DeepSeek, a Chinese AI lab, has garnered attention for its AI model R1, which recently made waves in the industry by performing at or beyond the level of top U.S. models like OpenAI’s o1 and Meta’s Llama.
The lab’s breakthrough comes with significant hardware support, as the CEO of Scale AI, Alexandr Wang, revealed in a recent CNBC interview that DeepSeek has access to approximately 50,000 of NVIDIA’s H100 GPUs for training its AI models.
Wang shared that DeepSeek’s R1 model performed exceptionally well in Scale AI’s “Humanity’s Last Exam,” a test that challenges AI models with complex questions from fields like math, physics, biology, and chemistry.
According to Wang, DeepSeek’s R1 model either outperformed or matched the top American AI models in this rigorous test.
Despite the U.S. government’s export restrictions on advanced NVIDIA chips, which limit China’s access to cutting-edge hardware, Wang believes DeepSeek has managed to secure a significant stockpile of H100 GPUs.
He noted that the Chinese lab likely has more of these GPUs than many anticipate, although future acquisitions may be limited due to ongoing export controls.
DeepSeek’s rapid rise in AI development, paired with its access to massive GPU resources, signals a shift in the AI race between the U.S. and China. The lab’s ability to compete with major American players in AI technology may have lasting implications for the global AI landscape.
Leave a comment