The successor to the R1 model will have to wait a little longer.
According to two anonymous sources from The Information, a shortage of Nvidia GPUs is hindering the production of the R2 model of DeepSeek.
R2 is Delayed
A few months ago, Chinese DeepSeek surprised the AI world with its R1 model, trained on fifty thousand Hopper Nvidia GPUs. 10,000 of these were H100 chips, another 10,000 H800 chips, and 3,000 H20 chips. The model performed as well as top models at a fraction of the cost. It soon became clear that security was lacking and the internal workings stirred controversy.
Since the export restrictions by the US to China, it has become difficult to obtain those H20 GPUs. Those already in China are being fully utilized by DeepSeek’s customers. The R1 model is reportedly used by Chinese companies and government agencies, which is rapidly depleting DeepSeek’s own H20 capacity.
Where is the Solution?
Without additional computing power, the R2 model cannot improve, and R1 is also facing issues. Usage is growing faster than the available chip supply, which could lead to reduced performance. Chinese alternatives like Huawei are not powerful enough and do not work with Nvidia’s CUDA software.
Where DeepSeek once made waves, it now seems to be losing its momentum compared to American rivals like OpenAI and Anthropic.