The Chinese generative artificial intelligence (AI) startup DeepSeek has garnered industry attention by mentioning 'next-generation Chinese chips' while unveiling an upgraded version of its flagship AI model 'V3.'
According to Reuters on the 22nd, DeepSeek announced 'DeepSeek V3.1' on the 21st, stating that the model features functions optimized for Chinese chips and boasts faster processing speeds. Amid the intensifying U.S.-China semiconductor hegemony competition, Reuters analyzed that DeepSeek has emphasized compatibility with Chinese chips in line with the recent movement by the Chinese government toward semiconductor technology independence.
DeepSeek explained that the V3.1 model, which integrates inference and non-inference modes into a single model, utilizes a data format called 'UE8MO FP8.' The UE8MO FP8 format significantly reduces memory usage while enhancing AI training and inference efficiency. DeepSeek noted, 'UE8MO FP8 has been designed to match the next-generation domestically produced chips that will be distributed soon.'
The Hong Kong daily South China Morning Post (SCMP) reported, 'This one-line introduction has prompted various speculations about what achievements China may have made recently in terms of semiconductor technology independence.'
If DeepSeek's technology combines with Chinese chips, it could pave the way for new breakthroughs in the combination of hardware and software. Moreover, this reference suggests that China has made significant progress in building an AI stack with indigenous technology, which SCMP added would also help in escaping U.S. chip export controls.
Some evaluations suggest that this V3.1 upgrade is relatively minor. What the industry is focusing on is the timing of the release of R2, the successor to the 'low-cost, high-efficiency' inference-focused model R1 that shocked the global AI market at the beginning of this year.
Since the launch of R1, DeepSeek founder Liang Wenfeng has led a life close to seclusion, and the company has maintained a quiet atmosphere. Some analyses have suggested that DeepSeek's usage rate has sharply declined.
As the release timing for R2 has been delayed, various suspicions have arisen. The Financial Times (FT) recently reported that DeepSeek delayed the launch after encountering issues using Huawei chips for R2 training. FT cited sources stating, 'DeepSeek adopted Huawei chips for R2 training at the recommendation of authorities but ultimately switched to NVIDIA chips due to technical issues.' This underscores the limitations of China's efforts to replace cutting-edge American technology with domestic technology, according to FT.
Previously, DeepSeek stated that it trained the V3 model using 2,048 NVIDIA H800 chips. However, it did not disclose the chips used for training V3.1.