Chinese corporations are seeking collaborative solutions to become independent from NVIDIA's AI semiconductor ecosystem, focusing on DeepSeek. Until now, Chinese corporations have relied on NVIDIA's AI semiconductors as well as software designed by NVIDIA's 'CUDA platform' to develop AI models.
According to the South China Morning Post (SCMP) on the 13th, following the launch of DeepSeek's V3 and R1 models, Chinese semiconductor corporations are strengthening collaborations to break away from NVIDIA's ecosystem. It has been reported that seven semiconductor corporations, including Huawei, which supplied its semiconductors to enhance the inference performance of DeepSeek's AI models, are in discussions about collaboration, including Biren Technology and Haigon Information Technology.
DeepSeek has noted that it developed AI models comparable to those of global big tech companies at a lower expense. Initially, it was known that many low-spec chips designed by NVIDIA for the Chinese market were utilized, but it has been confirmed that Huawei's next-generation AI accelerator 'Ascend 910C' has been used to enhance the inference performance of DeepSeek's AI models, proving China's semiconductor ambitions. The AI chip is reported to be at about 60% of the performance of the H100, the top model in NVIDIA's Hopper series, the model prior to the recently released Blackwell platform.
Separately, it is recognized that the dependency on NVIDIA regarding the software ecosystem used to develop AI models remains quite high. DeepSeek also expressed concerns that NVIDIA's ecosystem still plays a vital role in developing AI models right after announcing its AI models, noting that U.S. industrial regulations could be an obstacle to the development of the Chinese AI industry. Jensen Huang, the Chief Executive Officer (CEO) of NVIDIA, who visited China in January, revealed that about 1.5 million developers in China are using the CUDA platform, and stated plans to collaborate with over 3,000 Chinese startups.
A semiconductor industry official said, "Not only did DeepSeek's AI models extensively use NVIDIA's AI accelerators during the process of learning vast amounts of data, but the CUDA platform was also utilized to develop them, and the reason Jensen Huang is investing in the Chinese market is that the demand from Chinese developers and AI corporations cannot be overlooked. If U.S. regulations on China intensify, China's 'AI ambitions' will inevitably face setbacks, leading them to strive for an independent ecosystem."
Chinese corporations are reportedly tightening efforts to supply solutions to support DeepSeek's AI models. Chinese media, including CCTV, reported that at least 15 Chinese semiconductor companies are upgrading their products to better train and execute DeepSeek. Huawei's cloud computing division also announced that it is collaborating with Beijing-based AI infrastructure startup Siliconflow to provide DeepSeek's existing V3 large language model and R1 inference model through its Ascend Cloud service. Longsoon Technology, a CPU design corporation in China, stated that it will support consumers using products equipped with its CPUs to smoothly utilize DeepSeek's AI models.
SCMP reported that "the emergence of DeepSeek has ignited the determination of Chinese corporations to build an independent AI ecosystem" and that "efforts to reduce reliance on U.S. technology will accelerate."