DeepSeek-R2 model development is hindered: Nvidia H20 chip shortage

It is reported that the US chip export restrictions are causing a chain reaction in China's AI industry. According to foreign media reports, the shortage of Nvidia's H20 chips customized for the Chinese market has delayed the development of the DeepSeek-R2 artificial intelligence model. This bottleneck stems from the strategic adjustment of the US authorities, which cut off the key source of computing power and forced Chinese local AI companies to face R&D difficulties.

Specifically, DeepSeek relied on 50,000 Nvidia Hopper GPUs when training the previous generation R1 model, including 30,000 H20, 10,000 H800 and 10,000 H100 chips, all of which were purchased by the investor Gaoyi Asset. However, in the current intensive development stage of the R2 model, although the team is fully committed to optimizing, CEO Liang Wenfeng is still not satisfied with the model performance, and internal adjustments are continuously made to improve capabilities, resulting in the model not being ready for official launch. In addition, the shortage of H20 chips not only hinders the development of new models, but also affects the actual deployment of existing DeepSeek-R1 models. Because most users rely on these chips to run applications, supply interruptions have exacerbated overall business challenges.

It is worth noting that this shortage is the latest manifestation of the Sino-US technological game. The US government stopped the export of H20 chips in April and pressured Nvidia to adjust its China strategy and stop launching Hopper products. This move is aimed at curbing the rise of Chinese AI, especially for local companies like DeepSeek, whose innovations have caused concerns about US technological hegemony.