For years, Chinese chip developers have admired Nvidia as the American giant established an unrivaled position in the industry thanks to its hardware capabilities and the proprietary CUDA toolkit that engineers use to develop applications on the company's graphics processing units (GPUs).

Still, with the release of DeepSeek V3 and R1, they're getting more momentum to move away from Nvidia's orbit.

deepseek bloomberg
DeepSeek gives Chinese chip companies a boost to reduce their dependence on Nvidia. Photo: Bloomberg

Infinigence AI, a computing infrastructure platform provider, said it is working with seven leading Chinese chip developers: Biren Technology, Hygon Information Technology, Moore Threads, MetaX, Enflame, Iluvatar CoreX and Ascend.

During a visit to China in January, Nvidia CEO Jensen Huang said that about 1.5 million developers in the country were using CUDA. The company also partnered with more than 3,000 startups to help establish the technology industry there.

Exactly how many chips DeepSeek used to train its V3 and R1 models remains a mystery, but there has been speculation that Huawei chips played a key role in the startup's success.

According to Tom's Hardware, models running on the Huawei Ascend 910C can achieve 60% of the inference performance of the Nvidia H100. Dependence on the Nvidia software ecosystem is likely to decrease as AI inference becomes more important in the future.

Meanwhile, Chinese chipmakers are stepping up efforts to accommodate DeepSeek and help reduce their dependence on U.S. chips. A report by China Central Television (CCTV) showed that at least 15 domestic chipmakers have adapted their products to better accommodate DeepSeek training and operations, allowing DeepSeek models to be run on “domestic-developed computing infrastructure.”

It took Kunlun Xin, Baidu's AI chip company, less than a week to adapt its hardware to DeepSeek's open-source model, significantly faster than open-source models, according to CEO Ouyang Jian.

Last week, the company said it began supporting the DeepSeek V3 and R1 models, allowing developers to quickly deploy applications at low cost.

DeepSeek’s AI models are seen as a sign of China’s resilience in the face of US clampdowns. Some of the startup’s groundbreaking predictions will help spur AI development across a range of domestic industries.

Chip designer Loongson Technology has supported the DeepSeek model so users can run it on their computers. Huawei's cloud computing division has also partnered with AI infrastructure startup SiliconFlow to provide the DeepSeek V3 and R1 models on its Ascend cloud service. Huawei claims the performance is comparable to running the model on high-end GPUs.

(According to SCMP)