魏少军：摒弃英伟达 GPU ！

Core Viewpoint - The reliance on American hardware, particularly NVIDIA GPUs, poses long-term risks for China and its Asian partners in the AI sector, necessitating a shift towards independent technological pathways [1][6]. Group 1: Imitation Trap - Major tech companies in China and AI research institutions in South Korea and Japan predominantly use NVIDIA GPUs, such as the A100 and H100 series, for training large language models, which has become the standard in Asian AI labs [3]. Group 2: Upstream Hollowing - The core technologies of NVIDIA GPUs, including architecture design and software ecosystem, are controlled by American companies. For instance, the H100 chip is manufactured using TSMC's 4nm process, but its core IP and instruction set architecture are dominated by NVIDIA, creating significant technological barriers [4]. Group 3: Computing Cost Trap - The demand for computing power has surged exponentially as the parameter scale of large language models has increased from billions to trillions. Training a model with hundreds of billions of parameters requires thousands of NVIDIA A100 GPUs, with server costs exceeding one million yuan, and total training costs reaching hundreds of millions [5]. Group 4: Data Security Risks - The AI training and inference processes involve handling sensitive data, raising concerns about potential data breaches due to the reliance on NVIDIA's hardware and software, which are designed by American companies [7]. Group 5: Risk of Technological Innovation Stagnation - The imitation of the American model limits regional autonomy and stifles innovation. NVIDIA's CUDA ecosystem has become the dominant framework, with over 90% of AI developers using it, which reinforces the need for Asia to pursue independent innovation and establish a self-sufficient AI computing infrastructure [8].