Nvidia(US:NVDA) 是说芯语·2025-07-29 09:39Core Viewpoint - The article discusses the supply-demand imbalance of NVIDIA's H20 AI chips in the Chinese market following the lifting of export restrictions, highlighting the significant demand from Chinese tech companies and the challenges in meeting this demand due to geopolitical factors and supply chain constraints [2][3][4]. Supply and Demand Dynamics - After the lifting of export restrictions, there is a strong demand for H20 AI chips in China, with NVIDIA placing an order for 300,000 units from TSMC due to this demand [2]. - Jefferies estimates that the total demand for H20 chips in the Chinese market is around 1.8 million units, while NVIDIA's inventory is projected to be only between 600,000 to 900,000 units by early 2025, indicating a significant supply gap [2]. - Chinese tech giants have already placed pre-orders worth over $12 billion for H20 chips to secure future supply [3]. Geopolitical Influence - The H20 chip is not a standard product but rather a result of geopolitical influences, particularly U.S. export control policies that have fluctuated, impacting its availability in China [4]. - The U.S. government updated its export control list in April 2025, leading to a temporary halt in sales to China, which resulted in NVIDIA writing down inventory values by $4.5 billion to $5.5 billion in its Q1 2026 financial report [6]. Production and Supply Chain Challenges - NVIDIA's CEO announced in July 2025 that the company received U.S. government approval to resume shipments of H20 chips to China, but the supply chain is unable to respond immediately due to prior adjustments [7]. - TSMC faces challenges in fulfilling NVIDIA's orders, with a potential nine-month timeline required to restart the supply chain for H20 chips, which may not meet the immediate needs of Chinese customers [10][11]. Market Positioning and Competitiveness - The H20 chip is positioned to provide a legal and compatible high-performance computing solution for Chinese tech companies, especially in AI inference and large model training, amidst restrictions on advanced AI chips from the U.S. [5]. - Despite the H20's limitations in performance compared to the H100, its specifications, such as 96GB HBM3 memory and 4TB/s bandwidth, make it suitable for many applications in the Chinese market [12][13]. Challenges for Domestic Alternatives - Domestic AI chips struggle to replace the H20 due to the extensive reliance on NVIDIA's CUDA platform, which requires significant code rewriting and adaptation for alternative frameworks [12]. - The pricing strategy of the H20 chip is designed to create a competitive edge, being priced between $6,500 and $8,000, which is strategically positioned above the government procurement price of Huawei's Ascend 910B [14]. - The current state of domestic AI chips is at a critical point, requiring ongoing technological innovation and policy support to achieve broader market acceptance [15].