The focus of AI is pivoting from training to inference, according to TechInsights. As AI companies like DeepSeek-R1 highlight the growing need for inference computational power, Nvidia's market dominance remains strong, with demand for its GPUs continuing to rise. However, this shift presents a chance for local Chinese AI chipmakers to step up, as several have already aligned with DeepSeek's requirements.
During the Lunar New Year, companies such as Huawei Ascend, MetaX, Biren Technology, and others were quick to ensure compatibility with the DeepSeek model. Yet, local manufacturers are still far from matching Nvidia's capabilities in running popular open-source large language models (LLMs), with Nvidia GPUs supporting over 10,000 models on platforms like HuggingFace, while local chips only support a fraction of that.
Yue Ma, chairman of Gitee AI, admitted that while Chinese AI chips have made strides, they lag behind Nvidia in open-source model compatibility. Meanwhile, experts like Guo-hao Dai of Shanghai Jiao Tong University and Infinigence AI stress the importance of optimizing algorithms through PTX language to unlock hardware potential, although the technical barriers are high.
Despite local advancements, challenges persist. Wei Chen of Memory Compute cautioned that DeepSeek's optimization of GPU power has reached its technical limits, with proprietary technologies like mixed-precision storage further complicating replication efforts. Achieving compatibility with Nvidia's CUDA platform is key to enabling DeepSeek's large model to run on local hardware. However, time is crucial: prolonged integration delays may hinder commercial viability for local manufacturers.
MetaX CTO Jian Yang noted that for Chinese chipmakers like Huawei and Biren Technology, the ability to quickly achieve compatibility with DeepSeek's evolving models will determine their success. This reflects a broader shift in the AI chip market, with local companies increasingly focused on inference power for AI applications in 2025 and beyond.
