The True Story of DeepSeek: Insights from Semi analysis’s Extensive Report

Debunking Rumors and Providing In-depth Analysis of DeepSeek’s Costs, Talent, and GPU Resources

Image by Author

Non-members can read here or subscribe to stay updated whenever I publish.

Over the past week or months, depending on when you find this article, DeepSeek has become a hot topic worldwide. Its daily active users (reportedly over 19 million) now surpass those of Claude, Perplexity, and Gemini. For those who have heeled the AI industry, this news is not entirely surprising. We’ve been discussing DeepSeek for months at my university as an international student in China and are familiar with the company. However, the intense hype surrounding it is unexpected.

Public opinion has shifted dramatically.

Last month, we debunked the myth when scaling laws were broken. Now, algorithm improvements are happening so quickly that they negatively impact Nvidia and GPU usage. Discussions have emerged about DeepSeek’s efficiency, suggesting we no longer need as much computing power, leading to significant overcapacity because of model transformations.

You’re reading Nov Tech, I’m Novy, and today we’re examining MLA Mode, Performance Metrics, and the Mis-estimated Costs of…