The DeepSeek R1 model was launched in January this year and it started a paradigm in training artificial intelligence (AI) models. DeepSeek R1 is said to have cheaper training costs and uses lower-powered chips, causing shares of companies like NVIDIA to suffer at the time. With the success of R1, many are waiting for the next model DeepSeek R2 which is believed to be launched in May.
May has come and gone but R2 has not been launched to this day. In a report by the Financial Times today, issues with Huawei Ascend chips are believed to be the reason for the delay of this AI model. Huawei Ascend chips caused several issues that forced DeepSeek to use NVIDIA chips for training purposes and Ascend only for inference tasks.
Industry players in China say that Chinese-made chips are still behind the capabilities of NVIDIA chips in terms of stability, software and connectivity. Last April Huawei CloudMatrix 384 was launched as a competitor to NVIDIA GB200 NVL72 systems with supposedly higher performance but requiring more electricity.
Huawei developed its own AI chips not only because of US technology barriers but also to enable China to be self-reliant with locally made technology. The strained relations between China and the US have led the government to advise local companies not to use NVIDIA H20 artificial intelligence (AI) chips, especially for government-related projects.