DeepSeek has rocked the tech world with its AI model R1 because it delivers the performance of OpenAI’s o1 model, at only 5% of the cost.
How can Scale AI, the company behind DeepSeek, deliver that performance at such a low cost? The CEO and founder of Scale AI, Alexandr Wang, stated in an interview that AI model DeepSeek “has about fifty thousand H100s.” That caused controversy because Nvidia’s H100 GPUs are not allowed to be sold to China because of U.S. export restrictions.
Hitting the stock market
DeepSeek ‘s low-cost approach could lead to democratization of AI tools, but it is too early to talk about that. NVIDIA has since responded at WCCFTech: “DeepSeek is a perfect example of Test Time Scaling, with fully export compliant techniques. Inference requires significant numbers of GPUs, but DeepSeek implemented everything legally.”
Despite this statement, NVIDIA’s shares plummeted 17 percent, fearing that DeepSeek will make hyperscalers less relevant. It remains to be seen how revolutionary DeepSeek’s effective results will be.