中国宏观经济月报：DeepSeek对英伟达长期股价的潜在影响

Group 1: Market Impact - DeepSeek's launch led to a significant drop in the semiconductor market, with the Philadelphia Semiconductor Index (SOX) falling by 9.2%, marking the largest single-day decline since March 2020[2] - NVIDIA's stock price dropped nearly 17%, resulting in a market value loss of approximately $600 billion, one of the largest single-day market cap reductions in U.S. history[2] - The success of DeepSeek may have a more profound impact on companies relying on self-developed large models, such as OpenAI and ByteDance, rather than directly on NVIDIA[12] Group 2: Cost Efficiency - DeepSeek achieved a training cost of under$6 million using approximately 2,000 H800 GPUs, with a total of 2,788K GPU hours utilized for training[6] - The input cost for DeepSeek's reasoning model is only 1 yuan per million tokens, significantly lower than OpenAI's model, which charges $15 per million tokens for input[7] - DeepSeek's innovative training strategies, including the use of pure reinforcement learning, have led to a substantial reduction in training costs and improved model performance[9] Group 3: Future Implications - The reduction in training and inference costs may lead to increased demand for computational resources, contrary to the expected decrease, as per Jevons Paradox[14] - DeepSeek's advancements could lower the barrier for small and medium enterprises to develop private models, potentially leading to a surge in demand for inference capabilities[15] - Despite the short-term impact on NVIDIA's stock, the long-term outlook suggests a potential increase in demand for NVIDIA's chips as AI commercialization accelerates[16]