In Beijing on January 27th, Chinese startup DeepSeek made waves with the launch of its latest AI models, claiming performance on par with or better than leading U.S. models at a fraction of the cost. This development has disrupted the technology hierarchy. By stating in a recent paper that training DeepSeek-V3 required under $6 million in Nvidia H800 computing power, the company gained global AI attention.
DeepSeek's AI Assistant, powered by DeepSeek-V3, surpassed ChatGPT and became the top-rated free app on Apple's U.S. App Store. This success has sparked questions about the rationale behind the multi-billion-dollar AI investments made by U.S. tech firms, leading to repercussions for major players like Nvidia.
DeepSeek's models, DeepSeek-V3 and DeepSeek-R1, have garnered acclaim for their quality and cost efficiency, challenging the belief in the AI prowess gap between the U.S. and China. The company highlighted that its models are competitively priced compared to industry giants like OpenAI and Meta.
However, some voices have expressed doubts about DeepSeek's achievements. Scale AI CEO Alexandr Wang, in an interview with CNBC, raised concerns about potential violations of export controls related to the Nvidia H100 chips allegedly used by DeepSeek. Furthermore, Bernstein analysts noted discrepancies in DeepSeek’s training costs transparency.
DeepSeek, based in Hangzhou, is backed by Liang Wenfeng, co-founder of High-Flyer hedge fund, which shifted focus to AGI research before launching DeepSeek. The company's success has attracted attention from top echelons in China, with Liang attending a symposium hosted by Chinese Premier Li Qiang, hinting at Beijing's strategic interest in DeepSeek's achievements.