely used in legitimate AI development, it has become a flashpoint in U.S.-China tech tensions following allegations that Chinese AI companies have accessed American frontier models at scale, extracted their outputs, and leveraged them to build competing systems.

The U.S. government takes a fundamentally different view. Treasury Secretary Scott Bessent warned in July that if foreign AI companies steal American IP through model distillation, the U.S. would classify it as theft and potentially impose sanctions. He specifically flagged Chinese companies' "covert, industrial-scale distillation attacks" as potentially subject to sanctions and export control designations if they cross the line into IP theft.

This concern has become increasingly concrete. Earlier in September, the Cybersecurity and Infrastructure Security Agency (CISA), the National Security Agency (NSA), and the Federal Bureau of Investigation (FBI) issued a joint advisory alleging that Chinese AI companies are conducting an "industrial-scale knowledge distillation campaign" against American AI firms. The agencies stated that these companies are systematically extracting the functions and capabilities of frontier models in violation of American AI companies' terms of service.

American AI companies have echoed these concerns. Anthropic reported earlier this month that Alibaba and DeepSeek, among others, engaged in "illicit distillation" of its AI models. China has denied U.S. allegations regarding improper model distillation by its AI companies.

Huang's remarks carry particular weight given Nvidia's imminent re-entry into China's AI market. The Information reported on September 27th that the Chinese government is moving toward final approval for domestic companies to purchase Nvidia's new RTX Pro 5500 graphics processing unit. The RTX Pro 5500 is a professional-grade GPU designed for AI model development and inference, featuring 84GB of GDDR7 memory—14GB less than the higher-end RTX Pro 6000—with a maximum power consumption of 600W. Chinese pricing has been set at 85,000 to 90,000 yuan, approximately $13,000.

Nvidia plans to ship approximately 500,000 chips per quarter to China beginning in late December, and has advised customers to place orders by September 30th to secure supply. If the Chinese government grants final approval, Nvidia will recover a significant growth engine in the Chinese AI market following a market freeze lasting more than a year.

Analysts interpret China's move toward permitting RTX Pro 5500 purchases as reflecting the judgment that rapid AI service proliferation makes it impractical to rely solely on domestically produced chips. The decision also reflects the impact of intensified U.S. enforcement against Nvidia chip smuggling through Southeast Asia, which has severely disrupted black-market supply channels into China.

China once represented a substantial portion of Nvidia's revenue. According to Nvidia's financial reports, revenue from Chinese companies (including Hong Kong) totaled $25.05 billion in the year ending January 2025, representing 19% of total revenue. This has declined sharply: $19.68 billion (9% of total) in the year ending January 2026, and $7.88 billion (8% of total) in the May-July 2026 quarter.

The contraction is far more severe when examining only AI data center chips, Nvidia's core business. Huang stated last October that "China's AI chip market share has fallen from 95% to 0%."

This trajectory originates in October 2022, when the Biden administration imposed export restrictions on Nvidia's high-performance AI chips to China to constrain the country's AI development. Nvidia responded by designing China-specific AI chips with reduced performance, such as the H20, to circumvent regulatory restrictions.

Beginning in April 2025, shortly after the Trump administration took office, even the H20 required government authorization for sales to China. As criticism emerged that U.S. export controls were driving China's push for technological self-reliance, the U.S. government conditionally permitted sales—but China then reversed course. The Chinese government pressured domestic companies to refrain from purchasing H20 chips, citing security concerns about potential information leakage to the United States and a desire to strengthen its domestic semiconductor industry.

In September 2025, Chinese authorities escalated pressure by announcing preliminary antitrust findings against Nvidia. The Trump administration conditionally approved exports of the H200, a higher-performance AI chip, in January 2026. After the Chinese government partially approved purchases in July, ByteDance and Tencent each received 10,000 H200 chips in August.

Huang's remarks underscore a fundamental debate in AI competition: where does legitimate learning end and unauthorized technology extraction begin? As the U.S. government seeks to restrict not only China's access to advanced semiconductors but also its acquisition of American AI model capabilities, Huang has articulated a more permissive perspective that views competitor product analysis and learning as inherent to market competition.

The broader context includes a development following the U.S.-China summit in Washington, D.C. on September 24th, where the two countries agreed to establish a "hotline" for communication regarding AI-related risks.