Chinese AI Chip Companies Seek Training Alternatives

The rapid growth of DeepSeek, a private artificial intelligence startup from China, highlights a new path for Chinese AI chip firms aiming to innovate despite stricter technology regulations from the United States. This shift aims to challenge the dominance of Nvidia, a leading U.S. semiconductor company, and potentially transform the global computing sector.

DeepSeek's recent development of a competitor to ChatGPT at a significantly lower cost compared to similar products from U.S. companies demonstrates that high-performing language models do not solely rely on expensive Nvidia chips. This trend may open doors for increased use of domestic Chinese chip alternatives in the near future.

Further reinforcing this shift, major Chinese technology firms, including Huawei Technologies, have successfully adapted their products to run DeepSeek's large language model. Yang Jian, the chief technology officer of MetaX, a chip manufacturer based in Shanghai, noted that although DeepSeek utilized Nvidia graphics processing units (GPUs) for training, their overall expenditure on Nvidia technology was substantially less than the investments made by their American counterparts.

According to analysts from Jefferies, the cost of training a recent version of DeepSeek's models was estimated at just $5.6 million, which is less than 10% of what Meta spent for its Llama model.

Yang emphasized that this evidence suggests that sophisticated AI programs can be created using more affordable chips, igniting a renewed sense of optimism among Chinese semiconductor enterprises who are making significant strides.

Currently, Nvidia holds an impressive 98% share of the global AI training market. However, Yang predicts that by the end of this year, several large language models will be trained on Chinese AI chips, with this trend becoming significantly more pronounced by 2026. He stated, "The Chinese market will gradually evolve. Nvidia will continue to be part of the computing infrastructure alongside domestic chips, leading to a scenario where global computing power will consist of two parallel paths."

Adoption of DeepSeek's technology by cloud service providers like SiliconFlow showcases growing trust in local chips. SiliconFlow announced it has successfully run DeepSeek's AI models on its cloud infrastructure using Huawei's Ascend AI chips, achieving performance that is competitive with services based on Nvidia's high-end GPUs.

Wang Hua, vice president of AI and cloud computing at Moore Threads, mentioned that their self-developed GPUs enable quick deployment of DeepSeek's AI model inference services.

In light of these advancements, Pan Helin, an expert affiliated with the Ministry of Industry and Information Technology, stated that DeepSeek has created a new technical methodology for large language models, aiming to reduce computing power requirements. This method could encourage broader adoption of domestic processors, especially in the context of ongoing U.S. restrictions on advanced Nvidia chip exports to China.

According to Pan, "DeepSeek's success exemplifies a means to navigate the geopolitical barriers posed by U.S. technology bans, underlining the significance of innovation and adaptability in pursuing strategic objectives."

Experts recognize that while Nvidia's GPUs maintain an advantage in computing power and ecosystem development, AI models like DeepSeek are not expected to fully disengage from the Nvidia ecosystem in the immediate future. Nonetheless, improvements in domestic alternatives, more efficient algorithms, and a heightened focus on supply chain security will gradually lessen reliance on a single supplier.

Wang from Moore Threads commented, "This transition will take time and necessitates additional breakthroughs in domestic technology, but it is an ongoing trend. The global AI chip market is unlikely to remain under the control of any single company in the future."

AI, China, Semiconductors