Nvidia Loses $400 Billion in Market Value as Chinese AI Company DeepSeek Shakes the Industry

  • DeepSeek claims its AI training infrastructure uses 2,048 Nvidia H800 chips.

  • Analysts speculate it may also rely on 50,000 H100 GPUs acquired through intermediaries.

Nvidia loses $400 billion in market value
No comments Twitter Flipboard E-mail
juan-carlos-lopez

Juan Carlos López

Senior Writer

An engineer by training. A science and tech journalist by passion, vocation, and conviction. I've been writing professionally for over two decades, and I suspect I still have a long way to go. At Xataka, I write about many topics, but I mainly enjoy covering nuclear fusion, quantum physics, quantum computers, microprocessors and TVs. LinkedIn

This is the news of the day—and perhaps the week. The Chinese open-source AI model DeepSeek R1 is sending shockwaves through the U.S. tech industry, and it’s not just because of its open nature. While its business model is already stirring concern among U.S. AI and semiconductor companies, what’s even more surprising is the relatively modest infrastructure driving DeepSeek’s success.

To grasp what’s happening, it’s important to remember that Nvidia H100 chip is currently the go-to GPU for AI workloads in data centers. The company has already begun shipping its successor, the B200 platform, which is even more powerful on paper. However, U.S. government sanctions prevent Nvidia from selling these GPUs to its Chinese customers. This is where DeepSeek finds its edge.

Deepseek’s Efficiency and Openness Are Making Waves in Silicon Valley

Chinese companies developing and training AI models have had to get creative, often purchasing Nvidia’s top GPUs through intermediaries or parallel markets, though these channels may not provide the quantities they require. According to the Financial Times, DeepSeek relies on an infrastructure that pairs 2,048 Nvidia H800 chips to train its model, which contains 671 billion parameters, at a cost of just $5.6 million.

If these figures are accurate—and they seem to be—they reveal a compelling truth: DeepSeek’s engineers have built a highly competitive AI model at a fraction of the cost it took OpenAI or Google to create similar models. Much of that cost efficiency can be attributed to the H800 GPU, a version of Nvidia H100 chip created in response to U.S. sanctions under the Biden administration.

Nvidia’s engineers intentionally reduced the performance of the chip to meet the Commerce Department’s export restrictions.

When the U.S. government prohibited Nvidia from exporting its most advanced GPU at the time, the H100 chip, to Chinese customers, Nvidia’s engineers devised a workaround. They intentionally reduced the performance of the chip to meet the Commerce Department’s export restrictions, enabling them to sell the modified version in China. The result was the H800 GPU—a scaled-down, less powerful variant of the H100.

The situation escalated on November 16, 2023, when the U.S. government introduced new sanctions against China, extending restrictions to include the export of Nvidia's H800 GPUs. By that time, DeepSeek’s engineers likely had already secured the H800 chips they needed. However, some analysts suggest that DeepSeek’s infrastructure includes an aggregated 50,000 H100 GPUs, allegedly acquired through intermediaries.

If true, the escalating tensions between the U.S. and China would make it highly unlikely for DeepSeek to openly admit to possessing thousands of GPUs obtained through questionable channels. Meanwhile, the fallout has shaken the tech industry. Nvidia, Microsoft, ASML, and other major technology companies are seeing their stock prices drop sharply, with Nvidia alone losing $400 billion in market value.

This stems from the possibility that DeepSeek’s success could demonstrate that the most powerful GPUs from Nvidia—or any other company—are not essential for training state-of-the-art AI models. If DeepSeek is indeed achieving groundbreaking results with just 2,048 H800 chips, it could fundamentally challenge the high-cost approach of OpenAI, Google, and others, forcing the industry to prioritize optimization and efficiency.

Image | Nvidia

Related | These Are the Companies That Dominate the Chips and Semiconductor Industry. Unsurprisingly, the U.S. Takes the Cake With Nvidia

Home o Index