[Breaking News] Elon Musk Says, “DeepSeek, AI Revolution With Low Value? In fact, I secretly wrote a huge amount of H100”

Elon Musk’s reasons for questioning DeepSeek’s claims and expressing skepticism can be largely explained by the realistic limitations of AI model training, inconsistencies in chip use, economic drivers and strategic factors.

First, training large-scale AI models with low-value (H800) is technically very difficult. DeepSeek claimed to have used H800 rather than Nvidia’s latest chip H100, but H800 is likely to experience bottlenecks in effectively training the latest large AI models, especially because of its lower computational performance by 30-40% compared to H100, and limited communication bandwidth. In other words, saying DeepSeek was “enoughly able to train with cheap chips” is an unrealistic argument given the experience of existing AI researchers and the chip’s specifications.

Also skeptical is the cost of AI training. DeepSeek said it costs $6 million (W8.6 billion) to develop AI models, which is unrealistically low compared to the cost of training the latest large language models (LLMs). For example, GPT-4 training cost more than $100 million, and Anthropic’s Claud model also required tens of millions of dollars of GPU resources. Thus, Gavin Baker’s argument, which Musk agreed to as an “interesting analysis,” suggests that DeepSeek likely announced a reduction in costs incurred in the actual AI research and training process.

The most decisive factor is the claim that DeepSeek actually has a large amount of Nvidia H100, but is hiding it. According to a CNBC interview with Alexander Wang (Scale AI CEO) shared by Musk, DeepSeek has 50,000 H100s, but it is likely that the U.S. export regulations have not been able to officially reveal them. In response, Musk left a comment, “Obviously,” because although DeepSeek claims to have used H800, it is likely to have utilized H100, which actually performs much better.

Musk is convinced that the claim that DeepSeek “made cheap AI out of cheap chips” is different from reality, and that it has actually used the latest high-priced chips and announced a reduction in costs. If these allegations are true, not only will the confidence in DeepSeek’s technological innovation falter, but its price competition strategy in the AI market could also be fictitious.