US Commerce Secretary nominee Howard Lutnick has accused DeepSeek, a Hangzhou-based artificial intelligence (AI) firm, of stealing US technology and getting around US export controls to obtain high-end Nvidia chips.
In the nomination hearing with the US Senate on January 29, Lutnick said DeepSeek could create its AI models “dirt cheap” because it was able to purchase a large quantity of Nvidia chips and steal data from Meta’s open platform.
“I take a very jaundiced view of China,” he said. “They only think about themselves and seek to harm us, and so we need to protect ourselves. We need to drive our innovation, and we need to stop helping them. Meta’s open platform let DeepSeek rely on it. Nvidia’s chips – which they bought tons of, and they found their ways around [export controls] – drive their DeepSeek model. It’s got to end.”
Lutnick said he will coordinate and empower the Bureau of Industry and Security (BIS)’s export controls and tariffs to stop China from using American tools to compete with the US.
Beijing has so far not responded to Lutnick’s comments as it is celebrating the Chinese New Year from January 28 to February 4.
David Sacks, former PayPal Chief Operating Officer and an advisor to the White House on AI and cryptocurrency matters, told Fox News that there was “substantial evidence” that DeepSeek used a method called “distillation” to extract data from Microsoft’s OpenAI models for its own use.
IT experts said “distillation” or “knowledge distillation” is commonly used in AI training. It is a technique where outputs from a larger AI model are used to train and improve a smaller one.
DeepSeek, in this process, can be understood as a student who keeps asking questions to a knowledgeable teacher, for example ChatGPT, and uses the answers to fine-tune its logic. At some point, DeepSeek will be as smart as ChatGPT.
The “distillation” process requires much less computing power than what OpenAI has used to train ChatGPT.
OpenAI told the Financial Times that it had seen some evidence suggesting that DeepSeek may have tapped into its data through “distillation.” It criticized DeepSeek for violating its intellectual property.
Some Chinese IT experts agree that DeepSeek was created through “distillation.”
Wang Zhiyuan, a Beijing-based IT columnist, writes in an article that it’s obvious DeepSeek V-3, released on December 26, 2024, had used the “distillation” technique in training. He says he came to that conclusion after analyzing the characteristics of DeepSeek.
He says a lot of other Chinese AI models have also used distilled data from ChatGPT o1, released on September 12 last year. He says that an academic paper published by a group of Chinese researchers on November 25, 2024, has already explained the distillation process and its effectiveness in detail.
He says an AI model made with distilled data may not be able to answer very difficult questions but is enough to solve high school-level problems. In his view, all small AI models should improve themselves with distilled data before entering the markets.
“Don’t laugh at those who took a short-cut!” Wang says. “DeepSeek used a special method to save computing power. After all, its training cost is only US$5.58 million, 1.1% of US$500 million of Meta’s Llama 3.1.”
After releasing the DeepSeek-R1 on January 20, 2025, a group of DeepSeek researchers published a paper on January 22, saying that its latest AI model achieves performance comparable to ChatGPT-o1.
They said the training of DeepSeek-R1 used the distilled data from Alibaba’s Tongyi Qianwen (Qwen) and Meta’s Llama. They said the DeepSeek-R1-Distill-Qwen version outperforms ChatGPT-4o.
50,000 H100 chips?
DeepSeek said it used only 2,000 units of Nvidia’s H800 chips to train its AI model. Its parent High Flyer, a Chinese hedge fund, said that it had amassed a cluster of 10,000 A100 chips before the US banned the exports of the chips to China in October 2022.
But now Lutnick suspects that DeepSeek bypassed the US export controls by importing high-end Nvidia chips via third-countries, such as Singapore.
The Wall Street Journal reported last July that some dodgy institutions in Singapore purchased Nvidia’s A100 chips and paid Chinese students to bring them back to China.
Without providing any evidence, Alexandr Wang, chief executive of the US-based Scale AI, told CNBC that DeepSeek has 50,000 units of H100 chips, the most advanced Nvidia chips on the market.
Xiang Zhiping, a Hubei-based IT writer, finds that plausible. “It’s no surprise if DeepSeek has 50,000 H100 chips. Any Chinese internet giant could have accumulated a lot of Nvidia chips,” Xiang says.
Even if DeepSeek has many chips, he says, it will use software and new IT frameworks to win the game, instead of following in the US firms’ footsteps to boost computing power endlessly.
The US banned the exports of A100 and H100 chips to China in October 2022, and then the slower A800 and H800 chips in October 2023. After this, Nvidia tailor-made the even-slower H20 chips for the Chinese markets.
On January 14, 2025, the Biden administration announced a regulatory framework to restrict the exports of American AI chips and models. The framework took effect on January 31.
Some observers have said the step-by-step strengthening of the chip ban gave China too much time to accumulate high-end AI chips.
Yong Jian is a contributor to the Asia Times. He is a Chinese journalist who specializes in Chinese technology, economy and politics.
Read: Beijing calls Biden a ‘liar’ and says hello to Trump
