Additional Coverage:
- DeepSeek caused a $600 billion freakout. But China’s AI upstart may not be the danger to Nvidia and U.S. export controls many assume (newsbreak.com)
Local investors panicked Monday as Nvidia’s stock plummeted nearly 17%, erasing almost $600 billion in market value. The drop was sparked by the emergence of a Chinese AI startup, DeepSeek, whose new AI models seem to rival American-made versions, despite reportedly being developed at a fraction of the cost. This raised concerns about Nvidia’s future, as its high-powered GPUs are a major expense in AI development.
DeepSeek claims to have used older Nvidia A100 GPUs, acquired before US export restrictions, and newer H800 chips, designed to comply with those restrictions. However, the number of chips used is significantly less than what American companies typically employ.
For example, Elon Musk’s Xai uses 100,000 advanced H100 GPUs in its Tennessee computing cluster. DeepSeek’s R1 model, designed to compete with OpenAI’s reasoning model, can even run on a laptop CPU.
Critics of US export controls argue that DeepSeek’s success proves these restrictions are ineffective and even counterproductive, forcing Chinese companies to innovate and create more efficient AI models. Some suggest these restrictions may have inadvertently boosted Chinese AI development.
However, DeepSeek’s success could actually increase demand for advanced AI chips, including Nvidia’s. This is due to the Jevons Paradox, an economic principle stating that increased efficiency in resource use often leads to higher overall consumption. Cheaper AI could make it more accessible for businesses, potentially driving up demand for computing power overall.
The key question is what kind of computing power will be needed. While Nvidia dominates the market for training large language models, competitors may have an edge in “inference,” the process of using trained models to perform tasks. This could erode Nvidia’s market share, but the company’s strong software and developer community give it a significant advantage.
Reasoning models like DeepSeek’s R1 also present a unique situation. They perform better with more computing power during inference, meaning advanced chips are still crucial for faster results. This is vital for business applications where speed is key.
Therefore, restricting China’s access to advanced chips may still be a valid national security strategy. Even if less advanced chips can create powerful AI, having access to the best technology remains a significant advantage. This could give the US an edge in economic and military AI applications.
Finally, there are doubts about DeepSeek’s claims regarding chip usage. Some speculate they have secretly acquired more advanced chips than admitted. If true, this would strengthen Nvidia’s position and suggest the issue lies not with the export control policy itself, but with its enforcement.