The DeepSeek storm that erupted during the Lunar New Year left many in the tech industry shaken. Although Nvidia and TSMC’s stock prices initially took a hit but have since rebounded, the rise of this Hangzhou-based AI company, now with over 30 million users worldwide, continues to have far-reaching implications—not only for the tech sector but also for the U.S.-China technological balance of power.

For example, at China’s Two Sessions (National People’s Congress and Chinese People's Political Consultative Conference)—a key event for understanding China’s policy directions, which concluded in early March—AI and DeepSeek were hotly debated. The Financial Times even jokingly dubbed it the "DeepSeek Conference."

Many Chinese media outlets and influencers have used DeepSeek’s success as an example to boast that U.S. export restrictions on AI chips have been completely ineffective, claiming that Washington's sanctions are "meaningless."

This is precisely what Taiwan’s semiconductor industry is most concerned about: Are U.S. semiconductor export bans on China truly ineffective?

A report published in March by the Center for Strategic and International Studies (CSIS), titled "DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race," provides a shocking partial answer.

The Major Loophole in U.S. AI Chip Export Sanctions

It turns out that the Biden administration’s AI chip export ban on China had a massive loophole from the very beginning, which indirectly contributed to DeepSeek’s rise.

The story begins with DeepSeek’s launch of its V3 large language model in December 2023, the company’s first product to attract significant Western attention.

In a research paper, DeepSeek claimed that V3 was trained using just 2,000 Nvidia H800 GPUs.

The H800 was specifically designed for the Chinese market, a downgraded version of Nvidia’s most advanced AI chip at that time, the H100.

What shocked the Western world was that DeepSeek managed to train a globally competitive AI model using far fewer GPUs than Meta and OpenAI, and that too with a weaker chip.

U.S. AI companies eagerly studied DeepSeek’s research, admiring the many optimization tricks it revealed to enhance computational efficiency.

However, from the U.S. government’s perspective, H800 should have never existed in the first place.

Share

Nvidia’s Astonishing Downgrade Capabilities Left U.S. Officials Stunned

On October 7, 2022, the U.S. Department of Commerce imposed its first wave of AI chip export restrictions on China, effectively banning Nvidia’s entire lineup of data center GPUs from being shipped to the country.

Yet, just one month later, Nvidia released a downgraded version of the A100, called the A800.

Then, four months later, Nvidia launched the H800, a downgraded version of the most advanced H100, with 40% lower memory bandwidth and reduced computing power, just enough to meet the U.S. Commerce Department’s export threshold.

U.S. officials were caught completely off guard. A government official told CSIS that they had miscalculated the situation.

They originally expected that if Nvidia wanted to develop a compliant, China-specific version of its GPUs, it would take several years of redesign and development—during which China would be stuck using older, far inferior GPUs like the V100 from 2017.

"They never imagined that Nvidia could modify an existing chip’s performance even after it was manufactured," the report states.

This report finally answered a long-standing question I had: Why did the U.S. AI export bans leave so many loopholes?

(Continue reading the full story on Tech Taiwan Substack.)

Have you read?