The development of DeepSeek’s next-generation AI model, DeepSeek-R2, has been delayed due to US export restrictions on NVIDIA’s H20 chips, The Information reported on Friday. The H20 chip has only been available in limited quantities due to tightened US trade policies, despite being designed specifically for the Chinese market. DeepSeek previously used 50,000 NVIDIA Hopper GPUs, including 30,000 H20s, to train its R1 model. The H20 shortage is reportedly also affecting the deployment of the existing DeepSeek-R1 model, which many users run on the restricted chip. Sources told The Information that DeepSeek has been aggressively working on the R2 model for months, but CEO Wenfeng Liang remains dissatisfied with its current performance, and the model is not yet ready for deployment. The export controls are part of broader US efforts to limit China’s access to advanced AI hardware, creating mounting pressure on Chinese AI firms reliant on high-performance GPUs from overseas. [The Information]
Related
